Multi-Objective Forecast-Informed Reservoir Operation at different time scales using a Neuro Evolutionary algorithm

The growing global demand for water, energy, and food, driven by rapid population and economic growth and a changing climate, is placing unprecedented pressure on water resources. Recent advances in forecasting water availability at various temporal and spatial scales offer an opportunity to improve reservoir operations by incorporating hydrometeorological forecasts into traditional information systems, helping operators prepare for extreme hydrologic conditions. However, selecting the optimal forecast products from a wide range of options with distinct biases and lead times is challenging. In multi-purpose systems with conflicting water uses, processes with varying dynamics and predictability levels further complicate this selection, as the optimal forecasts depend on the trade-offs between diverse operating objectives. To address these challenges, this study integrates an Input Variable Selection routine into a multi-objective Direct Policy Search framework using Reinforcement Learning, dynamically determining the optimal forecast inputs during the learning process. The framework employs the Neuro-Evolutionary Multi-Objective Direct Policy Search (NEMODPS) algorithm, which uses evolutionary strategies to adapt policy shapes and inputs to develop Pareto-optimal policies. Forecast inputs include medium- to long-term ensemble forecasts from the European Flood Awareness System (EFAS), known for their accessibility despite some biases, and deterministic, locally calibrated short-term forecasts from the local consortium. This approach is applied to the Lake Como case study in Northern Italy, where the lake is regulated for flood control, water supply, and low-level avoidance. This work provides two primary contributions. First, we present a refined version of NEMODPS that significantly improves computational efficiency and modularity for faster and more complex policy optimizations. Second, we integrate real ensemble forecasts into the decision-making process, rather than deterministic perfect forecasts used in previous studies, offering a more realistic analysis that accounts for forecast biases. These improvements show clear benefits in computational performance and highlight the practical utility of the updated NEMODPS in forecast-informed reservoir operations.