PowerGen Europe Milano, June 2005 Marc Antoine ABB Power Technology Systems Parameter Estimation for Diagnosis and Optimization in Power Plants ABB Power Technology Systems -1-ma
Introduction Goal of parameter estimation: Generate real-time estimation of plant parameters and internal system states for control and optimization purposes States" = time varying magnitudes (e.g. temperatures or pressures), physical constants (e.g. masses or areas), functions of other plant states (e.g. efficiencies, characteristics), etc. Technology Kalman Filter: Predict model states and outputs given the previous estimation of the states Estimate the model states by correcting state prediction according the latest measurement ABB Power Technology Systems -2-ma
Parameter Estimation Boilers Degradation of heat transfer due to soot deposits Estimate surface effectiveness factors, furnace fouling factors, slagging, etc. ABB Power Technology Systems -3-ma
Parameter Estimation Cement Kiln Volatiles: Sulphur, Chlorine, etc. Vapour Transportation Condensation Volatile Vapourisation Solid Transportation Volatility Loss Unknown parameters to be estimated Measured Volatility (State) -C(t) ABB Power Technology Systems -4-ma
Parameter Estimation GT The consumed energy by a GT compressor typically accounts for 50% of the fuel. Compressor Fouling due to pollen, dust, hydrocarbon aerosols, salt... Turbine Erosion due to abbrasive removal of blade material. Turbine Fouling due accumulating combustion residuals on blades. Foreign Object Damage (FOD) due to detached parts, components breakdown. Changes Flow capacity, Temperatures, Pressures, Efficiency, Power, Cooling GT controller will try to compensate (e.g. VIGV; TAT ) Compressor Washing to partially recover from Compressor Fouling. ABB Power Technology Systems -5-ma
Examples around the GT Parameter Estimation Wash Optimizer Gas Path Diagnosis ABB Power Technology Systems -6-ma
Example of Estimation Result 0.95 Be fore Outa ge 0.90 After Outa ge 0.85 0.80 0.75 0.70 7 8 9 10 11 12 13 14 15 16 17 Pressure Ratio Different corrected estimates of efficiency vs. pressure ratio. The upper curve was estimated one week after the lower curve. During the week between the two estimates, the plant was down for maintenance. The estimated improvement in efficiency is about 2% over the entire range. ABB Power Technology Systems -7-ma
Compressor Wash Optimizer Features Identifies compressor maps based on parameter estimation Determines optimal washing schedule based on costs and Mixed Logical Dynamics (MLD) Possible to define and modify optimization constraints Benefits Decision-support tool Significant O&M Cost Savings Predictive instead of Preventive GR: Normal state, no washing BL: Online washing RE: Offline washing YE: Idle state WH: No constraint ABB Power Technology Systems -8-ma
Compressor Wash Optimizer Method Model relates state variables (m, p,...) to measured variables and unknowns (η,...) Estimator for time-varying unknown parameters Efficiency Degradation Model to be linked with an Optimizer engine Hybrid Dynamic Model links the physical model and the economic model ( hybrid because of Boolean decision variables) Performance model Measured data Parameter Estimation with Kalman Filter Hybrid Dynamic Model Boundary conditons Degradation model Fuel cost savings Optimization ABB Power Technology Systems -9-ma
Continuous Variables η : Actual degradation η 2 : Non-recoverable degradation α : Degradation rate for η α 2 : Degradation rate for η 2 Actual degradation vs. time Non-recoverable degradation Help variable Help variable ABB Power Technology Systems -10-ma
Introduction of Boolean Variables Day 0 δ 1 δ 2 δ 3 δ 4 δ 5 δ 6 η η 2 α α 2 z 1 z 2 z 3 z 4 Day 1 δ 1 δ 2 δ 3 δ 4 δ 5 δ 6 η η 2 α α 2 z 1 z 2 z 3 z 4 Day 2 δ 1 δ 2 δ 3 δ 4 δ 5 δ 6 η η 2 α α 2 z 1 z 2 z 3 z 4 δ 1 : Normal state, no washing δ 2 : Online washing δ 3 : Offline washing δ 4 : Idle state, no power generated δ 5 : help variable for α (degradation rate for η) δ 6 : help variable for α 2 (degradation rate for η 2 ) η : Actual degradation η 2 : Non-recoverable degradation z 1 -z 4 : Variables for objective function (to be defined) ABB Power Technology Systems -11-ma
