Further study of control system for reservoir operation aided by neutral networks and fuzzy set theory M.Hasebe*, T.Kumekawa** and M.Kurosaki*** *Energy and Environmental Science, Graduate School of Engineering, University of Utsunomiya, 2753, Ishii, Utsunomiya, Japan. **Utsunomiya Technical High School, Utsunomiya, Japan. ***Tokyo Erectoronic Power Co.LTD.,Groups of Dam Control System, Nagano Prefecture, Omachi, Japan. Email: hasebe@cc.utsunomiya-u.ac.jp Abstract In Japan, generally, the optimum operation of dam reservoir is desired to lighten the burden for the control system for reservoir operation. We have studied to apply the optimum control system of reservoir operation of the release discharge to fuzzy set theory and neural networks. Generally, the dam control is mainly classified into two groups by the object of utilization, namely, irrigation for water resources and flood control.
Now, in the past, we have investigated as to the dam control system that the object of the use is clear, for example, the reservoir operation for irrigation or flood control. In this study, as the subject of this paper, it is investigated for the control system of the reservoir operation of the multipurpose dam, which the storage volume have comparatively much capacity. This control system of reservoir operation is that neural networks is applied to the decision of the operational line and fuzzy set theory is applied to that of operational volume, that is, release discharge from reservoir of dam. It is investigated that the control system for reservoir operation of multipurpose dam, using both fuzzy set theory and neural networks system, is effective or not. 1 Introduction The hydrological information for dam management is abundant in quality and quantity because the technique of hydrological observation has made great advance recently. Thus, if the runoff mechanism of hydrological process in the river basin can be obvious, runoff prediction of inflow into dam reservoir may be performed precisely (e.g. Hasebe[4]). On the other hand, the control system for reservoir operation has been struggled for betterment. But, generally, the reservoir operating system of multipurpose dam in Japan have been carried out under the operational rule on the basis of inflow into reservoir. If that is the case, the control of the present rule for reservoir operation is inadequate to make better effective. Therefore, in the control system for reservoir operation, the author s join together two systems, namely, neural networks system and fuzzy system (e.g. Hasebe[3],[6]). The reasons why two systems are applicable is that the mathematical expressions of dam operating system are difficult and vague in a minor point as for reservoir operation of
multipurpose dam. For the above mentioned reasons, we considered that it is effective to apply two systems of both fuzzy set theory and neural networks to the control system for reservoir operation of multipurpose dam. In this paper, the operating system for multipurpose dam, based on the information obtained by inquiries to expert operators, is constructed and then, it is investigated that the control system of reservoir operation aided by both fuzzy system and neural networks system is effective or not. 2 The outline of the control system for reservoir operation of multipurpose dam This control system for reservoir operation of multipurpose dam is composed of two subsystems. One is subsystem of the decision of operational line for the selection of operating, (1) to release discharge from reservoir, (2) to keep constant water level in reservoir and (3) to storage discharge which flow in reservoir. The neural networks system can be applied to this subsystem which is the operational line. The other is one of the decision of operational volume to determine the volume of release discharge from reservoir. The fuzzy set theory can be applied to this subsystem which is the operational volume. 2.1 Fundamentals of operational line and operational volume First, the operational line aided by neural networks is determined by referring to both the manual where expert reservoir operators have operated actually and the information obtained by inquiries to actual reservoir operators. Neural networks system is composed of three layer perceptrons. Sensory units which is first layer are composed of seven neurons, that is, precipitation, river discharge, inflow into reservoir, inflow predicted by filter separation AR method (e.g. Hino[7]), changing
inflow, water level in reservoir and release discharge from reservoir. Association units to be second layer are composed of three neurons, that is, neuron to respond to the hydrological system ( dam basin ), one to respond to discharge and one to respond to the state of reservoir. Response units to be third layer is composed of one neuron, i.e., neuron for the selection of release discharge, storage volume and conservation of water level in reservoir. Secondly, the decision of operational volume, using fuzzy set theory, for release discharge from reservoir is determined to succeed to the operational line determined above. At this stage, fuzzy system is used, namely, (1)membership functions are made from the intervals of fuzzy labels of inflow and changing inflow as input variables, (2) from these membership functions, the optimum conformity value is determined by fuzzy reasoning, fuzzy composition and defuzzification, and then (3) the optimum release volume from reservoir is determined by membership functions from predicted inflow etc., as output variables and the conformity value calculated before. Block diagrams of the part of operational line and that of operational volume are shown in Figs 1 and 2. 2.2 Criterion of appraisal of the control system for reservoir operation of multipurpose dam The criterion for the fundamental appraisal of control system for reservoir operation are as follows. (1) to decrease the peak value of release discharge from reservoir. (2) to delay the beginning time of peak release discharge compared with that of inflow. (3) to make smooth the curve of release discharge. (4) to secure storage volume for effective use of reservoir as the lake. (5) to satisfy the actual operational rules of this multipurpose dam.
