Apprentissage profond appliqué à la prévision saisonnière des volumes d’apport dans les réservoirs hydroélectriques

Abstract

Hydropower reservoirs require inflow forecasts to allow water resources managers to optimize drawdown rates and improve infrastructure efficiency. Usually, operators use physically-based or conceptual hydrological models to forecast streamflow for a lead-time ranging from one month to a year, depending on the planning horizon. These forecasts are then aggregated to evaluate cumulative inflow volumes, which are essential for seasonal operations. Reliable medium-range volume forecasts over a few weeks support not only the optimization of reservoir filling and hydropower generation, but also preparedness for seasonal floods and low-flow periods. To address these needs more directly, this study presents a novel method that directly forecasts 45-day cumulative inflow volumes using a deep learning model trained on ensemble meteorological forecasts and observed volume data, avoiding the need to simulate daily streamflow and targeting the total volume over the full 45-day period in a single operation. Furthermore, the model leverages large-scale datasets during its training, by using data from 200 basins in Canada. The model is then applied to eight Canadian basins to estimate inflow volume forecasts. For the Lac-Saint-Jean hydroelectric system in Quebec, the performance of LSTM networks, a type of recurrent neural network designed to capture temporal dependencies, is also compared with an operational conceptual hydrological model. The results show that forecasts produced by LSTM models are generally accurate but lack reliability. Ensemble forecasts tend to be overconfident and underdispersed. Regional training improves performance in some cases, but no universal pattern emerges across all seasons or basins. Compared to the operational conceptual model, LSTM models outperform it during the autumn season, both in terms of accuracy and reliability, highlighting their complementary potential. Additionally, the direct training of a locally trained LSTM model using ensemble meteorological forecasts helped assess the limitations of this approach. Overall, this study suggests that LSTM-based models offer a promising alternative for strategic reservoir planning, particularly when combined with physically based hydrological models to improve uncertainty representation.

Date	4 Dec 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor) & Jean-Luc Martel (Co-supervisor)