Regulation of Great Lakes Reservoirs System by a Neuro-Fuzzy Optimization Model

Great Lakes reservoirs system is a complex natural system containing alarge percentage of the fresh water resources of the world. It is located in Canada and U.S.A. serving about 40 Million people and is managed by an International Joint Commission made up of engineers from these two countries. Management of this system is still based on rule curves and much more could be done to improve this situation. The system is complex also due to highly differing scales of variables, nonlinearity, uncertainty, and dimensionality. An implicit stochastic method is applied using successive approximation optimization to obtain the optimal state and control variables of the reservoirs using the 90 years monthly data. However, when simulating the derived policies a re-optimization in each time period is needed due to inequalities and nonlinear relations existing among variables. The optimal values obtained from simulation are used as input-output data in training an Adaptive-Network-based Fuzzy Inference System (ANFIS) model for one of the months that required a non-constant release policy. ANFIS derives the general operating rules of the reservoirs in the form of fuzzy "if-then" rules. The parameters of a Sugeno-type Fuzzy Inference system (FIS) are optimized through an Artificial Neural Network (ANN) using back-propagation learning algorithm and least square method. The model of our system is anticipatory in nature given the fact that we base our current decision from the expectation of a future state.In this paper, we discuss the various aspects related to our implementation and the computational issues. Simulation of operating policies obtained from the ANFIS model, and comparison of its performance with other policies shows the potential capability of the proposed approach to tackle optimal operations of the system.