An Evolution of Step Testing and its Impact on Model Predictive Control Project Times

Size: px
Start display at page:

Download "An Evolution of Step Testing and its Impact on Model Predictive Control Project Times"

Transcription

1 White Paper An Evolution of Step Testing and its Impact on Model Predictive Control Project Times Executive Summary Bumping the process or step testing became a necessary part of implementing Advanced Process Control (APC) applications with the introduction of Model Predictive Control (MPC) in the 1970 s. MPC technology revolutionized APC in that it was able to use a model to make control moves based on the future predicted behavior of the process instead of relying solely on feedback from the current process measurement. The basis of these models were linear ordinary differential equations. Instead of having to derive these models mathematically, it was discovered that they could be obtained empirically from process data. To generate the required process data, it was necessary to step the process multiple times in an up and down fashion until the empirical models could be acquired by model identification algorithms. In the early days, manual methods were used to step the process. These methods, although sufficient to get models, typically required many weeks if not months of bumping the process. Deficiencies in these methods were discovered and new methods were invented to compensate and reduce project implementation time. The on-going practice of inventing new methods to improve on the previous one started in the 1990 s and continues today. The result has been a threefold reduction in project times over the last three decades. Efforts to improve on these methods continue today. Currently, intelligent stepping software is available that incorporates decades of experience. The software automatically performs many of the tasks previously done manually such as stepping the process, collecting data, identifying models, and validating models. The software even allows conducting these tasks while an MPC application is controlling the process. This enables step testing with less impact on the process operation and stepping while the process is being controlled. This modern stepping software is opening the door to implementing MPC on processes previously considered inapplicable to MPC. It also affords itself to easing the burden of maintaining existing MPC applications.

2 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 2 Table of Contents Introduction... 3 Brief History... 4 Manual Step Testing (1970s and 1980s)... 4 Open Loop Sequential Automated Stepping (1990) Open Loop Simultaneous Automated Stepping (2000) Open Loop Automated Stepping and Auto ID (2005) Closed Loop Automated Stepping with MPC and Auto ID (2005) Enhanced Closed Loop Automated Stepping with MPC and Auto ID (Since 2010) Recent Enhancements (Since 2010) Summary Table of Figures Figure 1. Table of various types of CV responses to an MV or DV unit change Figure 2. Illustration of manual stepping, process movement and model prediction Figure 3. Compensated MV moves correlated with stepping MV moves Figure 4. Result of making correlated compensated MV moves Figure 5. Effect of steps held for long periods Figure 6. Effect of large steps on distillation column Figure 7. Effect of large steps on a furnace Figure 8. Effect of large steps on a reactor Figure 9. Effect of large steps on a valve Figure 10. Open loop sequential step patterns Figure 11. Effect of compensated moves on model ID results using a PRBS pattern Figure 12. Effect of UMD and drift on model ID results using a PRBS pattern Figure 13. Comparison of sequential step patterns from appropriate and short settling times Figure 14. Effect on model ID of a PRBS step pattern from short settling times Figure 15. Open loop simultaneous PRBS step test patterns Figure 16. Effect of UMD and drift on model ID results using simultaneous PRBS patterns Figure 17. Comparison of simultaneous step patterns from appropriate and short settling times Figure 18. Effect on model ID of simultaneous PRBS step patterns from short settling times Figure 19. Automated stepping software Figure 20. Automated stepping software allowable step patterns Figure 21. Configuring the stepping software Figure 22. Collect data, step the process and monitor ID results Figure 23. Evaluate results and lock models Figure 24. Changing the MVs being stepped and stopping when done Figure 25. Connecting stepping software to an MPC application Figure 26. The stepping application and MPC application are connected Figure 27. Moderating interaction between steps and MPC moves Figure 28. Configuring the seed MPC application Figure 29. Specifying seed models for CV/MV pairs Figure 30. Downloading new models from stepping software to the seed MPC application Figure 31. Updated seed MPC model matrix

3 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 3 Introduction Model Predictive Control (MPC) was introduced in the 1970s. At its simplest it aims to use a model to predict the movement of the processes and inform control. Control moves can therefore anticipate future events, rather than simply reacting to current conditions. As implied in the name, the technology requires a model of the process. The basis of these models is a derivation of a differential equation, and more specifically a linear ordinary differential equation (ODE). Although the behavior of the process is not exactly linear, linear models do a good job of approximating it. The feedback mechanism built into all MPC technologies compensates for subtle nonlinear process behavior, and this is sufficient for controlling most processes. Linear ODEs also have many properties that make it easier to implement control algorithms and do closed-loop stability analysis. One is the principle of superposition: If an independent variable is made larger and then smaller by the same amount, the dependent variable response going in one direction will be a mirror image going in the other direction. The other key property is that the initial starting value of the independent variables and dependent variables is irrelevant in absolute value terms. Once starting values are established, the only thing that matters is how much the independent variables change. A simple form of linear ODE is expressed by the following equation: dcv/dt = (material, energy, momentum balance) To calculate the behavior of the controlled variable (CV) over time (t), it is assumed that a step input of an independent variable, also known as a manipulated variable (MV) or disturbance variable (DV), is made to the process. Once the step is made, the model describes the CV s behavior over time until it steadies out to the steady state (SS). In commercially available MPC technologies, derivation of the linear ODEs is not necessary. Approximations of the linear ODEs can be obtained empirically by bumping the MVs or DVs and using a least-squares data fitting algorithm to describe the behavior of the CVs. These empirically derived models come in many forms, such as step response models, finite impulse response models, Laplace transform models, various forms of auto regressive moving average models, state space models, and others. Irrespective of the form used by the commercial technology to describe the model characteristics, when used online, the models are all used in the same way to predict the future behavior of the CVs given changes in the MVs and DVs. Figure 1. Table of various types of CV responses to an MV or DV unit change

4 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 4 Figure 1 provides examples of typical responses. Various CV responses are shown along with their respective Laplace Transform. The top row illustrates first and second order models with and without deadtime. The second row illustrates inverse response and overshoot models. The third row illustrates zero gain (hump) models and ramp (integrating) models. The final row illustrates higher order (third and fourth order) models. Since all MPC technologies rely on process data to derive the models, it is necessary to go through the process of bumping the process (or conducting step testing), gathering data, identifying the model(s), validating them and commissioning the controller. Although these tasks are typically considered independent on a project, in reality, they are commingled. Step testing determines the basis for how long the other tasks take. Successful step testing tends to minimizes the time spent on the successive tasks. Improper step testing can create an iterative loop with successive tasks, wherein the engineer moves on to the next task only to find out that the models acquired from step testing was insufficient and the engineer is forced to revisit the step testing procedure. This iterative loop can initiate during model identification, model validation or commissioning. The net result of improper step testing is longer project times and a higher percentage of the project time spent on these tasks. This paper focuses not on the identification algorithms, online control algorithm, or form of model used. Instead it concentrates on methods used to step the process to develop models of sufficient quality to use in an online controller, and the effect these methods have on the time it takes to implement an MPC application. Brief History Step testing methods have evolved over 40 years: The 1970s and 1980s saw manual open loop step testing methods, conducted by process operators and engineers In 1990 open loop sequential automated step testing methods were first deployed 2000 marked the development of open loop simultaneous automated step testing methods In 2005 open loop automated step testing in conjunction with automated model identification methods were deployed Also in 2005 closed loop automated step testing in conjunction with an online MPC and automated model identification methods were deployed to maintain existing MPC applications 2010 saw enhanced closed loop automated step testing in conjunction with an online MPC and automated model identification methods. Enhancements continue to be developed. The main goal has always been to acquire quality models as quickly and with as little disruption to the process as possible. Progress over the past decades has been significant, with higher quality models acquired amid less disruption to the process and in shorter time frames. However, the evolution is on-going. It is useful to look at the methods in more detail. Manual Step Testing (1970s and 1980s) In these early years, the primary ambition was to simply get models that could be used online. Stepping the process was a new concept and met resistance. Engineers had to work closely with operators to implement the steps while ensuring key process variables did not violate operational constraints. Computing power was insufficient for sophisticated identification algorithms, so simple data sets were required. The general method deployed follows: 1. Set up the PIDs such that each MV corresponds to the SP of a PID or the OP of a PID (or a hand valve); in other words, make the underlying control system open loop. This requires breaking certain cascade structures or putting a PID in manual. 2. Allow the process to settle so it is not moving very much. 3. Negotiate the size of steps with the operator based on their tolerance for disruption to the process. 4. Make manual steps to a PID.SP or PID.OP, one MV at a time. The steps to the SP or OP are typically made by requesting the operator make the change. 5. Once stepping commences, the operator is responsible for ensuring CVs do not violate key operational constraints. The operator is instructed to make compensating changes to other MVs to prevent constraint violations.

5 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 5 6. Collect data off of an historian and take off-line to run through the identification (ID) algorithm. 7. Massage the data as necessary to remove bad data and rerun through the ID algorithm. 8. Repeat 6-7 until acceptable models are obtained. 9. Once acceptable models are obtained, commence stepping on another MV. 10. Repeat 2-9 until acceptable models are obtained for all the CVs and MVs. 11. Repeat 2-9 until acceptable models are obtained for all the CVs and DVs. 12. Validate all models using engineering knowledge and judgment. Models that do not make sense are either removed or stepping is conducted again to obtain a more realistic model. 13. Build the MPC application and run in prediction mode for the final validation of model quality. 14. With MPC online, step the MVs one at a time to evaluate the behavior of the predicted CVs against their respective process CV. 15. Massage data or re-step the process until the predictions follow the process closely enough for online control. 16. Begin commissioning by forcing key CVs to violate their constraints and evaluating how effectively the controller brings them back. If the response is insufficient and tuning cannot rectify the issues, massage the data or step the process once again. 17. Configure the optimizer and turn on. This typically causes the operation to slowly move from one operating state to another. 18. Evaluate the performance of the optimizer and further massage models or re-step the process as necessary. Figure 2. Illustration of manual stepping, process movement and model prediction Figure 2 illustrates a typical MV manual step sequence (top), the process movement result from stepping the MV (middle), and the expected prediction results overlaid on the process movement (bottom). These illustrations depict the expected results from the manual stepping procedure described above. However, processes are rarely so cooperative.

6 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 6 The explanation of the manual method shows the inherent inefficiency. There are a myriad of reasons for this, but three principal factors contribute to it: compensated MV moves are correlated with the MV being stepped; each step is held for a long time; and the steps made are too large. When steps are made manually, the intended step pattern is broken to address CV constraint violations. Unfortunately, breaking the step pattern often results as an attempt to counteract unmeasured disturbances or drift causing CVs to violate their constraints. This frequently results in erroneous data as the true movement of the process due to the steps is not captured. Another method used to address CV constraint violations during this period was making compensating changes with other MVs to keep key CVs from violating their constraints. Again, when people make the steps, it is natural for correlated compensated changes to be made between the MV being stepped and MVs used to counteract constraint violations. This tends to prolong the duration of step testing and results in repeat stepping of many of the MVs. The ID software can handle changes being made to multiple MVs but cannot deal with correlated changes that essentially cancel out steps. Figure 3 illustrates steps made to an MV (above), correlated compensated moves made to a different MV (middle), and the resulting process movement from all steps (bottom). Instead of the process moving in an easily identifiable pattern, as illustrated in Figure 2, the process is hardly moving at all and it is impossible to make sense of the contribution each MV is having on the process movement. To make matters worse, the real process movement will also have unmeasured disturbances and drift in it that renders the data essentially useless for obtaining models. Figure 3. Compensated MV moves correlated with stepping MV moves Figure 4 illustrates the results of making this type of move pattern. The correlated MV moves are shown (top), a comparison of the process movement with and without unmeasured disturbances and drift (middle blue and middle green, respectively), and the predicted results of the models obtained from both process movements (bottom). In the bottom graph, the green is the process movement without UMD or drift; the blue represents the prediction from the model ID performed on the data without UMD or drift; and the red represents the prediction from the model ID performed on the data with UMD and drift. The blue prediction seems to follow the process reasonable well, but it is difficult to discern. The red prediction does not represent the process movement at all, and the models that generated it are the likely results from real process data with these types of compensated MV moves.

7 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 7 Figure 4. Result of making correlated compensated MV moves. The result was step testing data that did not yield accurate enough models to use online and the need to redo the step testing for the MV. This is something that obviously needs to be avoided for any step testing method to be successful. Figure 5. Effect of steps held for long periods The next issue encountered with the manual method is that steps are held until the process settles out. Long periods between steps were best practice at the time, but this allowed UMD or drift to affect the quality of the data. Figure 5 illustrates the result. The top graph illustrates the process movement without UMD or drift (green), the process movement with UMD and drift (blue) and the resultant prediction from the model ID using the process movement with UMD and drift (red). The duration of the stepping was 1320 minutes.

