Testing Stochastic Dynamic Programming Models Conditioned on Observed or Forecasted Inflows

Abstract

This paper presents and compares four types of stochastic dynamic programming (SDP) for on‐line reservoir operation, relying on observed or forecasted inflows. The models are different because of the assumptions regarding the inflow in the next time period. If this inflow is known (or a forecast is possible with 100% reliability) models with expected value of the future returns are possible (present returns are deterministic). Otherwise, a simple forecast based on conditional probabilities is necessary, and present and future returns are random. The objective is to maximize expected annual hydropower generation. In a case study of the Feitsui Reservoir in Taiwan, SDP models appear to provide efficient long‐term operating policies. The simulation of on‐line operation of the reservoir reveals that the SDP model that relies on the observed inflows of the preceeding time step provides the best performance. Nevertheless, under different hydrological regimes this finding might be not universal, but dependent upon the characteristics of the particular water resources system.