Séminaire du Lagea ENSTP

08/05/2026

Séminaire du mois d'Avril 2026

08/05/2026

Séminaire du mois de Mars 2026

08/05/2026

Deuxième Séminaire du mois de Février 2026

08/05/2026

20/04/2026

Séminaire Scientifique-Equipe ECO (Génie de l'Eau, Climat et Océanographie) du LaGEA/ENSTP/UNSTIM

Email : [email protected]

Jeudi 23 Avril 2026 à partir de 19h


Titre : Optimisation multi-modèles pour la prévision.

Par : HOUNNONDAHO Z. Freddy (LaGEA/ENSTP & ED-STM/UNSTIM)

Pour participer au séminaire sur Google Meet, cliquez sur ce lien :
https://meet.google.com/swz-qdpf-xeh


Abstract:
The Ouémé basin, in Benin, is exposed to increasing hydroclimatic variability that weakens hydrological systems and amplifies flood risks, particularly at the Bonou outlet. In this context, improving the reliability of streamflow forecasts is a major challenge. This study proposes a multi-model approach aimed at reducing the uncertainties associated with the use of a single hydrological model. Three conceptual models (HBV, GR4J and SAC-SMA) were thus calibrated over seven sub-basins of the Ouémé using the Differential Evolution and L-BFGS-B algorithms, testing three objective functions (NSE, logNSE, KGE). Several combination strategies were then compared: the corrected GrangerRamanathan method (GRC), a Random Forest Regressor, and a hierarchical semi-distributed approach implemented through the airGRiwrm package. Several relevant criteria (NSE, KGE and CRPSS for the probabilistic dimension) were used to assess the robustness of our approach. The three models adequately reproduce the seasonal dynamics, with station-dependent performance (KGE around 0.85), and KGE emerges as the best calibration criterion. Multimodel combinations consistently outperform individual models: the Random Forest better captures flood peaks and seasonal transitions, while the GRC performs better in terms of interpretability. At Bonou, the combination markedly improves the representation of extreme events. The multimodel approach therefore constitutes an effective lever for operational hydrological forecasting in tropical basins, reconciling statistical robustness with process representation.