8 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 8 The bottom graph illustrates the process movement without UMD or drift (green), the prediction from the model ID using process movement without UMD and drift (blue), and the prediction from the model ID using process movement with UMD and drift (red). In the top graph, the prediction looks like it fits the process without UMD and drift fairly well. However, in the bottom graph, when compared to the prediction from the process without UMD and drift, the prediction is not quite accurate. The model that produced the red prediction differs in gain to the model that produced the blue by 25%, and the red s settling time is about 30% shorter. As illustrated, holding steps for long periods makes the data sensitive to UMD and drift that naturally occur in processes. This unusable data must be massaged out to get better models from the ID algorithm, or the step testing period prolonged. The result is very long step testing periods for each MV to ensure sufficient data for quality models. The final issue encountered with the manual method is that steps made to the process are many times too large and do not resemble the MV moves the MPC would make. Larger steps are implemented to ensure a response to the CVs can be seen. However, large steps tend to push the process well out of equilibrium, while smaller changes made by the MPC will not. Figure 6 illustrates a potential problem encountered when making large steps to distillation columns. The tower on the left illustrates two examples of what happens to the temperature profile when a large reflux move is made suddenly. The tower on the right illustrates what happens to the temperature profile when the MPC makes the same size reflux move but over a longer period. The upper curve on the left tower shows the temperature profile sagging more when the move is made suddenly. Figure 6. Effect of large steps on distillation column The model produced from the ID reflects the process behavior during step testing, which is the behavior expected when the MPC makes moves online. However, when the MPC is online, the actual process behavior differs from the model s prediction. In this example, the model gain is fairly close to reality, but the dynamic behavior differs significantly. The lower curve on the left tower shows similar attributes to the upper curve, but the sag is more exaggerated. The model gain is very different from the reality. Both examples lead to errant models, but the second example is much more serious and will likely result in having to correct the models for the MPC to effectively control the process. Figure 7 shows a similar potential problem encountered when making large steps to furnaces. The furnace on the left illustrates the effect on the furnace outlet temperature when a large, sudden feed move is made; the furnace on the right, when the same size feed move is made gradually.

9 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 9 Figure 7. Effect of large steps on a furnace The curve on the left furnace shows the outlet temperature sagging more than the curve on the right. The result of erroneous predictions of the process behavior when the MPC is online is similar here to that described above. Again it may lead to having to correct the models for the MPC to effectively control the process. Figure 8 illustrates a potential problem encountered when making large steps to reactors. The reactor on the left shows the severity when a large temperature move is made at once; the reactor on the right when the MPC makes the move gradually. The curve on the left shows the severity overshooting and moving back down, while using the MPC results in no overshoot. Again, the results of inaccurate predictions when the MPC is online are similar to those already described, and could mean having to correct the models. Figure 8. Effect of large steps on a reactor Finally, Figure 9 shows a potential problem in making large steps to setpoints. The left shows the valve output when a large setpoint move is made all at once; the right when it is made gradually. The left valve s output overshoots and falls back, while the right suffers no overshoot. Here again, models may need correcting for the MPC to effectively control the process.

10 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 10 Figure 9. Effect of large steps on a valve In summary, larger steps make it easier to ID models, but the models do not represent the process well. The above illustrations show that the large steps used in the past tended to excite the process in a manner that is inconsistent with the impact of smaller changes made by the online MPC. Once this was discovered during commission, model massaging or re-stepping proved necessary to rectify the problem. This prolonged commissioning time or required repeated step testing. This manual method typically led to stepping for two to six months or more (depending on the size of the application) to get enough for a base set of models. Model validation and correcting erroneous models added another few weeks or months to the project. The need to further modify models during commissioning added another two to six months. Projects typically took six months from start to finish for small applications (5-6MVs) and up to two years for larger applications (25-30MVs). The tasks of stepping the process, identifying models, validating models and commissioning took up to 75% of the overall project time. Open Loop Sequential Automated Stepping (1990) By 1990 MPC applications were more accepted by the process industry and interest in implementing them was growing. The need for step testing became more widely accepted. Attention turned to reducing the time it took to implement MPC applications. Faster computers enabled implementation of more sophisticated identification algorithms to handle larger data sets and data that had steps of an MV commingled with scattered steps of other MVs and DVs. The importance of steps that spanned the power spectrum to obtain better quality models was also recognized in other words, a mix of steps consisting of varying switching times (steps held for short, medium and long periods of time). The final development was the move to restrict human involvement in the step testing procedure as far as possible to reduce the effect of correlated moves. Instead, software could automatically step the process. The general method deployed follows: 1. The size of steps is negotiated with the operator based on their tolerance for disruption to the process. 2. This is input, together with an estimate of the process settling time, to an algorithm that designs the desired step pattern to span the entire power spectrum. Two types of step patterns were available a pseudo random binary sequence (PRBS) and a Schroeder-Phased sinusoid. 3. Transfer the designed step pattern to an online software module to conduct the steps.

11 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times Set up the PIDs such that each MV corresponds to the SP of a PID or the OP of a PID (or a hand valve) in other words, make the underlying control system open loop. This required breaking certain cascade structures or putting a PID in manual. 5. Allow the process to settle out so it is not moving much. 6. Make steps to a PID.SP or PID.OP one MV at a time using the designed step pattern. 7. Once stepping commences, the operator is still responsible for ensuring CVs do not violate key operational constraints. The operator is instructed to make compensating changes to other MVs to prevent constraint violations. 8. Collect data off an historian and take off-line to run through the ID algorithm. 9. Massage the data as necessary to remove bad data and rerun through the ID algorithm. 10. Repeat 8-9 until acceptable models are obtained. 11. Once acceptable models are obtained, commence stepping on another MV. 12. Repeat 2-11 until acceptable models are obtained for all the CVs and MVs. 13. Repeat 2-11 until acceptable models are obtained for all the CVs and DVs. 14. Validate all models using engineering knowledge and judgment. Models that do not make sense are either removed or stepping is conducted again to obtain a more realistic model. 15. Build the MPC application and run in prediction mode to do the final validation of model quality. 16. With MPC online, step the MVs one at a time to evaluate the behavior of the predicted CVs against their respective process CV. 17. Massage data or re-step the process until the predictions follow the process closely enough for online control. 18. Begin commissioning by forcing key CVs to violate their constraints and evaluating how effectively the controller brings them back. If the response is insufficient and tuning cannot rectify the issues, data massaging or stepping the process is again necessary. 19. Configure the optimizer and turn on. This typically caused the operation to slowly move from one operating state to another. 20. Evaluate the performance of the optimizer and further massage data or step the process again as necessary. Figure 10 illustrates a typical pseudo random binary sequence (PRBS), a Schroeder Phased sinusoid, and a filtered PRBS. All three step patterns are designed with switching intervals that cover the entire power spectrum for the settling time provided. To design the step pattern of choice, the engineer entered the desired step size and an approximate settling time for the CVs associated with each MV into a software program. The program then generated the step pattern that was deployed via an online running application. The PRBS was typically used unless large steps to the process were deemed necessary. The user would then choose either the Schroeder Phased sinusoid or a filtered PRBS step pattern to avoid the effect of implementing large steps described in the section on manual step testing, above. Despite the computer implementing most of the steps, an operator is still needed to make compensated moves to keep CVs from violating constraints. The structure of the automated step pattern helps reduce the incidence of humans making correlated moves, but they still need to be removed from data before the ID algorithm is run. The benefit is there are many more steps made and at different switch intervals so the proportion of correlated compensated moves is significantly reduced, and many times the ID algorithm could still get quality models with the correlated moves left in the data.

12 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 12 PRBS Schroeder Phased Filtered PRBS Figure 10. Open loop sequential step patterns Figure 11 illustrates a PRBS pattern and the same compensated MV moves illustrated in the section on manual step testing (above); a comparison of the process movement with unmeasured disturbances and drift (middle blue) and without (middle green); and the predicted results of the models obtained from both process movements (bottom). Figure 11. Effect of compensated moves on model ID results using a PRBS pattern In the bottom graph, the green is the process movement without UMD or drift; the blue represents the prediction from the model ID performed on the data without UMD or drift; and the red represents the prediction from the model ID performed on the data with UMD and drift. Both the blue and red predictions follow the process well and there is little difference between them. The gains and time constants are essentially the same. This example shows that compensated moves can be made with very little effect on the model ID results.

13 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 13 The next issue to address is the effect of UMD and drift on the model ID results. Figure 12 illustrates the result of UMD and drift on step patterns that cover the power spectrum. Figure 12. Effect of UMD and drift on model ID results using a PRBS pattern The top graph illustrates the process movement without UMD or drift (green), the process movement with UMD and drift (blue), and the resultant prediction from the model ID using the process movement with UMD and drift (red). The duration of the stepping was 655 minutes, compared to 1320 for the manual method illustrated in Figure 5. The bottom graph shows the process movement without UMD or drift (green), the prediction from the model ID using process movement without UMD and drift (blue), and the prediction from the model ID using process movement with UMD and drift (red). In the top graph, the prediction looks like it fits the process without UMD and drift. This is verified in the bottom graph when it is compared to the prediction from the process without UMD and drift. The model producing the red prediction differs in gain to the model that produced the blue by 10% and the settling times are essentially the same. The graph shows issues from making long steps are essentially rectified by using a step pattern that covered the power spectrum. With this step pattern, the model ID algorithms are able to get quality models without having to remove much of this naturally occurring phenomenon over the duration of the step testing period. The mix of short, medium and long steps not only reduces the effect of unmeasured disturbances and drift on the step test data, but the model is obtained with half the step testing time of the manual stepping method. Making steps that are too large is still an issue. Even though the step pattern is automated, there is still a desire to make big enough steps to see the movement in the CVs. To rectify this, the engineer is provided with choices in the type of steps that can be designed. He can use a standard PRBS signal to make conventional steps if the process is well behaved or does not require large steps to get the CVs to move. If steps have to be large, the engineer implements the standard PRBS but filters the step change in over time, or designs a Schroeder-Phased sinusoid as the step pattern. As illustrated in Figure 10, the Schroeder-Phased sinusoid is an oscillating pattern that does not make large changes to the process all at once. The downside to filtering the PRBS or using the Schroeder is that the engineer must trust that the ID algorithm will get a quality model from the data as it is difficult to detect sufficient CV movement in the data with the naked eye. Over time with this method engineers found that quality models could still be obtained and began to rely on these methods more often. Despite the benefits automated step patterns brought, new issues arose. First, to get quality models, the engineer had to provide a good approximation of the process settling time, but to reduce the step testing time, engineers tended to underestimate this. The result

14 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 14 was step patterns with switching intervals that were too close together. Figure 13 illustrates the difference between step patterns with appropriate settling times and those with settling times that are too short. Figure 13. Comparison of sequential step patterns from appropriate and short settling times The difference between the sequential step pattern resulting from an appropriate settling time (top) and from a short settling time (bottom) can be seen. Switching intervals in the top patterns are much further apart than those in the bottom. Figure 14 shows the effect on the prediction of using a shorter step pattern than appropriate in the presence of UMD and drift (top), and in the presence of compensated moves, UMD and drift (bottom). The top graph illustrates the process movement without UMD or drift (green), the prediction from the model ID using an appropriate settling time (blue), and the prediction from the model ID using a short settling time (red). The models differ in both gain and settling time, with the model producing the red prediction having a 40% smaller gain and 400% longer settling time than the model producing the blue prediction. The bottom graph illustrates the process movement without UMD or drift but in the presence of compensated MV moves (green); the prediction from the model ID using an appropriate settling time (blue); and the prediction from the model ID using a short settling time (red). These models also differ in both gain and settling time. The model that produced the red prediction is 35% smaller in gain and has a 50% longer settling time than the model that produced the blue. The model differences are the result of step patterns that did not cover the true power spectrum but only the higher frequency parts of it. This leads to erroneous models in both gain and settling time. To make matters worse, with short switching intervals it is nearly impossible for the engineer to evaluate whether the steps are generating data good enough for the ID algorithm to obtain quality models. Under these circumstances, the engineer has no choice but to redo the step testing exercise.

15 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 15 Figure 14. Effect on the model ID of a PRBS step pattern from short settling times Using automated step patterns did, however, reduce the duration of MPC projects compared to manual stepping, cutting the time taken to step a process in half. Model validation and correcting erroneous models was also twice as quick. The need to further modify models during commissioning was cut by about 25%. Projects now typically took up to four months from start to finish for small applications and up to 18 months for larger applications. Time spent stepping the process, identifying models, validating models and commissioning was reduced to 67% of the overall project time. Open Loop Simultaneous Automated Stepping (2000) The quest to reduce implementation time continued with simultaneous, rather than sequential, automated stepping. Stepping two or more MVs at the same time could reduce the overall time to step the process, while computers continued to get faster and ID algorithms continued to improve. The general method deployed follows: 1. Negotiate the size of steps with the operator based on the disruption to the process they are comfortable with. This method, however, allows larger steps to be made without causing severe constraint violations due to the inherent filtering resulting from stepping multiple MVs in an uncorrelated fashion. 2. Input the size of steps, an estimated settling time, and the group of MVs to be stepped simultaneously to an algorithm, which designs the desired step pattern. These patterns not only cover the complete power spectrum but are also uncorrelated. Only the PRBS step pattern was available since it is impossible to create uncorrelated Schroeder-Phased sinusoids. 3. Transfer the step patterns to an online software module to conduct the steps. 4. Set up the PIDs so that each MV corresponds to the SP of a PID or the OP of a PID (or hand valve). In other words, make the underlying control system open loop. This required breaking certain cascade structures or putting a PID in manual. 5. Allow the process to settle out. 6. Make steps to the group of MVs in accordance to their uncorrelated step patterns. 7. Once stepping commences, the operator remains responsible for ensuring CVs do not violate key operational constraints. The operator is instructed to make compensating changes to other MVs to prevent constraint violations. 8. Collect data from an historian and take off-line to run through the ID algorithm.

16 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times Massage the data to remove bad data and rerun through the ID algorithm. 10. Repeat 8-9 until acceptable models are obtained for the MV group. 11. Once acceptable models are obtained, commence stepping on another group of MVs. 12. Repeat 1-11 until acceptable models are obtained for all the CVs and MVs. 13. Repeat 1-11 until acceptable models are obtained for all the CVs and DVs. 14. Validate all models using engineering knowledge and judgment. Models that do not make sense are either removed or stepping is conducted again to obtain a more realistic model. 15. Build the MPC application and run in prediction mode to do the final validation of model quality. 16. With MPC online, step the MVs one at a time to evaluate the behavior of the predicted CVs against their respective process CV. 17. Massage data or re-step the process until the predictions follow the process closely enough for online control. 18. Begin commissioning by forcing key CVs to violate their constraints and evaluating how effectively the controller brings them back. If the response is insufficient and tuning cannot rectify the issues, data massaging or stepping the process should be done again. 19. Configure the optimizer and turn on. This typically caused the operation to slowly move from one operating state to another. 20. Evaluate the performance of the optimizer and further massage models or step the process again as necessary. Figure 15 illustrates a PRBS pattern for MV1 (green), a PRBS pattern for MV2 (brown), and the resultant process movement without UMD or drift (blue). The simultaneous step patterns are designed to be uncorrelated and with switching intervals that cover the entire power spectrum for the settling time provided. To design the step pattern of choice, the engineer entered the desired step size and an approximate settling time for the CVs associated with each MV into a software program. The program then generated the step patterns deployed via an online running application. The PRBS was typically used unless large steps to the process were deemed necessary. The user could choose a filtered PRBS step pattern to avoid the effect of implementing large steps as described in the manual step testing section. The Schroeder Phased sinusoid is no longer available due to the difficulty of creating uncorrelated patterns. Figure 15. Open loop simultaneous PRBS step test patterns

17 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 17 Despite a computer implementing most of the steps, the operator must still make compensated moves to keep CVs from violating constraints. The structure of the simultaneous step patterns essentially eliminates the risk of people making correlated moves. Removing correlated steps from the data before the ID algorithm is run becomes less of an issue. The effect of UMD and drift on the step test data is also much less of an issue. ID algorithms are able to obtain quality models without having to remove this naturally occurring phenomenon over the duration of the step testing period. Figure 16 illustrates the result of UMD and drift on simultaneous uncorrelated step patterns that cover the power spectrum. Figure 16. Effect of UMD and drift on model ID results using simultaneous PRBS patterns The top graph illustrates the process movement without UMD or drift (green), the process movement with UMD and drift (blue) and the resultant prediction from the model ID using the process movement with UMD and drift (red). The duration of the stepping was 550 minutes, compared to 1320 for the manual method (Figure 5) and 655 for the automated sequential method (Figure 12). The bottom graph illustrates the process movement without UMD or drift (green), the prediction from the model ID using process movement without UMD and drift (blue), and the prediction from the model ID using process movement with UMD and drift (red). In the top graph, the prediction looks like it fits the process without UMD and drift well. This is verified in the bottom graph, when it is compared to the prediction from the process without UMD and drift. The MV1 and MV2 models that produced the red prediction have the same gains as the models that produced the blue one. The red and blue model for MV2 have the same settling time, and the red model for MV1 differs in settling time from the blue MV1 by 20%. The sensitivity to UMD and drift resulting from long steps is essentially removed by using simultaneous uncorrelated step patterns that cover the power spectrum. With these simultaneous step patterns, the model ID algorithms are able to get quality models without removing much of this naturally occurring phenomenon over the step testing period. The models are also obtained with only 75% of the step testing time required for the sequential stepping method. Making steps that are too large is still possible but is less of an issue. Stepping multiple MVs at once tends to create a built-in filtering of the CV movement, dampening the ability of larger steps to throw the process out of equilibrium. Of course, the size of the steps still needs to be selected prudently to avoid upsetting the process operation too much. The option to filter the steps is still available in the event that steps need to be large. The downside of filtering the PRBS signals too much also still exists. While an additional reduction in implementation time is realized with this method, the need to specify an accurate settling time becomes crucial. Failing to specify an appropriate settling time leads to very low quality data that is essentially useless.

18 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 18 Figure 17. Comparison of simultaneous step patterns from appropriate and short settling times Figure 17 illustrates the difference between step patterns with appropriate settling times (top) and short settling times (bottom). The switching intervals for the former are much further apart than those for the latter. Figure 18. Effect on the model ID of simultaneous PRBS step patterns from short settling times In figure 18 the top graph illustrates the process movement without UMD or drift (green), the process movement with UMD and drift (blue), and the prediction from the model ID using short settling times (red). The bottom graph illustrates the process movement without UMD or drift (green), the prediction from the model ID using appropriate settling times (blue), and the prediction from the model ID using short settling times (red). It is clear from the top graph that the model ID results generated bogus models with short simultaneous step patterns. The bottom graph, which compares the predictions using models with short simultaneous step patterns and predictions using appropriate patterns, shows just how far off the models are. Short step patterns when stepping a single MV at a time produces low model quality results.

19 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 19 Using short step patterns when stepping multiple MVs simultaneously produces garbage. It is also impossible for the engineer to massage this data into something useful. If quality models are not acquired, the engineer has little choice but to redo the step testing. Overall properly implementing simultaneous automated step patterns further reduced the duration of MPC projects. The time to step a process was cut by another 25%; the model validation and correcting of erroneous models was cut by a further 50%; the need to further modify models during commissioning was cut by approximately another 25%. Projects now typically took up to 3 months from start to completion for small applications and up to 14 months for larger applications. The tasks of stepping the process, identifying models, validating models and commissioning were reduced to 60% of the overall project time. Open Loop Automated Stepping and Auto ID (2005) The next evolution to reduce implementation time was to fully automate the step testing procedure. The key drivers were the desire to remove human intervention from the procedure and reduce the time and skill level required to implement step testing. This was accomplished by creating stepping software that could step the process, collect data, identify models and validate them into a single methodology. Computers were now fast enough to allow very sophisticated ID algorithms to be developed and run in real-time, and to allow stepping the process, collecting data and running the identification algorithm online to all be automated. It was no longer necessary to collect data and transport it to an off-line software package to run the ID algorithm. Automatic data massaging algorithms are implemented to look for certain types of bad data and remove it. The ID algorithm automatically estimates the quality of models being identified and reports this to the engineer. The stepping software supports open loop sequential and open loop multivariable stepping procedures. The majority of the off-line model validation is incorporated as part of the procedure. The general method deployed follows: 1. Negotiate the size of steps with the operator based on how much disruption to the process they are comfortable with. 2. Configure the stepping software to communicate with the underlying PID.SPs or PID.OPs for each MV and the process variables associated with each CV and DV. 3. Build the stepping application to run in real-time at a specified interval. 4. Determine whether sequential stepping or simultaneous stepping is to be conducted. 5. For sequential stepping, configure the stepping application with the size of steps and an estimate of the process settling time for each MV. 6. For simultaneous stepping, configure the stepping application with the step sizes, an estimate of the process settling time and the first group of MVs to be stepped simultaneously. For this type of stepping, the step sizes can be larger, as discussed in the previous section. 7. Specify how often the ID algorithm executes. 8. Specify all the sub models (CV-MV and CV-DV) to be identified that corresponded to the MV or group of MVs to be stepped. 9. Set up the PIDs such that each MV corresponds to the SP of a PID or the OP of a PID (or hand valve) in other words make the underlying control system open loop. This requires breaking certain cascade structures or putting a PID in manual. 10. Allow the process to settle. 11. Enable the stepping software to begin data collection. 12. Enable the stepping software to begin making steps either sequentially or to the first group of MVs. The stepping software automatically designs the sequential step patterns for each MV or designs the uncorrelated step patterns for the group of MVs to be stepped. 13. Once stepping commences, the operator is still responsible for ensuring CVs do not violate key operational constraints. However, the stepping software enables the engineer to make compensating steps to any MV at any time and in any direction to address CV constraint violations. 14. The stepping software automatically triggers the ID algorithm to execute at the specified interval.

20 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times After each ID execution, the engineer evaluates the quality of the models obtained by comparing model predictions to the process movement and utilizing model quality information reported for each sub model by the ID algorithm. 16. Massage the data to remove remaining bad data and rerun through the ID algorithm or wait for the next automatic ID execution. The automated data removal algorithm is designed to look for certain types of data and automatically remove them, but other bad data may remain. 17. Observe the progress of the model ID until models of sufficient quality are obtained. 18. Move on to the next sequential MV or to the next group of MVs to be stepped. 19. Repeat until acceptable sub models are obtained for all the CVs and MVs. 20. Repeat until acceptable sub models are obtained for all the CVs and DVs. 21. Validate all models using engineering knowledge and judgment. The majority of this exercise is done during the stepping procedure. 22. Build the MPC application and run in prediction mode for the final validation of model quality. 23. With MPC online, step the MVs one at a time to evaluate the behavior of the predicted CVs against their respective process CV. 24. Massage data or re-step the process until the predictions follow the process closely enough for online control. This requirement is significantly reduced. 25. Begin commissioning by forcing key CVs to violate their constraints and evaluating how effectively the controller brings them back. If the response is insufficient and tuning cannot rectify the issues, data massaging or stepping the process is required again. This requirement is significantly reduced. 26. Configure the optimizer and turn on. This typically causes the operation to slowly move from one operating state to another. 27. Evaluate the performance of the optimizer and further massage models or step the process as necessary. This requirement is significantly reduced. Figure 19 shows the automated stepping software. The red square highlights where the engineer enters the size of steps and settling time for each MV to be stepped. The red circles highlight where the engineer specifies how often the ID algorithm runs (right most), the button to start the data collection (left most), and the button to start the stepping (middle). Figure 19. Automated stepping software

21 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 21 Figure 20 illustrates the allowable step patterns as either sequential or simultaneous. There are no restrictions to how many MVs can be specified. For sequential stepping, the engineer tells the software when an MV is done and it automatically moves onto the next one. For simultaneous stepping, the engineer specifies all the MVs or a group of MVs to be stepped. When a group of MVs is done, the user turns those off and configures the next group. Figure 20. Automated stepping software allowable step patterns In Figure 21, the configuration page for specifying the size of steps and settling time for each MV (top) is shown. This is where sequential stepping of MVs (not shown) or simultaneous stepping is specified. All MVs can be stepped simultaneously or in specified groups. Also illustrated is the configuration page for specifying which sub models are to be identified (bottom). Here the intended matrix of models for the MPC is specified. The ID algorithm only attempts to obtain models for sub models not labeled as NULL. Figure 21. Configuring the stepping software

22 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 22 Figure 22. Collect data, step the process and monitor ID results Figure 22 illustrates the data collection and model ID interface (top) and the model ID progress (bottom). Once the data collector is started, process data is automatically collected and inserted into the model ID software. The historical data is shown to the left of the horizontal line in the top chart. The future step patterns are shown to the right of the horizontal line. The horizontal line shifts to the right each time the stepping software executes. Every time the model ID algorithm executes, the results and model quality data are updated in the Model Highlights page of the stepping software, illustrated in the bottom chart. The engineer is provided with a model rank (1-5), the gain, the settling time and the deadtime of each sub model. The rank is color coded to display good quality models (green), medium quality models (yellow), and poor quality models (red). Model prediction for all CVs (top) and the lock model function (bottom) are shown in Figure 23. Each time the model ID algorithm executes, the current models are stored in the model ID software. The engineer can then generate plots of how well the current models predict the process movement for each CV, as illustrated in the top chart. The green curves represent the process movement and the red the prediction based on the current models. Figure 23. Evaluate results and lock models

23 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 23 As can be seen, the plots for most of the CVs have red curves superimposed on the green curves. This is an indication to the user that the models for those CVs are high quality and there is no need to identify better models. In two of the plots, there is a distinguishable difference between the red and green, indicating the need for further stepping to acquire better models. In the bottom chart a dialogue box has popped up on the Model Highlight page. This alerts the engineer to all the sub models that have a high rank and have been essentially the same for the past few ID executions. It recommends that these can be locked. Using the plots in the top chart, the engineer can assess if the recommendation to lock the model is valid. As illustrated on the Model Highlight page, the user can select the sub models to be locked. The model ID software will no longer attempt to get better models for these and will use them for predictions on subsequent model ID executions in the event that not all sub models are locked for a CV. If all models for a CV are locked, the ID algorithm will stop attempting to ID better models. The engineer can lock and unlock models at anytime. Figure 24 shows the display for altering the MV stepping configuration (top) and when stepping is complete (bottom). Once all the models for an MV being stepped are locked, it is no longer necessary to step that MV. From the configuration page of the stepping software, the engineer has multiple options for how to treat an MV being stepped. They can hold or resume the MV, or step the MV up or down at will. Here the engineer can make compensating moves to any MV to address CV constraint violations, relieving the operator from the task. The engineer can also put an MV on hold indefinitely or for a period. If all sub models for that MV are locked, there is no reason to continue stepping it. If the engineer deems there is too much interference between MVs being stepped simultaneously, one or more MVs can be put on hold to resume later. In the bottom chart, the engineer has assessed that all models are good and has locked them. The testing procedure is then stopped and the engineer can move onto final model validation and commissioning. Figure 24. Changing the MVs being stepped and stopping when done Since this method utilizes both sequential and simultaneous step patterns, all the benefits described in the Open Loop Sequential and Open Loop Simultaneous sections are realized. These include reduced sensitivity to UMD, drift, and correlated compensating moves, and elimination of the need to make excessively large moves. An additional advantage with this method is that it essentially removes the need for the operator to make compensating moves to keep CVs from violating constraints. The engineer can make compensating changes as necessary directly from the stepping software as illustrated in the top display of Figure 24. The creation of software fully automating the step testing tasks of configuring step patterns, data collection, stepping the process, model identification, and model validation radically changed the way step testing was implemented:

24 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 24 The ability to observe ID results on the fly, along with the ease of changing step sizes, allowed the engineer to moderate the step size early to ensure quality data was generated. Users could filter and moderate the amount of filtering in the event that the steps had to be excessively large. Certain types of bad data were detected and automatically removed, reducing the need for the engineer to track and remove them. The development of more sophisticated ID algorithms reduced the need to massage data to get quality models. If bad data was seen to be affecting the ID results, the engineer could easily remove it instead of going through an off-line iterative procedure of massaging the data and running the ID again. Since the engineer had the ability to pause MVs in a group to validate steps on other MVs, the concern of getting useless data from stepping multiple MVs at the same time was reduced. The improved model ID algorithms obtained much better results with multivariable data than previous algorithms. The need to repeat the step testing procedure was essentially eliminated. The engineer knows when quality models are obtained for all sub models, and this is the indication that step testing can cease. The method cut implementation times and reduced the experience required. One major issue remained: Not specifying an appropriate settling time of the process could still lead to very low quality data that is essentially useless. However, even this was partly addressed since the engineer can observe the ID results on the fly and change the settling time online. The worst case now was that some step data was useless and the stepping procedure prolonged. Overall the time to step a process was cut by another 25%; model validation and correcting erroneous models, after stepping and model ID were complete, was no longer necessary; and further modifying models during commissioning was cut by an additional 50%. Projects now typically took 2 months from start to finish for small applications and up to 10 months for larger applications. The tasks of stepping the process, identifying models, validating models and commissioning were reduced to as little as 40% of the overall project time. Closed Loop Automated Stepping with MPC and Auto ID (2005) The new software to fully automate the open loop step testing procedure was also implemented to allow automated closed loop stepping. A multitude of implemented MPC applications prompted efforts to make it easier to maintain the applications, especially when it came to updating models that were no longer predicting the process well. Users wanted to be able to step test without having to turn off the online running MPC application. Faster computers along with very sophisticated ID algorithms that could effectively deal with closed loop data meant they could. All the automated procedures discussed in the open loop section were still available. In addition, the stepping software coordinated step changes to the MVs with changes required by the MPC application to keep CVs from violating limits. Off-line validation of models was unnecessary as the engineer could compare the prediction capability of the new models against previous ones. The general method is similar to that discussed in the previous section with a few differences: 1. Determine the starting size of steps and settling time of the process based on experience from the existing operation of the MPC application. 2. Configure the stepping software to communicate with the MPC application. This automatically connects the stepping software to all CVs, MVs and DVs in the MPC application. 3. Build the stepping application to run in real-time at a specified interval. 4. Specify whether to step a single MV or simultaneous step a group of MVs. 5. Input the starting size of steps and an estimate of the process settling time for each MV to be stepped. 6. Specify how often the ID algorithm executes. 7. Specify the target sub models (CV-MV and CV-DV) that needed to be identified for the MV or group of MVs being stepped. Unlike open loop stepping where all sub models need to be identified, for closed-loop stepping some sub models may still be of sufficient quality, so only a portion may need to be identified (the target sub models). 8. Enable the stepping software to begin data collection.

25 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times Enable the stepping software to begin making steps. The stepping software automatically designs the step patterns for a single MV or designs the uncorrelated step patterns for a group of MVs to be stepped. 10. Once stepping commences, the operator still has overall responsible for ensuring CVs do not violate key operational constraints. However, since the MPC application is running online, it makes changes as necessary to address CV violations. If deemed necessary, the engineer can still make compensating steps to any MV at any time and in any direction to help the MPC application address CV constraint violations. 11. The stepping software automatically triggers the ID algorithm to execute at the specified interval. 12. After each ID execution, the engineer evaluates the quality of the models obtained by comparing model predictions to the process movement and utilizing model quality information reported back for each sub model by the ID algorithm. 13. Massage the data as necessary to remove any remaining bad data and rerun through the ID algorithm or wait for the next automatic ID execution. The automated data removal algorithm looks for certain types of bad data and automatically removes them, but other bad data may remain. 14. Observe the progress of the model ID until models of sufficient quality are obtained. 15. Move on to the next MV or the next group of MVs to be stepped. 16. Repeat 9-15 until acceptable sub models are obtained for all the CVs and MVs. 17. Repeat 9-15 until acceptable sub models are obtained for all the CVs and DVs. 18. Validate all models using engineering knowledge and judgment. The majority of this exercise is done during the stepping procedure. 19. Update the MPC application with new models. 20. Re-commission the MPC application with new models. This is straight forward as the engineer has previous models to compare against. The procedure to connect the stepping software to an MPC application is shown in Figure 25. The engineer first enters the name and description of the stepping application to be built (upper most red circle); then browses for an existing running MPC application (middle red circle); and selects the target MPC application (bottom red circle). The stepping application is now built and ready for configuration. Figure 25. Connecting stepping software to an MPC application When using the step testing software for open loop stepping, the engineer needs to directly connect each CV, MV and DV to its respective process variable. With closed loop stepping, this is no longer necessary. Instead, the engineer connects the stepping

26 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 26 software directly to the MPC application that requires model maintenance, and uses the connections to the process already established by the MPC application. From this point the engineer needs to configure the stepping software as illustrated in Figure 21, and follow the rest of the procedure for open loop stepping. Figure 26. The stepping application and MPC application are connected Figure 26 shows the user interface for the MPC application (top), and the user interface for the stepping application (bottom). When stepping an MPC application, a Stepping button appears on the MPC application user interface. This button allows the operator to pause stepping if, for example, it is causing problems with controlling the process. Once the problems are addressed, the operator pushes the button again to resume stepping. Figure 27. Moderating interaction between steps and MPC moves

27 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 27 Figure 27 illustrates the interactions between the stepping application making steps, and the MPC application making compensating moves. Under normal operating conditions, the MPC application is configured to push the process aggressively towards its optimal operating state. In the figure, under the Optimizer Normal window, the steps to MV1 and MV2 look more like pulses than conventional steps. This is due to the optimizer making compensating changes to MV 1 and MV2 to push CV1 and CV2 back to their optimal operating states (165 low limit for CV1 and 9.5 high limit for CV2). The resulting data is low quality and little use for obtaining new models as it resembles short settling time step test patterns discussed above. To rectify this problem, it is necessary to moderate the aggressiveness of the optimizer to yield better quality step test data and more accurate models. The Optimizer Slowed data shows the optimizer is less aggressive. CV1 and CV2 stay close to their optimal limits but are allowed to drift away further than under Optimizer Normal operation. The steps for MV1 and MV2 are now elongated pulses, and are better than before, but still might yield inaccurate models. The Optimizer Slower data shows the optimizer moderated even more. The MV1 and MV2 moves are further elongated and approaching square. ID results from this data should yield accurate models. Finally, with the optimizer off and no longer pushing the operation to its optimal operating state, the steps are square and the movement of the CVs, even though the controller is running, closely resemble that of open loop. The data quality is as good as it gets. The downside is the cost of operation is higher. With closed loop stepping, the engineer must decide the proper balance between operating close to optimal conditions and obtaining high enough quality step test data to generate accurate model ID results. The more aggressively the optimizer is allowed to make compensating moves to push the operation back to optimal, the longer the step test period will have to be. Moderating the trade-off simply requires adjusting a single tuning parameter. The engineer can adjust this parameter as necessary to achieve the desired level of interaction. The closed loop method built on the benefits of the automated open loop with ID method described above. The need for the operator to address CV constraint violation is eliminated unless the MPC has to be turned off due to severe process upsets; the option to filter and moderate the amount of filtering is still available; and so is the ability for automatic and manual engineering removal of bad data on the fly. Validating how well the new models predicted process behavior compared to the previous models is easily done; and it is no longer necessary to turn off the MPC controller to re-step the MVs. The interaction between the steps made to the MVs and the MPC application s compensation can be moderated to give precedence to obtaining better quality step data. This method provided additional options for implementing step testing. Revising models that were no longer predicting the process well could be done without having to conduct open loop steps from scratch. Allowing the MPC application to stay online meant key economic operating variables could continue to operate close to optimal levels while MVs were being re-stepped. Updating the MPC with the new models required minimal, if any, effort in terms of retuning the MPC. The new closed loop method reduced the time to acquire and re-commission new models by up to 75% compared to turning the MPC application off and re-conducting open loop stepping. It also opened the door for the next big evolution in step testing methodologies. Enhanced Closed Loop Automated Stepping with MPC and Auto ID (Since 2010) This next step was to reduce implementation time by extending the capability of the closed loop step testing methodology to open loop stepping. This would eliminate the need for open loop stepping and model validation, and reduce commissioning time. The major problem with open loop stepping is that it requires moving the process away from its most economic operating condition and away from key operating constraints to allow the process to swing, within operating constraints, while steps are implemented. For example, it involves reducing the charge rate to the process or reducing the load on operating equipment that typically operates close to full capacity. There are also processes that do not behave well under open loop operation resulting in a reluctance to conduct any form of open loop step testing on them. Finally, open loop stepping requires close monitoring of the process operation to ensure MV steps do not lead to unacceptable constraint violations. This is particularly true if ramp CVs are included in the MPC application design, which typically requires somebody to be present while stepping is taking place. This means either putting in place 24 hour coverage or stopping the step testing when engineers leave and restarting when they return.

28 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 28 Extending the ability to step the MVs with an MPC application running online to open loop stepping requires a seed MPC application. There must be at least a minimum set of models to build this seed application. To address any CV constraint violation, every CV must have at least one seed model with respect to an MV. The seed MPC will initially be configured to just push CVs back within their limits upon a limit violation. This reduces the need for highly accurate models at the start. It is possible to update gains on any sub model at any time in the seed MPC as well as download new models from the stepping software into the seed MPC without turning the seed MPC off. This enables the MPC model structure to be built on the fly as the closed loop step testing method previously described is worked through. The general method is similar to that discussed in the previous section with a few differences: 1. Design the basis of the seed MPC by defining the CVs, MVs and DVs from the process. 2. Create an expectation matrix that identifies all the required sub models. 3. For each CV, select one or more sub models to include in the seed MPC matrix. 4. Obtain model parameters for each seed model via historical data or engineering/operator judgment. 5. Build the seed MPC application using the seed models. 6. Implement the seed MPC application online and specify the high and low limits for the CVs and MVs, as well as other relevant parameters. 7. Validate the seed models by ensuring the gain sign of each model is correct and trends with its respective CV. The seed models do not have to be highly accurate, just trend in the direction of their respective CVs. 8. Determine the starting size of steps and settling time of the process for each MV. 9. Configure the stepping software to communicate with the seed MPC application. This automatically connects the stepping software to all CVs, MVs and DVs in the MPC application. 10. Build the stepping application to run in real-time at a specified interval. 11. Specify whether to step a single MV or a group of MVs simultaneously. 12. Input the starting size of steps and an estimate of the process settling time for each MV to be stepped. 13. Specify how often the ID algorithm executes. 14. Specify all the sub models (CV-MV and CV-DV) to be identified. 15. Enable the stepping software to begin data collection. 16. Enable the stepping software to begin making steps. The stepping software automatically designs the step patterns for a single MV or designs the uncorrelated step patterns for a group of MVs. 17. Once stepping commences, the operator still has overall responsible for ensuring CVs do not violate key operational constraints. However, since the seed MPC application is running online, it makes changes as necessary to address CV violations. If deemed necessary, the engineer can make compensating steps to any MV at any time and in any direction to help the seed MPC application address CV constraint violations. 18. The stepping software automatically triggers the ID algorithm to execute at the specified interval. 19. After each ID execution, the engineer evaluates the quality of the models obtained by comparing model predictions to the process movement and utilizing model quality information reported back for each sub model by the ID algorithm. The engineer also evaluates the signal to noise ratio for each sub model to ensure step sizes are large enough and settling times not too short. 20. Massage the data as necessary to remove remaining bad data and rerun through the ID algorithm or wait for the next automatic ID execution. The automated data removal algorithm looks for certain types of bad data and automatically removes them, but others may remain. 21. Observe the progress of the model ID until models of sufficient quality are obtained. 22. Download newly obtained models directly to the seed MPC application for better CV constraint management. (This task is optional.) 23. Once models are downloaded for key CVs, configure the optimizer to run the process closer to optimal operating conditions. Again, this is optional.

29 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times Move on to the next MV or group to be stepped. 25. Repeat until acceptable sub models are obtained for all the CVs and MVs. 26. Repeat until acceptable sub models are obtained for all the CVs and DVs. 27. Validate all models using engineering knowledge and judgment. The majority of this exercise is done during the stepping procedure. 28. Run the seed MPC application in prediction mode for the final validation of model quality. 29. With the seed MPC online, step the MVs one at a time to evaluate the behavior of the predicted CVs against their respective process CV. 30. Massage data or re-step the process until the predictions follow the process closely enough for online control. This requirement is significantly reduced. 31. Begin commissioning by forcing key CVs to violate their constraints and evaluating how effectively the controller brings them back. If the response is insufficient and tuning cannot rectify the issues, data massaging or stepping the process is again necessary. This requirement is significantly reduced. 32. Configure the optimizer and turn on. This typically causes the operation to slowly move from one operating state to another. 33. Evaluate the performance of the optimizer and further massage data or step the process as necessary. This requirement is significantly reduced. Figure 28 illustrates the CV and MV configuration for a seed MPC application. The application is configured with all the CVs and MVs, along with high and low limits for all variables. The engineer may also need to enter values for tuning and other important parameters. The operator interfaces with the application in the same way as other applications. Figure 28. Configuring the seed MPC application The beginning model matrix for the seed MPC application is shown in Figure 29. It is sparse, with the minimum for a seed MPC application for most CVs just one sub model for each CV. The engineer can specify more than one sub model if desired as is done for CV3 (it has 2). These seed models are obtained from historical data or engineering or operator experience. The models do not have to be very accurate; as long as the gain sign is correct, the seed MPC will make proper compensating moves to address CV constrain violations. If the compensating moves fail to properly address a CV constraint violation, the engineer can increase the gain multiplier to

30 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 30 slow down the response or reduce the gain multiplier to accelerate the response. The idea is not to accurately control the CVs, merely to keep them between the limits. Figure 29. Specifying seed models for CV/MV pairs Once the seed MPC application is configured and commissioned, the engineer can build the stepping application as illustrated in Figure 25. From this point the engineer needs to configure the stepping software as illustrated in Figure 21, and follow the rest of the procedure for open loop stepping. As Figure 26 shows, the seed MPC application is connected to the stepping application and the operator can pause and resume stepping from the seed MPC user interface as necessary. The optimizer of the seed MPC should be turned off to begin with. The stepping results should be similar to the Optimizer Off data from Figure 27. The engineer should utilize the calculated signal to noise ratio for each sub model to ensure the MV step sizes are large enough. The engineer should also verify that the settling times specified for the MVs are not too short. Once the stepping procedure begins to produce quality models the engineer can verify the model prediction and then lock the models as illustrated in Figure 23. The engineer can then begin to download models into the online running seed MPC application (Figure 30). As shown, the engineer can select all models or a portion of the models to be downloaded (blue background). The engineer pauses the stepper, downloads the models and then resumes stepping. The stepping software combines the selected models with the models already in the seed MPC, builds the new model matrix, and notifies the seed MPC to read in the new set of models. No visit by the engineer to the off-line software is necessary. The seed MPC reads in the new model matrix, performs bump-less initialization and begins using the new matrix as shown in Figure 31

31 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 31 Figure 30. Downloading new models from stepping software to the seed MPC application Figure 31. Updated seed MPC model matrix The seed MPC matrix at the beginning of stepping (top) and after a new set of models is downloaded (bottom) can be seen. The bottom matrix is much denser. At this point, the engineer has the option to configure the optimizer to maintain the process closer to ideal economic operating conditions as the remainder of the stepping is conducted. The engineer must maintain an appropriate balance in the interaction between stepping and the optimizer making compensating moves, as illustrated in Figure 27. The engineer continues to monitor for lockable quality models and download them until all sub models in the seed MPC application are updated. At this point, the stepping procedure is complete.

32 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 32 This method provides an alternative to open loop step testing. By building a seed MPC application up front, the engineer can realize all the benefits of running the stepping software with an MPC application online to address any CV constraint violations. The ability to download newly acquired models into the seed MPC application, without having to turn it off, enables model validation to be incorporated into the step testing procedure. Once all the models for a CV are obtained and downloaded, it is possible to observe in part how well the seed MPC application can handle CV constraint violations during the step testing procedure, reducing the need to do this after commissioning commences. As key models in the seed MPC application are obtained, the engineer can configure the seed MPC to operate closer to ideal economic conditions and key operating constraints while the remainder of the step testing is conducted. The engineer can then moderate how aggressively the seed MPC pushes toward ideal economic conditions to ensure quality step test data for the remaining models can still to be obtained. The engineer also has a measure of the signal to noise ratio for each sub model to provide early confirmation on the sufficiency of the MV move sizes. The new method has reduced the duration of MPC projects even further. Time taken to step a process is cut by another 25%. Model validation and correcting erroneous models are incorporated into the stepping procedure, as is preliminary validation of how well the models address CV constraints. This cuts commissioning time by an additional 25%. Projects now typically take under 2 months from start to finish for small applications and up to 8 months for larger applications. Stepping the process, identifying models, validating models and commissioning are reduced to as little as 30% of the overall project time. Recent Enhancements (Since 2010) Work to reduce implementation times and the skill required for step testing continues today, with a number of the improvements being developed: Technology to assist the engineer in obtaining seed models from historical data. Combined manual and automated stepping (allows simultaneous stepping of fast responding or well behaved processes manually) Techniques to improve automated bad data removal Techniques to improved model quality assessment Techniques to estimate an appropriate starting step size and settling time Techniques to alert the engineer earlier that step sizes are too small for getting good quality data. It is not yet clear how much more can be cut from the time required to implement MPC applications or what is the point of diminishing returns. However, progress here and in efforts to reduce the skill level required to implement step testing continues to be made. Summary MPC technologies were introduced in the 1970s. Instead of deriving linear ODEs mathematically to predict the process behavior, it was discovered that mathematical algorithms could approximate the models empirically. Empirical model generation, however, required process data to correlate the movements in MVs and DVs to CVs. Methods for bumping or stepping the process were thus introduced. This started with manually stepping MVs and DVs. A key problem was correlated moves between the MV being stepped and compensating moves on other MVs. This was driven by the need to manually address limit violations of key operating variables. The correlated moves were inadvertent, but common due to human nature and the desire to avoid upsetting the process operation too much. Another key issue was the tendency to make a step change and hold it until the process settled out. This allowed unmeasured disturbances and drift in the process to corrupt the step test data. The final problem was that step changes were much larger than the changes the MPC application made to control the process. The MPC application typically made small, smooth changes over time, but to generate enough movement for the ID algorithm to get models manual stepping made large changes. This tended to throw the process out of equilibrium much more than the MPC application moves and rendered the models inadequate for prediction. The result was prolonged step testing periods, a lot of time spent massaging data and re-running the ID algorithm, or repeated stepping. All these combined to elongate project times. The 1990s introduced sequential automated, rather than manual, step patterns for each MV. This addressed the issues of holding steps until the process settled out by creating step patterns that had steps of varying length, reducing the impact of unmeasured disturbances

33 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 33 and process drift on data quality. These methods also reduced the incidence of correlated moves rendering data useless and issues resulting from making steps too large, through filtering of conventional step patterns or deploying a sinusoidal based pattern. However, automated step patterns made it more difficult to decipher whether quality data was produced so erring on the side of large steps was usual. New issues also arose, with engineers tending to underestimate process settling times to speed up step testing. This led to models with erroneous gains and settling times. Shorter switching intervals also made it difficult if not impossible to massage the data during the off-line ID procedure. The next methods, introduced around 2000, consisted of deploying simultaneous automated, rather than sequential, step patterns. This essentially eliminated the impact of unmeasured disturbances and drift as well as correlated moves due to human intervention. Some steps that were too large still occured, but it was much less of an issue. Moving multiple MVs simultaneously created a built in filtering affect allowing larger steps without significantly throwing the process out of equilibrium. The result was richer data sets to run the ID algorithm against and models that gave good predictions, even allowing for the smaller, smoother changes deployed by the MPC application. Moreover, the ability to filter in steps was still available if very large steps had to be made. However, the issue of providing a representative settling time became crucial to avoid low quality or useless data. Massaging data off-line during the model ID procedure was also essentially impossible, so re-stepping was necessary if ID results were unfavorable. From around 2005 stepping software could combine stepping the process, collecting data, identifying models and validating models into a single procedure. Both sequential stepping and simultaneous stepping were supported. The need to collect data and transport it to an off-line software package to run the ID algorithm was eliminated. Automatic data massaging algorithms were implemented to look for certain types of bad data and remove them, and the engineer could also remove bad data on the fly. The ID algorithm automatically estimated the quality of models being identified and reported it back to the engineer. Correlated moves were also all but eliminated as the engineer could make compensating changes to MVs as necessary from the stepping software. Likewise, steps that were too large were avoided since the engineer could observe ID results on the fly and make adjustments to step size as appropriate. The option to filter steps is still available but used less often. There were also improvements in the ability to assess data quality when stepping MVs simultaneously as the engineer could stop stepping MVs to isolate a single MV or group, and then resume them. The ability to view periodic ID results allowed the engineer to make adjustments early to get better quality data, and to perform the model validation on the fly. Finally, the engineer no longer had to guess when to stop testing as the procedure indicates when quality models are obtained and stepping can cease. These and other advances reduced times and experience levels required, with the software s built in intelligence helping guide the engineer. The biggest issues remaining were the risk of not specifying a long enough settling time or specifying steps sizes that were too small. However, both could be changed online. At worst stepping was prolonged. Another method introduced in 2005 was to use automated stepping software in conjunction with an online MPC application to perform closed loop stepping. Stepping software that connected easily to an existing MPC application was developed so re-stepping of MVs could take place to repair errant sub models while the MPC application was running. This extension to the stepping software allowed new models to be acquired without having to turn the MPC application off and conduct open loop stepping. In addition to the benefits delivered for open loop stepping, the stepping software coordinated step changes to the MVs with changes required by the MPC application to keep CVs from violating limits. This eliminated the need for human intervention unless a severe process upset occurred. The engineer could also moderate the interaction between the steps being made to the MVs and the MPC application making countering moves to prioritize obtaining better quality step data. Overall the new closed loop method reduced the time to acquire and re-commission new models by up to 75% compared with turning the MPC application off and re-conducting open loop stepping. This new method also opened the door to the next big evolution in step testing methodologies. These methods, introduced around 2010, extended the capability of closed loop step testing to open loop stepping by using a seed MPC application. The engineer first defines the application s CVs, MVs and DVs and adds seed models for key CV/MV pairs; implements the seed MPC application online and conducts simple model validation; and then connects the stepping software to the seed MPC application and follows the rest of the closed loop procedure.

