Optimisation de la prévision hydrologique d'ensemble par l'intégration de l'apprentissage profond dans une méthode multi-modèle hybride

Abstract

The recent application of deep learning in the field of hydrology has shown promising potential. This research focuses on the development of a hybrid multi-model approach, implementing eight lumped and conceptual hydrological models, a semi-distributed model, and a deep learning (DL) model, for forecasting inflows to a hydroelectric reservoir. The introduction of the Long Short-Term Memory (LSTM) model within this multi-model framework is particularly highlighted for its potential to improve the accuracy of short-term flow forecasts in the Lac-Saint-Jean watershed. The results show that the combination of hydrological models with LSTM significantly improves forecast performance, especially for short-term forecasts up to 9 days, outperforming other multi-model combinations by reducing, on average, up to 45% the CRPS values. The study also emphasizes that deep learning, when combined with other modeling approaches, can achieve maximum performance, particularly in terms of forecast distribution where ABDU scores decreased by 64%. The hybrid multimodel method proves effective in capturing hydrological variability at the watershed scale. In addition, a section of the thesis focuses on the operationalization of these tools and presents the application of this method in a practical context, particularly for improving hydrological forecasting systems at Rio Tinto. Using the CEQUEAU7 hydrological model, already in place at Rio Tinto, the study proposes a simple integration method, consisting of adding and merging complementary hydrological models to the existing system. This approach increases the robustness and accuracy of the forecasts. The inclusion of the LSTM model, in particular, has shown a significant improvement in forecast accuracy, producing on average CRPS values 50% lower, confirming the effectiveness of deep learning in better capturing the complex dynamics of hydrological systems. The integration of the GR4J model with LSTM offers the best performance, illustrating the effectiveness of a multi-model approach that leverages the specific strengths of different hydrological models for more accurate and robust forecasts.

Date	12 Feb 2024
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor)