Utilisation d’un modèle de mémoire long-court terme entrainé sur des rétro-prévisions pour la prévision des apports en eau dans le réservoir hydroélectrique du Lac Saint-Jean

Abstract

The rise of artificial intelligence (AI) is transforming numerous key sectors of our society, including hydrology. Among the most impactful and promising advances is deep learning, a branch of AI. This study leverages a deep learning (DL) model to forecast cumulative water inflows over a 14-day period into a hydroelectric reservoir supplied by the Lac Saint-Jean watershed. The model employed is based on Long Short-Term Memory (LSTM) networks, which have recently gained attention for their strong performance in various hydrological forecasting applications. Additionally, meteorological reforecasts are integrated into the training of some models to assess their influence on model performance and the quality of volumetric forecasts. The results show that LSTM models generate high-quality forecasts while highlighting specific benefits and limitations. Overall, the ensemble forecasts produced by the models are accurate but lack reliability, exhibiting persistent under-dispersion across all seasons. The models notably struggle to accurately capture the timing and magnitude of the spring freshet, frequently failing to predict extreme flows. Furthermore, the research indicates that incorporating reforecasts during training enhances forecast precision and sharpness but at the expense of robustness. In addition, a seasonal evaluation was conducted to analyze model behavior under the unique hydroclimatic conditions of each season. Model performance is more limited in winter and spring, where complex hydrological conditions, such as snow presence and rapid flow variations, pose additional challenges. During these seasons, models trained with reforecasts outperform the baseline model in terms of accuracy. Conversely, summer and autumn forecasts are of higher quality, but the use of reforecasts during training tends to reduce ensemble precision. Regarding reliability, significant under-dispersion is observed across all models. However, the baseline model without reforecasts demonstrates greater overall robustness, particularly during summer and autumn. The operational relevance of reforecasts thus depends on the specific forecasting objectives, such as managing extreme hydrological conditions.

Date	15 Apr 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor) & Jean-Luc Martel (Co-supervisor)