Comparing Model-Based and Model-Free Streamflow Simulation Approaches to Improve Hydropower Reservoir Operations

Abstract

This paper presents a comparison between an inflow model-based and an inflow model-free optimization method applied to a hydropower system. Widely used stochastic dynamic programming (SDP) and the evolutionary multiobjective direct policy search (EMODPS) methods are used, respectively, as model-based and model-free methods. Main results show that the model-free approach provides a better representation of the complex inflow correlations. Stochastic dynamic programming suffers from the temporal decomposition of the problem that allows only autoregressive exogenous (ARX) models to be used. However, because inflow uncertainty is implicitly represented through simulation with EMODPS, long time series are required to accurately characterize the probability of all possible events. To tackle this drawback and to avoid only learning the data set, a new regularization framework is introduced to improve the policy robustness on unseen data sets. Moreover, this study highlights how the preselected family of functions defining the policy can reduce the performance of the EMODPS method. This study is based on the real-world case of Kemano, located in British Columbia, Canada. This system is challenging because of the long-term inflow correlation due to long snow accumulation periods and its multiple objectives.