Rainfall River Discharge Inflow Predicted Inflow Changing Inflow Water Level Neurons to respond basin Neurons to respond inflow Neurons to respond the state of reservoir Neuron to respond the selection of release discharge, storage and preservation water level in reservoir Outflow Fig. 1 Block diagram of operational line ( Neural Networks ) Inputs Membership Function Fuzzy Reasoning Fuzzy Composition Optimum Conformity Value Outputs Fig. 2 Block diagram of operational volume ( Fuzzy System )
3 Application of this control system to the real dam The above-mentioned control system for reservoir operation is applied to the real multipurpose dam. The analysis has been carried out on the hydrologic data of the catchment area of the Nishi-Arakawa Dam (24.8 km 2 ). The volume of storage of this dam is 3.5 10 6 m 3. The utility object of this dam is multipurpose one. 3.1 Identification of parameters of both operational line and volume Parameters of two subsystems, that is,the operational line and the operational volume, are identified for controlling the reservoir operation of this multipurpose dam. 3.2 Operational line In the case of identification of the operational line for reservoir operation, the perceptron which belongs to the hierarchically connected type is used and then parameters of this control subsystem are identified through Backpropagation which is a systematic method for multilayer artificial neural networks. The squashing function is chosen to be the sigmoidal logistic function. This function is expressed mathematically as follows;
f(u)= 1 1+ exp(- u) where &wjxj-h I.1 in which xi is a set of inputs, w i is an associated weights and 0 is a threshold. In the case of the operational line, w and 8 of a part of artificial neuron obtained by this analysis is shown in Table 1. 3.3 Operational volume In the case of identification of this subsystem, the fuzzy system is composed of some fuzzy inference and the optimum fuzzy inference is determined from fuzzy plain explained (e.g. Hasebe[5]). This fuzzy plain represents the characteristics of fuzzy inference and the difference in the fuzzy reasoning. Though the judgment is difficult because there is no criterion of evaluation to judge rightly. From the viewpoint that fuzzy control is very consistent with human thinking, we judge that the smoothness of fuzzy plain is the best. The intervals of fuzzy labels as input variables are. determined from inflow, predicted inflow by filter separation AR method (e.g. Hino[7]) and fuzzy labels as output variables are determined from storage volume and release discharge. An example of two dimensional fuzzy matrix which represents the relation between fuzzy labels as input variables and those as output variables is shown in Fig.3. This fuzzy matrix is determined with reference to both the control rules for reservoir operation and the information obtained by inquires to expert reservoir operator. The intervals of membership functions as input variables are divided into four. On the other hand, those of membership functions as output variables to
Neuron to respond to the basin associated value threshold rainfall 0.8 river discharge 0.1 38 inflow 1.0 Neuron to respond to the inflow inflow 0.3 predicted inflow 0.3 2.1 changing inflow 1.0 Neuron to respond to the state of reservoir inflow 0.1 water level in reservoir 0.1 360 outflow 0.2 Neuron to respond to the release discharge response to the basin 0.2 response to the inflow 0.3 2 response to the state of reservoir 1.0 Neuron to respond to the storage response to the basin 1.0 response to the inflow 0.6 2 response to the state of reservoir 0.7 Neuron to respond to the water level in reservoir response to the basin 0.2 response to the inflow 0.5 2 response to the state of reservoir 0.8 Table 1 Values of associated weight and threshold
In the case of the low of water level in reservoir (1) Inflow is large Vertical axis is predicted inflow and horizontal axis is rainfall. Zero Small Big Negative Keep Zero Storage Positive (2) Inflow is middle Zero Small Big Negative Preservation of Zero water level Positive Storage (3) Inflow is small Negative Zero Positive Zero Small Big Storage Preservation of water level Fig.3 Fuzzy matrix
