Développement de modèles d'apprentissage automatisé pour la prévision hydrologique d’ensemble au lac Saint-Jean

Abstract

The objective of this study is to develop and evaluate advanced artificial intelligence models to improve flood volume forecasting in the context of hydroelectric reservoir management. Focusing on the Lac-Saint-Jean watershed, this research aims to optimize the accuracy of 15 day hydrological forecasts by integrating ensemble-based approaches. To achieve this, it leverages historical and meteorological data from sources such as ERA5 and ECMWF, enabling better anticipation of hydrological variations associated with climatic events, particularly snowmelt and extreme precipitation. The adopted methodology is based on the comparison of two main architectures: a standard encoder-decoder model and a version incorporating an attention mechanism. These models are designed to capture complex hydrological dynamics through supervised learning. The evaluation relies on metrics such as the KGE and CRPS, as well as Talagrand diagrams to assess the calibration of probabilistic forecasts. Additionally, this study explores the impact of transfer learning, which involves pre-training the model on a long historical dataset before fine-tuning it with more recent data, thus enhancing generalization. The combination of the attention mechanism and transfer learning leads to a reduction in errors, with a KGE maintained at 0.91 on day 15. The effect of transfer learning is particularly pronounced during spring floods, when streamflow dynamics are more complex. The addition of the attention mechanism provides an additional gain, although its impact varies by season. It proves beneficial in winter and autumn, while it can increase forecast variability in spring and summer. Although the models perform well on average, Talagrand diagrams reveal a calibration issue, characterized by an under-dispersion of forecasts, especially as the lead time increases. This indicates that the models are overly confident, as they underestimate the uncertainty associated with their predictions. Overall, this study highlights the benefits of transfer learning and attention for hydrological forecasting, while also emphasizing the need to enhance the reliability and calibration of ensemble predictions, paving the way for future research.

Date	17 Apr 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor)