Optimization of the operation of multireservoir systems, a Great Lakes case study

Abstract

In order to recognize the intrinsic stochastic features of the natural inputs and take them into consideration explicitly, Stochastic Dynamic Programming is employed to generate long-term operation policies. Because of the well-known "curse of dimensinoality", that can affect the optimization of large systems, the technique called Multi-Level Approximate Aggregation/Decomposition - Stochastic Dynamic Programming (MAM-SDP) Methodology is employed. The performance of this technique can be enhanced by using the suggested alternate approximation to the conditional distribution of the releases from the reservoirs. So far, MAM-SDP performs physical diagnosis to determine the aggregation scheme. A means of applying Principal Components Analysis is presented, therefore adding a different perspective to solving the problem, i.e., a statistical decomposition.

This work also aims at obtaining the relationship performance of the system versus its respective variance. To this effect, an extension of the Expected Return-Variance of Return Rule was developed, applied to a multistage decision type of problem. This technique was called Two-Pass Mean-Variance Approach. The algorithm for doing so is described. It was possible to show a significant range of performances at the variances associated with them for the operation of reservoir systems.

This work ends with the application of the techniques above mentioned to a real case study. In it, an alternate closed-loop type operation policy is presented for the North American Great Lakes. These policies and those from the Two-Pass Mean-Variance Approach are then compared with the ones obtained from a simplified model of the actual operation, based on heuristics. Two sets of synthetic Net Basin Supplies for the five lakes are studied, generating two different sets of release policies.