Revealing Deep Learning Model Preferences for Spatio‐Temporal Drivers of Runoff Forecasting: A <scp>SHAP</scp> ‐Based Comparative Study

ABSTRACT Accurate runoff forecasting is essential for flood prediction and disaster preparedness amid increasing hydrological extremes driven by climate change. While deep learning models offer high efficiency, most interpretability studies focus on single models. This limits understanding of how model architecture influences feature sensitivity and model applicability under different conditions, posing a challenge for developing robust urban flood forecasting systems. To address this issue, this study compares four deep learning models with a flood‐weighted loss for daily runoff forecasting: LSTM, CNN, Transformer and Informer, using SHapley Additive exPlanations (SHAP) to link predictions with local meteorological drivers at flood and non‐flood scales. Among the models, CNN is the only model that reproduces the most extreme peak, whereas LSTM captures many other high peaks. Transformer and Informer are more stable, and Informer best tracks temporal fluctuations. All models increase the importance of rainfall during floods and give more weight to temperature in non‐flood periods. Based on these patterns, we propose a framework for evaluating and selecting data‐driven models in urban flood prediction. This framework links forecasting objectives to suitable architectures and supports the development of adaptive, interpretable tools for real‐time flood forecasting and risk management in complex urban settings.