Comparison of Stochastic Optimization Algorithms for Hydropower Reservoir Operation with Ensemble Streamflow Prediction

Abstract

Stochastic optimization methods have been developed over the last few decades to help water managers who are regularly confronted with making complex decisions about reservoir releases in the context of streamflow uncertainties. However, a comparative evaluation of the performance of the methods in an operational context is not an easy task, which makes it difficult to select the approach that offers the best performance. This paper presents a comparison between four optimization algorithms in a test bed in which ensemble streamflow predictions (ESPs) are updated each time a decision is taken. The comparison was performed on the Rio Tinto Alcan (RTA) hydropower system in Québec, Canada, which consists of six generating stations in series and three major reservoirs. The tested optimization algorithms are the deterministic optimization approach currently used by RTA and three explicit stochastic optimization approaches, i.e., stochastic dynamic programming, sampling stochastic dynamic programming, and a scenario tree approach. The results showed that methods on the basis of scenarios prove superior to methods on the basis of probability distributions. Moreover, using an anticipative deterministic approach to calculate the release decisions for the first period was found to be an inadequate strategy. Artificially introducing underdispersion in ESPs was also found to affect the quality of the results, and the optimization methods were affected differently. Given that hydrological dispersion will likely differ in the future as a consequence of climate change, further evaluation of optimization techniques should be carried out before selecting approaches that best meet managers’ needs in a climate change context.