Membership function of inflow QS QM QB QVB 0 10 20 30 40 50 (m 3 /s) Membership function of predicted inflow QS : Small QM : Middle QB : Big QVB : Very big QPS QPM QPB QPVB 0 10 20 30 40 50 (m 3 /s) Membership function of changing inflow DQN DZ0 DQP QPS : Positive-Small QPM : Positive-Middle QPB : Positive-Big QPVB: Positive-Very-big DQN : Negative DZ0 : Zero DQP : Positive -10 0 10 (m 3 /s) Membership function of release discharge VS SM SMD MD BMD BG VB BG : Big VS : Very small SM : Small VB : Very big MD : Middle SMD : Small-Middle BMD : Big-Middle 0 2 4 6 8 10 12 14 (m 3 /s) Fig.4 Optimum membership functions
determine the release discharge from reservoir are divided into seven. The optimum structures of membership function is shown in Fig.4. 4 The results of analysis Here, the results, to be applied to the control system for reservoir operation of the multipurpose dam ant to be investigated, is explained. Comparison the resulted operated by expert operator with the simulated by this dam control system which is driven by both neural networks and fuzzy set theory is shown in Fig.5. In the case of this control system aided by fuzzy set theory, the fuzzy inference as shown in Fig.5 is applied to Min-Max type. From figure 5, it is understood that, as peak value of release discharge by neural-fuzzy system is almost equal, the effective storage volume which is inflow minus release discharge is many and the release curve of the discharge is smooth, the dam control by this neural -fuzzy system is better than that by actual reservoir operator for this multipurpose dam. Last, comparison effective storage volume in reservoir and peak value of release discharge, operated by the control system using both neural networks and fuzzy set theory, with those operated by actual expert operator is shown in Table 2. For reference, results to be controlled by only fuzzy system and examples of result of analysis of another floods are also shown in the same table. Consequently, it is suggestible that the dam control system for reservoir operation using both neural networks and fuzzy set theory is effective for the multipurpose dam where we have analyzed. key words: reservoir operation, fuzzy set theory, neural network system, dam control system, rainfall-runoff system. References
Peak release discharge Effective storage volume (m 3 /s) (10 4 m 3 ) Control system Simulation Actual Simulation Actual (Flood 1) Fuzzy system(only) 26 68.70 Neural networks and fuzzy system 30 35 28.08 5.04 (Flood 2) Fuzzy system(only) 8 27.30 Neural network and fuzzy system 12 15 25.56 4.32 Table 2 Values of peak release discharge and effective storage volume
[1] Clifford.L.(Editor); Neural Networks, Theoretical foundations and analysis, New York, IEEE Press,1993. [2] Kosko.B.,; Neural networks and fuzzy system, A dynamical systems approach to machine intelligence, New Jersey, Prentice Hall,1992. [3] Hasebe.M.,Kumekawa.T. and Satoh.T.; The application to the dam gate operation rule from a viewpoint of water resources using various reasoning method of fuzzy theory, Applications of Artificial Intelligence in Engineering VII,pp.561-578,1992. [4] Hasebe. M., Hino. M., and Hoshi.K.,; Flood forecasting by the filter separation AR method and comparison with modeling efficiencies by some rainfall-runoff models, J.Hydrol.,110,pp.107-136,1989. [5] Hasebe.M., and Nagayama. Y.,; Application of fuzzy set theory to the dam control system, Trends in Hydrology,pp.35-47,1994. [6] Hasebe.M.,Kumekawa.T. and Nagayama. Y.,; Reservoir operation for flood control by both fuzzy and neural networks system, Applications of Artificial Intelligence in Engineering XII, pp.221-234,1997. [7] Hino.M.and Hasebe.M.,; Identification and prediction of nonlinear hydrologic systems by the filter separation auto-regressive (AR) method: Extension to hourly hydrologic data, G.E.Stout and G.H.Davis(ed), Global Water; Science and Engineering, The Ven Te Chow Memorial Volume, J.Hydrol.,68,pp.181-210,1984.
[8] Philip D. Wasserman; Neural Computing, Theory and Practice, New York,VAN NOSTRAND REINHOLD Press, 1989.