Advancing the operation of multipurpose water reservoirs with multi-timescale forecasts: Application to Lake Como

Water management can benefit from recent advances in hydroclimatic forecasting, like the growing accuracy of forecasts over longer lead times. However, it is critical to understand how to use the wealth of information available over multiple timescales in the most effective way. For water system control, hydrological forecasts can be used with two approaches, either an on-line one, which acquires the forecasts in real-time in an optimization routine at each control time step, or a data-driven off-line one, learning a policy that includes forecasts as inputs based on historical reforecasts. In this study, we compare the performance of an on-line and off-line control scheme to understand how multi-timescale forecasts can be best used to inform the optimal operation of multipurpose water reservoirs. The on-line control scheme is a new nested multi-stage stochastic Model Predictive Control (MPC) framework integrating the use of forecasts over multiple timescales. The MPC framework can use multiple forecasts proactively and satisfy competing objectives but is affected by the forecast bias and optimization challenges arising with longer control horizons. Off-line control schemes like Direct Policy Search (DPS) can help overcome these challenges. The off-line framework combines an Input Variable Selection algorithm, for extracting the most informative policy inputs from a set of forecast variables over different lead times, with the Evolutionary Multi-Objective DPS method, for designing Pareto-optimal control policies conditioned on forecast information. We test the performance of these two control schemes for the Lake Como system in Northern Italy, where a large lake is regulated mainly for irrigation supply and flood control. The forecasts used include deterministic and ensemble hydrological forecasts over multiple timescales from short-term (60 h) to sub-seasonal and seasonal lead times (7 months). The performance of the two forecast-informed operating policies is compared with different benchmarks, including the historical management with no forecasts and a perfect operating policy. Beside improving the policy performance with respect to single timescale forecasts, results provide insights on the role of the forecast bias and uncertainty at different timescales in reservoir policy design.