34 An Evolution of Step Testing and its Impact on Model Predictive Control Project Times 34 In addition to the benefits of closed loop stepping, the engineer now has the ability to change seed model gains on the fly if the seed MPC is not addressing CV constraint violations appropriately. The engineer can also download newly acquired models from the stepping software into the seed MPC application without turning the application off. As new models are downloaded to the seed MPC application, the engineer can begin to observe how well the seed MPC application handles CV constraint violations, reducing the need to do this after commissioning commences. As key models in the seed MPC application are obtained, the engineer can configure the seed MPC to operate closer to ideal economic conditions and key operating constraints while the remainder of the step testing is conducted. This enables engineers to build the MPC model structure as they go, extending the closed loop step testing method previously described. The following table summarizes the project timeline for each method. Further enhancements are still being made to reduce project times and reduce skill levels required. Progress is also being made to lessen the impact step testing has on the normal process operation and to reduce errant practices that creep into step testing procedures. The final line on the evolution of step testing has yet to be written. Author John A. Escarcega Business Consultant for APC and Optimization Honeywell Process Solutions For More Information Learn more about how Honeywell s alarm management solutions can improve your plant safety and profitability, visit our website or contact your Honeywell account manager. Honeywell Process Solutions Honeywell 1250 West Sam Houston Parkway South Houston, TX Honeywell House, Arlington Business Park Bracknell, Berkshire, England RG12 1EB Shanghai City Centre, 100 Junyi Road Shanghai, China WP ENG June Honeywell International Inc.

Making APC Perform 2006 ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc.

Making APC Perform 2006 ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc. Making APC Perform 2006 ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc. Summary Advanced Process Control (APC) promises to deliver optimal plant performance by optimizing setpoints, decoupling interactions,

More information

Process Control and Optimization Theory

Process Control and Optimization Theory Process Control and Optimization Theory Application to Heat Treating Processes Jake Fotopoulos, Lead Process Controls Engineer Air Products and Chemicals, Inc. Process Control and Optimization Theory --

More information

ADVANCED PROCESS CONTROL

ADVANCED PROCESS CONTROL CHAPTER-11 ADVANCED PROCESS CONTROL Chapter No 11 Page No 1 INTRODUCTION The traditional control philosophy, what is called instrumentation in the chemical industries, is based on single loop control (sometimes

More information

Efficiency up, emissions down

Efficiency up, emissions down Efficiency up, emissions down Achieving higher power plant performance with advanced process control Pekka Immonen, Ted Matsko, Marc Antoine Coal-fired power plants have come a long way in the last decades

More information

Best Practices for Controller Tuning

Best Practices for Controller Tuning Best Practices for Controller Tuning George Buckbee, P.E. ExperTune, Inc. 2013 ExperTune, Inc. Page1 Best Practices for Controller Tuning George Buckbee, P.E., ExperTune Inc. 2013 ExperTune Inc Summary

More information

WHITE PAPER Achieving Operational Excellence in the Chemicals Industry Using APC

WHITE PAPER Achieving Operational Excellence in the Chemicals Industry Using APC WHITE PAPER Achieving Operational Excellence in the Chemicals Industry Using APC Tushar Singh, Product Marketing Manager, Aspen Technology, Inc. Kate Kulik, Sr. Principal Business Consultant, Aspen Technology,

More information

Optimize your Process with Advanced Process Control

Optimize your Process with Advanced Process Control Honeywell.com 1 Document control number 2014 Honeywell Users Group Europe, Middle East and Africa Optimize your Process with Advanced Process Control Richard Salliss, Honeywell Honeywell Proprietary Agenda

More information

Control Systems Investment and Return Part 2 of 3

Control Systems Investment and Return Part 2 of 3 Control Systems Investment and Return Part 2 of 3 By F.G. Shinskey Sponsored by: 2012 ExperTune 1 Objective CONTROL SYSTEMS: INVESTMENT AND RETURN with Examples from Industry F. G. Shinskey Process Control

More information

SimSci APC. Model Predictive Control to Improve Your Process Economics

SimSci APC. Model Predictive Control to Improve Your Process Economics SimSci APC Model Predictive Control to Improve Your Process Economics In today s economic environment, capital budgets and overhead are constantly being cut. Companies are faced with rising manufacturing

More information

Alternatives to Optimize Gas Processing Operations

Alternatives to Optimize Gas Processing Operations Mark E. Roop esimulation, Inc. Houston, Texas, U.S.A. Jean Leger Enogex, Inc. Oklahoma City, Oklahoma, U.S.A. Steve Hendon esimulation, Inc. Houston, Texas, U.S.A. CHALLENGES AND REQUIREMENTS Gas processing

More information

Dynamic Simulation for APC projects A case study on a Reformate Splitter with side draw. Dr Sebastien OSTA TOTAL Jose Maria FERRER Inprocess

Dynamic Simulation for APC projects A case study on a Reformate Splitter with side draw. Dr Sebastien OSTA TOTAL Jose Maria FERRER Inprocess Dynamic Simulation for APC projects A case study on a Reformate Splitter with side draw Dr Sebastien OSTA TOTAL Jose Maria FERRER Inprocess Introduction Steady-state simulation is used traditionally for

More information

ADVANCED PROCESS CONTROL FOR YARA UREUM PLANT BRUNSBÜTTEL

ADVANCED PROCESS CONTROL FOR YARA UREUM PLANT BRUNSBÜTTEL ADVANCED PROCESS CONTROL FOR YARA UREUM PLANT BRUNSBÜTTEL SVEN KOHNKE, YARA BRUNSBÜTTEL, GERMANY BRAM LEEN, IPCOS, BELGIUM TOM BOELEN, IPCOS, BELGIUM CHRISTIAAN MOONS, IPCOS, BELGIUM 1 Abstract IPCOS,

More information

Unsteady-State Simulation Using CHEMCAD Steady-State

Unsteady-State Simulation Using CHEMCAD Steady-State Unsteady-State Simulation Using CHEMCAD Steady-State State Jerry Hartis, Sr. Process Engineer Grant Duncan, Process Engineer Process Engineering Associates, LLC June 2008 Copyright 2009 Process Engineering

More information

Optimization of a CAN Granulation Process with INCA Advanced Process Control

Optimization of a CAN Granulation Process with INCA Advanced Process Control Optimization of a CAN Granulation Process with INCA Advanced Process Control Steve Helfensteyn, Process Manager Fertilizers, EuroChem Antwerpen NV, Antwerpen, Belgium Heinz Falkus, APC Consultant, IPCOS,

More information

OPERATOR INTERACTION WITH MODEL-BASED PREDICTIVE CONTROLLERS IN PETROCHEMICAL REFINING

OPERATOR INTERACTION WITH MODEL-BASED PREDICTIVE CONTROLLERS IN PETROCHEMICAL REFINING OPERATOR INTERACTION WITH MODEL-BASED PREDICTIVE CONTROLLERS IN PETROCHEMICAL REFINING ABSTRACT Greg A. Jamieson Honeywell Technology Center Stephanie Guerlain University of Virginia In this paper, we

More information

FLOTATION CONTROL & OPTIMISATION

FLOTATION CONTROL & OPTIMISATION FLOTATION CONTROL & OPTIMISATION A global leader in mineral and metallurgical innovation FLOATSTAR OVERVIEW Flotation is a complex process that is affected by a multitude of factors. These factors may

More information

Chapter 3 Assembly Systems. Screen Titles

Chapter 3 Assembly Systems. Screen Titles Chapter 3 Assembly Systems Screen Titles System Input / Output Methods of Assembly Manual Assembly Automated Assembly Flexible Assembly Choice of Assembly Assembly Economics Assembly Line Components Assembly

More information

A Holistic Approach to Control and Optimisation of an Industrial Crushing Circuit

A Holistic Approach to Control and Optimisation of an Industrial Crushing Circuit A Holistic Approach to Control and Optimisation of an Industrial Crushing Circuit D Muller*, P.G.R de Villiers**, G Humphries*** *Senior Process Control Engineer, Anglo Platinum, Control and Instrumentation

More information

Simplify APC Tuning and Configuration

Simplify APC Tuning and Configuration WHITE PAPER Simplify APC Tuning and Configuration Advanced Process Control Tushar Singh, APC Product Marketing, Aspen Technology, Inc. Lucas Reis, APC Product Management, Aspen Technology, Inc. Introduction

More information

Shift Your Focus from Analysis to Optimization Using Active Model Capture Technology

Shift Your Focus from Analysis to Optimization Using Active Model Capture Technology Shift Your Focus from Analysis to Optimization Using Active Model Capture Technology 2005 ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc. Summary Process Industries are scrambling to keep up with

More information

OPTIMAX Advanced Control and Optimization of Power Plants

OPTIMAX Advanced Control and Optimization of Power Plants Pekka Immonen, Manager, Plant Optimization, OPTIMAX Advanced Control and Optimization of Power Plants OPTIMAX Suite for Real-Time Optimization Fleet-wide Optimization Unit and Component Optimization Performance

More information

Shift Your Focus from Analysis to Optimization Using Active Model Capture Technology

Shift Your Focus from Analysis to Optimization Using Active Model Capture Technology Shift Your Focus from Analysis to Optimization Using Active Model Capture Technology 2005 ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc. Summary Process Industries are scrambling to keep up with

More information

3 Steps to Better Steam Temperature Control Proven Strategies to Minimize Thermal Stress and Tube Leaks

3 Steps to Better Steam Temperature Control Proven Strategies to Minimize Thermal Stress and Tube Leaks December 17 3 Steps to Better Proven Strategies to Minimize Thermal Stress and Tube Leaks www.emerson.com/ December 17 Commercial power generators are having to significantly change how they operate their

More information

CT425 Optimize Your Process to Make Profits

CT425 Optimize Your Process to Make Profits - 5058-CO900H 1 CT425 Optimize Your Process to Make Profits Rockwell s Model Predictive Control Technology Aaron Dodgson Pavilion Business Development Western Region aadodgson@ra.rockwell.com PUBLIC 2

More information

Lecture 1. In practice, most large systems are developed using a. A software process model is an abstract representation

Lecture 1. In practice, most large systems are developed using a. A software process model is an abstract representation Chapter 2 Software Processes Lecture 1 Software process descriptions When we describe and discuss processes, we usually talk about the activities in these processes such as specifying a data model, designing

More information

Don t Tune These Four Loops!

Don t Tune These Four Loops! Don t Tune These Four Loops! George Buckbee 2012 ExperTune Inc. Page 1 Don t Tune These Four Loops! George Buckbee, ExperTune Inc. 2012 ExperTune Inc Summary Some control loops cannot be improved by tuning.

More information

ECNG 3032 CONTROL AND INSTRUMENTATION Introduction To Process Control

ECNG 3032 CONTROL AND INSTRUMENTATION Introduction To Process Control ECNG 3032 CONTROL AND INSTRUMENTATION 1 1. Introduction To Process Control The Aim of this module is to provide students of Electrical and Computer Engineering with a useful and practical introduction

More information

Software Processes 1

Software Processes 1 Software Processes 1 Topics covered Software process models Process activities Coping with change 2 The software process A structured set of activities required to develop a software system. Many different

More information

Control System Design for HVAC System and Lighting System using PID and MPC Controller

Control System Design for HVAC System and Lighting System using PID and MPC Controller Journal of Engineering and Science Research 1 (2): 66-72, e-issn: RMP Publications, DOI: Control System Design for HVAC System and Lighting System using PID and MPC Controller Nur Azizah Amir and Harutoshi

More information

Product Documentation SAP Business ByDesign February Business Configuration

Product Documentation SAP Business ByDesign February Business Configuration Product Documentation PUBLIC Business Configuration Table Of Contents 1 Business Configuration.... 4 2 Business Background... 5 2.1 Configuring Your SAP Solution... 5 2.2 Watermark... 7 2.3 Scoping...

More information

12th International Conference on Fluidized Bed Technology

12th International Conference on Fluidized Bed Technology 12th International Conference on Fluidized Bed Technology DYNAMIC BEHAVIOR OF A COMMERCIAL CFB UNIT MEASUREMENTS AND RELATED STUDIES Jen Kovács 1*, Ari Kettunen 1, István Selek 2 1 Amec Foster Wheeler

More information

This under-utilized approach can enhance operations and the bottom line.

This under-utilized approach can enhance operations and the bottom line. This under-utilized approach can enhance operations and the bottom line. By David Huffman, ABB Inc. Plants often overlook their automation system as a resource for improving overall equipment effectiveness

More information

Daniel Feltes Phone: Fax:

Daniel Feltes Phone: Fax: Coordination of Controls of Renewable Power Plants to Meet Steady State and Dynamic Response Requirements for Voltage Control and Reactive Power Supply Dinemayer Silva Phone: 518-395-5169 Fax: 518-346-2777

More information

Introduction to Artificial Intelligence. Prof. Inkyu Moon Dept. of Robotics Engineering, DGIST

Introduction to Artificial Intelligence. Prof. Inkyu Moon Dept. of Robotics Engineering, DGIST Introduction to Artificial Intelligence Prof. Inkyu Moon Dept. of Robotics Engineering, DGIST Chapter 9 Evolutionary Computation Introduction Intelligence can be defined as the capability of a system to

More information

How to Read Sine Language : Reduce Process Cycling to Save Money

How to Read Sine Language : Reduce Process Cycling to Save Money How to Read Sine Language : Reduce Process Cycling to Save Money 2006 ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc. How to Read Sine Language : Reduce Process Cycling to Save Money George Buckbee,

More information

ADVANCED CONTROL STRATEGIES FOR POLYOLEFIN GAS PHASE PROCESSES

ADVANCED CONTROL STRATEGIES FOR POLYOLEFIN GAS PHASE PROCESSES Proceedings of the American Control Conference Anchorage, AK May 8-1 0,2002 ADVANCED CONTROL STRATEGIES FOR POLYOLEFIN GAS PHASE PROCESSES Chih-An Hwang, Doug Johnson, and Stephen Goff Pavilion Technologies,

More information

Optimizing Batch Process Control

Optimizing Batch Process Control Optimizing Batch Process Control In Your Spare Time Standards Certification Education & Training Publishing Conferences & Exhibits Presenter Danaca Jordan Mentee of Greg McMillan (Coauthor) Bachelor s

More information

Department of Chemical Engineering, University of California Santa Barbara, Santa Barbara, CA 93106, U.S.A.

Department of Chemical Engineering, University of California Santa Barbara, Santa Barbara, CA 93106, U.S.A. Plantwide Optimization of a Pulp Mill Process Mehmet Mercangoz, Francis J. Doyle III Department of Chemical Engineering, University of California Santa Barbara, Santa Barbara, CA 93106, U.S.A. Abstract

More information

Getting Started with OptQuest

Getting Started with OptQuest Getting Started with OptQuest What OptQuest does Futura Apartments model example Portfolio Allocation model example Defining decision variables in Crystal Ball Running OptQuest Specifying decision variable

More information

More is Better? Analyzing the Relation between Metering Duration and Accuracy of Results for Residential Lighting Evaluations

More is Better? Analyzing the Relation between Metering Duration and Accuracy of Results for Residential Lighting Evaluations More is Better? Analyzing the Relation between Metering Duration and Accuracy of Results for Residential Lighting Evaluations Joel D. Pertzsch, Michaels Energy Inc., La Crosse, WI Ryan M. Kroll, Michaels

More information

Chapter 3 Prescriptive Process Models

Chapter 3 Prescriptive Process Models Chapter 3 Prescriptive Process Models - Generic process framework (revisited) - Traditional process models - Specialized process models - The unified process Generic Process Framework Communication Involves

More information

Application for process automation

Application for process automation Application for process automation Guidelines for the use of the SIMATIC PCS 7 embedded MPC Notes on use and limitations Guarantee, liability and support APC guidelines Note The application examples are

More information

VARIABLE STRUCTURE CONTROL

VARIABLE STRUCTURE CONTROL When I complete this chapter, I want to be able to do the following. Understand why many applications of process control require variable structure Implement a design using more than one valve in a control

More information

Automated Black Box Testing Using High Level Abstraction SUMMARY 1 INTRODUCTION. 1.1 Background

Automated Black Box Testing Using High Level Abstraction SUMMARY 1 INTRODUCTION. 1.1 Background Automated Black Box Testing Using High Level Abstraction Dake Song, MIRSE, USA Dr Uli Dobler, FIRSE, Germany Zach Song, EIT, Canada SUMMARY One of the big bottlenecks of modern signalling projects lies

More information

Evolving Control for Micro Aerial Vehicles (MAVs)

Evolving Control for Micro Aerial Vehicles (MAVs) Evolving Control for Micro Aerial Vehicles (MAVs) M. Rhodes, G. Tener, and A. S. Wu Abstract This paper further explores the use of a genetic algorithm for the purposes of evolving the control systems

More information

Gregory Rogers, PE June HCP-0101 Fundamental of Advanced Controls

Gregory Rogers, PE June HCP-0101 Fundamental of Advanced Controls Gregory Rogers, PE June 2018 HCP-0101 Fundamental of Advanced Controls HCP (Advanced Controls) Journey Training Lineup 1 09:15 10:00am HCP-0101 Fundamentals of APC Greg Rogers Enterprise Products 10:00

More information

WHITE PAPER. Analytics Software. Find what Matters

WHITE PAPER. Analytics Software. Find what Matters WHITE PAPER Analytics Software Find what Matters We Have The Data: Lets Find What Matters The past decade has seen dramatic advances in automation systems and smart devices. From IP connected systems employing

More information

The concept of statical determinacy

The concept of statical determinacy Appendix 3 The concept of statical determinacy 140 A3.1 Introduction It has been shown that the conditions for equilibrium of a set of coplanar forces can be summarised in the three equations of equilibrium

More information

CASE STUDY ANSWERIQ SUPPORT ENABLES 100% AUTOMATED TICKET CLASSIFICATION AND 50% TIME-TO-RESPONSE REDUCTION FOR

CASE STUDY ANSWERIQ SUPPORT ENABLES 100% AUTOMATED TICKET CLASSIFICATION AND 50% TIME-TO-RESPONSE REDUCTION FOR CASE STUDY ANSWERIQ SUPPORT ENABLES 100% AUTOMATED TICKET CLASSIFICATION AND 50% TIME-TO-RESPONSE REDUCTION FOR TRIAGE SMART CLASSIFICATION MACHINE LEARNING ThredUP is an ecommerce shop selling secondhand

More information

Batch Reactor Temperature Control Improvement

Batch Reactor Temperature Control Improvement Batch Reactor Temperature Control Improvement Flávio P. Briguente (Monsanto) and Greg McMillan (CDI Process & Industrial) Key words: batch reactor, temperature control, solubility curve. ABSTRACT Temperature

More information

Predictive Intelligence Optimizing Performance & Reliability. Neeraj Agarwal Manish Khetrapal

Predictive Intelligence Optimizing Performance & Reliability. Neeraj Agarwal Manish Khetrapal Predictive Intelligence Optimizing Performance & Reliability Neeraj Agarwal Manish Khetrapal 1 Outline.. Introduction Basic Architecture The MPC Model Potential Areas Conclusion 2 Sustainable Growth Optimize

More information

Affordable plant layout and process simulation software

Affordable plant layout and process simulation software White Paper Affordable plant layout and process simulation software contact Visual Components Ltd. Korppaanmäentie 17 CL6 FI-00300 Helsinki, Finland Office phone +358 9 323 2250 info@visualcomponents.com

More information

Interaction Detection Using Oscillation Analysis

Interaction Detection Using Oscillation Analysis Interaction Detection Using Oscillation Analysis by Tom Kinney, ExperTune Keywords: performance supervision, manufacturing intelligence, power spectrum, cycling, optimization, troubleshooting, variability,

More information

Optimization of the Aeration of Sewage Treatment Plants with Advanced Process Control

Optimization of the Aeration of Sewage Treatment Plants with Advanced Process Control Siemens AG, 26.11.2015 Optimization of the Aeration of Sewage Treatment Plants with Advanced Process Control Dr. Andreas Pirsing Sewage Treatment Plants cause High Energy Consumption An increase in energy

More information

Schedule Compression

Schedule Compression Schedule Compression The need to reduce the time allowed for a schedule, or a part of a schedule is routine, some of the times the need arises include: When the initial schedule is too long to meet contractual

More information

MILLING CONTROL & OPTIMISATION

MILLING CONTROL & OPTIMISATION MILLING CONTROL & OPTIMISATION MillSTAR OVERVIEW Due to the complex nature of milling circuits, it is often found that conventional control does not address many of the common problems experienced. These

More information

Optimization Solution White Paper. An Overview of Honeywell s Layered Optimization Solution

Optimization Solution White Paper. An Overview of Honeywell s Layered Optimization Solution Optimization Solution White Paper An Overview of Honeywell s Layered Optimization Solution Optimization Solution White Paper 2 Table of Contents Table of Contents...2 Table of Figures...3 Introduction...4

More information

Schedule Compression

Schedule Compression Schedule Compression The need to reduce the time allowed for a schedule, or a part of a schedule is routine, some of the times the need arises include: When the initial schedule is too long to meet contractual

More information

Technical Note OPTIMIZATION OF THE PARAMETERS OF FEEDWATER CONTROL SYSTEM FOR OPR1000 NUCLEAR POWER PLANTS

Technical Note OPTIMIZATION OF THE PARAMETERS OF FEEDWATER CONTROL SYSTEM FOR OPR1000 NUCLEAR POWER PLANTS Technical Note OPTIMIZATION OF THE PARAMETERS OF FEEDWATER CONTROL SYSTEM FOR OPR1000 NUCLEAR POWER PLANTS UNG SOO KIM *, IN HO SONG, JONG JOO SOHN and EUN KEE KIM Safety Analysis Department, KEPCO Engineering

More information

Software Engineering Part 2

Software Engineering Part 2 CS 0901341 Software Engineering Part 2 In this part, we look at 2.1 Software Process 2.2 Software Process Models 2.3 Tools and Techniques for Processing Modelling As we saw in the previous part, the concept

More information

Livia Tizzo, Process Control and Automation Engineer, Braskem. Case Study

Livia Tizzo, Process Control and Automation Engineer, Braskem. Case Study Livia Tizzo, Process Control and Automation Engineer, Braskem Case Study We saved around six months on APC implementation using Aspen DMC3. - Livia Tizzo, Process Control and Automation Engineer, Braskem

More information

Connoisseur. Software Datasheet MODEL PREDICTIVE CONTROL TO IMPROVE YOUR PROCESS ECONOMICS MAXIMIZE YOUR PROFITS. Summary.

Connoisseur. Software Datasheet MODEL PREDICTIVE CONTROL TO IMPROVE YOUR PROCESS ECONOMICS MAXIMIZE YOUR PROFITS. Summary. Invensys is now Software Datasheet Summary Connoisseur model predictive control is a comprehensive offering for any size control problem or process. By exploiting its unequaled features including online

More information

Real-Time Asset Performance in Process Plants ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc.

Real-Time Asset Performance in Process Plants ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc. Real-Time Asset Performance in Process Plants 2007 ExperTune, Inc. George Buckbee, P.E. ExperTune, Inc. Real-Time Asset Performance in Process Plants George Buckbee, P.E., ExperTune Inc. 2007 ExperTune

More information

ISA-106. and concepts of procedural automation

ISA-106. and concepts of procedural automation ISA-106 and concepts of procedural automation By Maurice Wilkins and Marcus Tennant Manual operations live on and even thrive in largely automated process plants, often causing safety incidents or all

More information

Simple Rules for Economic Plantwide Control

Simple Rules for Economic Plantwide Control Krist V. Gernaey, Jakob K. Huusom and Rafiqul Gani (Eds.), 12th International Symposium on Process Systems Engineering and 25th European Symposium on Computer Aided Process Engineering. 31 May - 4 June

More information

CHAPTER 1 INTRODUCTION

CHAPTER 1 INTRODUCTION 1 CHAPTER 1 INTRODUCTION 1.1 MANUFACTURING SYSTEM Manufacturing, a branch of industry, is the application of tools and processes for the transformation of raw materials into finished products. The manufacturing

More information

Measuring Effort and Productivity of Agile Projects

Measuring Effort and Productivity of Agile Projects Measuring Effort and Productivity of Agile Projects Measuring Effort Measuring effort means the same thing in agile and non-agile methods: you need to know the hours spent by people working on the release.

