Résumé
Mediterranean basins are particularly exposed to flash floods caused by intense rainfall. Nîmes city, located in southern France, is frequently affected by these events, which cause significant damage and casualties. These floods are exacerbated by climate change and rapid urbanization. Addressing these challenges, decision-makers and managers require efficient flood forecasting models to optimize their crisis management strategies. However, traditional hydrological models often struggle to accurately predict such extreme events. In this context, artificial neural networks (ANNs) provide a promising alternative due to their ability to model complex nonlinear dynamics, making them particularly suitable for hydrological modeling and flood forecasting. In this study, a model based on multilayer perceptron was applied to simulate flash floods in the Alès Cadereau in Nîmes city, while extending the lead time. The KnoX method, a Physics Informed Machine Learning (PIML) method, was employed to extract internal information from the model, enhancing its interpretability and acceptability in operational contexts. The results demonstrate remarkable performance across various lead times, with criteria values indicating the reliability and accuracy of the proposed models. Contributions of input variables, calculated using the KnoX method, provide relevant and coherent insights consistent with the physical processes of the modeled basin. This approach highlights how prior knowledge can effectively guide ANN models to produce results in line with the physical processes of the basin.