From forecast-informed reservoir operations to integrated forecast-control design

Seasonal streamflow forecasts have been proven effective and are becoming widely used to improve water reservoir operations, especially in areas where climate teleconnections enable predictability on medium and long lead times. Most existing studies have focused on the assimilation of forecasts for informing operational decisions—an approach that typically banks on pre-developed forecasts to optimize water release decisions. However, such approach may overlook the potential synergies that stand in co-developing the forecast and decision-making models: in other words, a seamless design integrating both forecast and system operation models has not been explored yet. In this work, we contribute a novel approach building on the Evolutionary Multi-Objective Direct Policy Search algorithm to design forecast and operation models together, with the ultimate goal of improving the performance of water reservoir controllers. The proposed approach is benchmarked against closed-loop policies not informed by any forecast as well as forecast informed policies relying on data-driven or perfect seasonal inflow predictions. Numerical experiments are conducted on the Angat-Umiray water resources system, Philippines, which is operated primarily for ensuring municipal water supply to Metro Manila and irrigation supply to a large agricultural district. All forecast models rely on data-driven (linear regression) models linking large-scale climate drivers (e.g., ENSO) and local hydro-meteorological conditions (e.g., antecedent inflow) to future monthly inflows. Our results show that solutions informed by seasonal forecasts outperform the no-forecast baseline policies. Importantly, the integrated design of forecast models and control policies provides an additional gain with respect to solutions informed by pre-designed forecasts. This result is particularly interesting, because the accuracy of the integrated forecast models is lower than the pre-developed ones, thus demonstrating that more accurate forecasts do not necessarily produce better water system operations. Lastly, we notice that the differences between forecast models change over time depending on the underlying hydrological conditions, with the bias of the integrated forecast models that also vary along the Pareto front according to the operators’ preferences.