More information

Three Element Boiler Drum Level Control using Cascade Controller

Three Element Boiler Drum Level Control using Cascade Controller Three Element Boiler Drum Level Control using Cascade Controller A. Amarnath Kumaran 1, M. Ponni Bala 2, V. Sivaraman 3 1 PG Scholar, 2 Associate Professor, 3 Instrumentation Manager, Department of Electronics

More information

PROCESS DESIGN AND CONTROL. Guides for the Selection of Control Structures for Ternary Distillation Columns. William L. Luyben*

PROCESS DESIGN AND CONTROL. Guides for the Selection of Control Structures for Ternary Distillation Columns. William L. Luyben* Ind. Eng. Chem. Res. 2005, 44, 7113-7119 7113 PROCESS DESIGN AND CONTROL Guides for the Selection of Control Structures for Ternary Distillation Columns William L. Luyben* Process Modeling and Control

More information

CHILLED WATER SYSTEM OPTIMIZER

CHILLED WATER SYSTEM OPTIMIZER CHILLED WATER SYSTEM OPTIMIZER A White Paper by Steve Tom, P.E., Phd. Carrier Corporation Farmington, Connecticut July, 2017 INTRODUCTION When it comes to measuring HVAC energy use in buildings, it s

More information

Chapter 24. Software Project Scheduling

Chapter 24. Software Project Scheduling Chapter 24 Software Project Scheduling - Introduction - Project scheduling - Task network - Timeline chart - Earned value analysis (Source: Pressman, R. Software Engineering: A Practitioner s Approach.

More information

Monitoring Controller Performance and Production Reliability

Monitoring Controller Performance and Production Reliability Monitoring Controller Performance and Production Reliability Robert Rice Vice President, Engineering Agenda General Information - Industry Challenges and Opportunities Breakdown of Control Loop Performance

More information

The Myths. George Buckbee, P.E.

The Myths. George Buckbee, P.E. George Buckbee, P.E. The Myths Myth 1: Tuning is Everything Myth 2: Averages are More Important Than Dynamics Myth 3: There Isn t Time to Look at Everything Myth 4: Each Unit Operation is an Island Myth

More information

CHAPTER 6 A CDMA BASED ANTI-COLLISION DETERMINISTIC ALGORITHM FOR RFID TAGS

CHAPTER 6 A CDMA BASED ANTI-COLLISION DETERMINISTIC ALGORITHM FOR RFID TAGS CHAPTER 6 A CDMA BASED ANTI-COLLISION DETERMINISTIC ALGORITHM FOR RFID TAGS 6.1 INTRODUCTION Applications making use of Radio Frequency Identification (RFID) technology with large tag populations often

More information

[Ramli* et al., 6(2): February, 2017] ISSN: IC Value: 3.00 Impact Factor: 4.116

[Ramli* et al., 6(2): February, 2017] ISSN: IC Value: 3.00 Impact Factor: 4.116 IJESRT INTERNATIONAL JOURNAL OF ENGINEERING SCIENCES & RESEARCH TECHNOLOGY CONTROL STRATEGIES OF HEAT EXCHANGER Nasser Mohamed Ramli *, Haslinda Zabiri * Chemical Engineering Department, Faculty of Engineering,

More information

Global position system technology to monitoring auto transport in Latvia

Global position system technology to monitoring auto transport in Latvia Peer-reviewed & Open access journal www.academicpublishingplatforms.com The primary version of the journal is the on-line version ATI - Applied Technologies & Innovations Volume 8 Issue 3 November 2012

More information

PROCESS CONTROL LOOPS A Review Article

PROCESS CONTROL LOOPS A Review Article PROCESS CONTROL LOOPS A Review Article Lalit Goswami 1,2, Shagun Malhotra 3 1 Department of Electronics & Communication Engineering, Delhi Technological University, New Delhi. 2 Advanced Materials & Devices,

More information

PROFIT SUITE INNOVATIONS FOR NEXT-GENERATION APC AND REAL-TIME OPTIMIZATION

PROFIT SUITE INNOVATIONS FOR NEXT-GENERATION APC AND REAL-TIME OPTIMIZATION Richard Salliss October 2016 Richard Salliss 25 Oct 2016 PROFIT SUITE INNOVATIONS FOR NEXT-GENERATION APC AND REAL-TIME OPTIMIZATION Honeywell s Advanced Control Solutions Profit Controller Profit Sensor

More information

Time to Shipment MRP Guide DBA Software Inc.

Time to Shipment MRP Guide DBA Software Inc. Contents 3 Table of Contents 1 Benefits and Features 4 2 MRP Overview 6 3 MRP Phases 10 4 Phase 1 - Plan Times to Shipment 12 5 Phase 2 - Plan Supply Pipelines 25 6 Phase 3 - Generate Jobs and POs 34 7

More information

Software Next Release Planning Approach through Exact Optimization

Software Next Release Planning Approach through Exact Optimization Software Next Release Planning Approach through Optimization Fabrício G. Freitas, Daniel P. Coutinho, Jerffeson T. Souza Optimization in Software Engineering Group (GOES) Natural and Intelligent Computation

More information

Traffic insights ATSPM Flashcards

Traffic insights ATSPM Flashcards Traffic insights ATSPM Flashcards Traffic Insights Optimize your traffic network. Measure the impact. Traffic Insights are a suite of signal performance analytics and tools that help improve traffic safety

More information

BCS THE CHARTERED INSTITUTE FOR IT. BCS HIGHER EDUCATION QUALIFICATIONS BCS Level 6 Professional Graduate Diploma in IT SOFTWARE ENGINEERING 2

BCS THE CHARTERED INSTITUTE FOR IT. BCS HIGHER EDUCATION QUALIFICATIONS BCS Level 6 Professional Graduate Diploma in IT SOFTWARE ENGINEERING 2 BCS THE CHARTERED INSTITUTE FOR IT BCS HIGHER EDUCATION QUALIFICATIONS BCS Level 6 Professional Graduate Diploma in IT SOFTWARE ENGINEERING 2 Friday 30 th September 2016 - Morning Answer any THREE questions

More information

How to Transition From Annual Performance Reviews to Real-Time Feedback

How to Transition From Annual Performance Reviews to Real-Time Feedback How to Transition From Annual Performance Reviews to Real-Time Feedback The annual performance review is changing EMPLOYEES ORGANIZATIONS PERFORMANCE REVIEWS 51% 86% 30% Employees believe that annual performance

More information

AUTOMATE YOUR ORGANIZATION

AUTOMATE YOUR ORGANIZATION AUTOMATE YOUR ORGANIZATION Build an automation foundation for digital business operations Page 1 DIGITAL BUSINESS = AUTOMATED I.T. OPERATIONS Page 2 TODAY S REALITY FOR I.T. OPERATIONS Page 3 OBSTACLES

More information

Profit Suite R300. Release Highlights

Profit Suite R300. Release Highlights Profit Suite R300 Release Highlights Profit Suite Release 300 represents Honeywell s latest technological advances in plant-wide control and optimization, providing the tightest integration with the Experion

More information

Model Predictive Control. Rockwell Automation Model Predictive Control delivers results.

Model Predictive Control. Rockwell Automation Model Predictive Control delivers results. Model Predictive Control Rockwell Automation Model Predictive Control delivers results. The Challenge Today s manufacturing companies contend with intense global competition, reduced technical and operational

More information

DeltaV InSight. Introduction. Benefits. Gain new process insight from embedded process learning. Easily identify underperforming control loops

DeltaV InSight. Introduction. Benefits. Gain new process insight from embedded process learning. Easily identify underperforming control loops DeltaV Distributed Control System Product Data Sheet DeltaV InSight Gain new process insight from embedded process learning Easily identify underperforming control loops Quickly tune loops for improved

More information

Benefits of State Based Control

Benefits of State Based Control White Paper By David A. Huffman ABB Inc. February 16, 2009 Abstract State Based Control is a plant automation control design based on the principle that all process facilities operate in recognized, definable

More information

Process Control and Automation Systems Advancements for Reheat Furnaces

Process Control and Automation Systems Advancements for Reheat Furnaces Process Control and Automation Systems Advancements for Reheat Furnaces In the hot processing of steel to a formed product, the reheat furnace performs the function of raising the steel from the charge

More information

Pulping Process Analysis with Dynamic Delay Compensation

Pulping Process Analysis with Dynamic Delay Compensation Pulping Process Analysis with Dynamic Delay Compensation Mika Suojärvi: Product Manager, Savcor Oy, FI-50100 Mikkeli, Finland, mika.suojarvi@savcor.com Rodrigo Prado: Project Manager, Savcor Forest, São

More information

Automation System Optimization

Automation System Optimization Photo (optional) Automation System Optimization Achieve Superior System Performance After a DCS Migration by: Mike Vernak 2 Automation System Optimization Many process plants have an outdated Distributed

More information

Big Tent S&OP: Expanding the Scope of a Critical Process

Big Tent S&OP: Expanding the Scope of a Critical Process Big Tent S&OP: Expanding the Scope of a Critical Process Five Steps to Bridging the Gap Between Supply Chain Planning and Execution 1 Centralize 5 Improve 4 3 Execute Collaborate 2 Evaluate E2open ebook

More information

2014 Software Global Client Conference

2014 Software Global Client Conference Tips & Tricks from ROMeo Implementation Projects Robert S. Morrison Optimization Consultant September 30, 2014 Mobile App: Please take a moment Check into Session by: Select Detailed Schedule Select the

More information

Limitations of Model Predictive Controllers. Alan Hugo Senior Applications Engineer, Control Arts

Limitations of Model Predictive Controllers. Alan Hugo Senior Applications Engineer, Control Arts Introduction Limitations of Model Predictive Controllers Alan Hugo Senior Applications Engineer, Control Arts Multivariable Model Predictive Controllers (MPC) have been used to control process plants for

More information

Forecasting for Short-Lived Products

Forecasting for Short-Lived Products HP Strategic Planning and Modeling Group Forecasting for Short-Lived Products Jim Burruss Dorothea Kuettner Hewlett-Packard, Inc. July, 22 Revision 2 About the Authors Jim Burruss is a Process Technology

More information

Cyclical Dynamics of Airline Industry Earnings Online Supplement

Cyclical Dynamics of Airline Industry Earnings Online Supplement Cyclical Dynamics of Airline Industry Earnings Online Supplement Kawika Pierson* Atkinson Graduate School of Management, Willamette University, Salem OR John Sterman MIT Sloan School of Management, Cambridge,

More information

ADVANTAGES OF INLINE COLOUR MEASUREMENT Benefits for the printer and brand owner

ADVANTAGES OF INLINE COLOUR MEASUREMENT Benefits for the printer and brand owner ADVANTAGES OF INLINE COLOUR MEASUREMENT Benefits for the printer and brand owner RESEARCH ARTICLE Michael Sisco QuadTech, Inc., Michael.Sisco@quadtechworld.com ABSTRACT Management of color quality goes

More information

Chapter Six{ TC "Chapter Six" \l 1 } System Simulation

Chapter Six{ TC Chapter Six \l 1 } System Simulation Chapter Six{ TC "Chapter Six" \l 1 } System Simulation In the previous chapters models of the components of the cooling cycle and of the power plant were introduced. The TRNSYS model of the power plant

More information

Sawtooth Software. A Perspective on Adaptive CBC (What Can We Expect from Respondents?) RESEARCH PAPER SERIES. Rich Johnson, Sawtooth Software, Inc.

Sawtooth Software. A Perspective on Adaptive CBC (What Can We Expect from Respondents?) RESEARCH PAPER SERIES. Rich Johnson, Sawtooth Software, Inc. Sawtooth Software RESEARCH PAPER SERIES A Perspective on Adaptive CBC (What Can We Expect from Respondents?) Rich Johnson, Sawtooth Software, Inc. Copyright 2008, Sawtooth Software, Inc. 530 W. Fir St.

More information

Avoiding Common Pitfalls of SCM-ERP Integration

Avoiding Common Pitfalls of SCM-ERP Integration An Executive White Paper Avoiding Common Pitfalls of SCM-ERP Integration Harnessing the Advantages of a Custom Integration without the Drawbacks of Custom Development Table of Contents Best-of-Breed Solutions,

More information