Receeding Horizon Control δ 1 δ 2 δ 3 δ 4 δ 5 δ 6 η α α 2 z 1 z 2 z 3 z 4 Day 0 η 2 Day 1 δ 1 δ 2 δ 3 δ 4 δ 5 δ 6 η η 2 α α 2 z 1 z 2 z 3 z 4 Day 2 δ 1 δ 2 δ 3 δ 4 δ 5 δ 6 η η 2 α α 2 z 1 z 2 z 3 z 4 Day 3 δ 1 δ 2 δ 3 δ 4 δ 5 δ 6 η η 2 α α 2 z 1 z 2 z 3 z 4 Day N δ 1 δ 2 δ 3 δ 4 δ 5 δ 6 η η 2 α α 2 z 1 z 2 z 3 z 4 ABB Power Technology Systems -12-ma
Compressor Wash Optimizer Results Non-recoverable degradation Optimized solution No washing Washing Cycle No Washing Every Day Every week Optimal Additional Costs* +60% -22% 0-46% * normalized ABB Power Technology Systems -13-ma
Development of GT Diagnostics Progressive Progressive Maintenance Maintenance Strategy Strategy Quantitative Quantitative Fault Fault Diagnostics Diagnostics Fault Tree Qualitative Fault Matrix Qualitative Fault Fault Diagnostics Diagnostics Measurement Monitoring ABB Power Technology Systems -14-ma GT Gas Path Diagno sis GPD GPA Trending Monitoring Monitoring
GT Gas Path Diagnosis Symptoms cannot trace back to condition unequivocally: Uncertainty in causes Measurements are inexact, corrupted with noise: Uncertainty in observations PROBABILISTIC INFERENCE Estimator (DEKF ** ) takes noise of measurements and of process into account (time-varying health parameters) Determines covariances and calculates joint pdf * for the estimated health parameters Probabilistic Classifier quantifies component s level of degradation *Probability Distribution Function ** Discrete Extended Kalman Filter Measurement signal set Preprocessing Preprocessed measurement signal set G as Path Analysis Health parameter Gas Path Diagnosis Equipment fault likelihood ABB Power Technology Systems -15-ma
Discrete (Extended) Kalman Filter The ingredients: A discrete process model Change in state over time Difference equation A discrete measurement model Relationship between state and measurement Model Parameters Process noise characteristics Measurement noise characteristics u(k) x(k-1) v(k) D x(k) f h u: input signal x: state vector v: process noise z(k) + z: measurement noise f: model function h: output function y(k) ABB Power Technology Systems -16-ma
Fault Symptom Model Parameters Fault-conditional density p ( δ F i ) is parametrized by: subset of symptom elements j impacted by fault F i expected degradation (in case of fault F i ) µ i = E ( δ F i cloud center standard deviation of degradation cloud width correlation of degradations cloud orientation σ i = Var( δ F, j j i ρ i, jk = Corr ( δ j, δ k F i ), ρ i, jk 0.08 ) ) 0.07 0.06 0.05 0.04 0.03 0.02 TurbineFOD TurbineErosion TurbineFouling 0.01 0-0.08-0.06-0.04-0.02 0 0.02 0.04 0.06 0.08 Symptom S1 1 ABB Power Technology Systems -17-ma Symptom S2 Optimal
GT Gas Path Diagnosis - Example Monitored evolution of symptoms, coded by trajectory: Posterior probability, coded by colour map: P ( δ ) F i (t) δ Features Identifies and quantifies single and multiple physical faults (fouling, erosion, FOD, etc.) Real-time information on GT health Helps avoiding unnecessary trips Benefits Increases trust in instrumentation Reduces downtime Extends Maintenance Interval Provides early problem Recognition Movie ABB Power Technology Systems -18-ma
Conclusions Parameter estimation is a technique that can be used for identification of complex systems, in particular for control, diagnosis, and maintenance optimization. Combined with Model Predictive Control (MPC) this can be used for economic optimization (e.g. compressor wash optimizer). The same parameter estimation technique can be extended for equipment diagnosis. Equipment faults are specified by a confidence domain and the probabilistic classifier allows quantifying a component's level of degradation". ABB Power Technology Systems -19-ma