Real-time reservoir optimization using ensemble streamflow forecasts

The sequential nature of reservoir operating decisions and the variability of streamflow makes Stochastic Dynamic Programming an attractive optimization procedure for reservoir system operations. This study examines the use of Sampling Stochastic Dynamic Programming (SSDP) for single- and multi-reservoir system operation. SSDP models based on historical streamflows and snowmelt volume forecasts are compared to SSDP models employing the National Weather Service's (NWS) Ensemble Streamflow Prediction (ESP) forecasts. The SSDP optimization algorithm, which is driven by individual streamflow scenarios rather than a Markov description of streamflow probabilities, allows the ESP forecast traces to be employed intact, thus taking full advantage of their stochastic and statistical characterization of streamflow. This study demonstrates that the use of frequently updated ESP forecasts in a real-time SSDP optimization for a reservoir system can provide more efficient operating decisions than an SSDP model employing historical time series data and current snowmelt volume forecasts. Both models were driven by an appropriately weighted and representative subset of the original forecast and streamflow samples. This study further demonstrates that a simplification of SSDP which forms a simpler two-stage optimization model provides operations that are as efficient as those provided by the complete SSDP model. Both a single reservoir and a multi-reservoir system that provides water supply and hydropower were evaluated with various SSDP models.