AI-assisted framework using physically informed rainfall–drainage features for real-time urban flood risk forecasting

Urban flash flooding is becoming more severe as urban population density increases, and severe and frequent floods occur. As a result, early warning systems are required that translate raw sensor data into clear and actionable hazard situations, rather than simply providing ongoing forecasts. For this purpose, this study develops a hydrological- and hydraulics-informed framework for real-time multi-class AI-based flood warning across evaporation (low risk), drained (medium risk), and flooding (high risk) states. The methodology firstly derives physically guided rainfall features using a rule-based back-propagation neural network to estimate return-period signals alongside seasonality and antecedent-rainfall cues, while hydraulics-informed “water-level memory” (current class and class duration) captures system dynamics. These inputs feed seven weak-learner families whose outputs are fused by a time-series mixture-of-experts. The model performance is evaluated for lead times of up to 5 h using both multi-step metrics and event-based analyses. The framework is applied to the Ruislip urban drainage system, UK, using long-term IoT rainfall and water-level records (2011–2024). The results show discrimination improved compared with voting and averaging ensembles. It also increases hit rates for flood and non-flood decisions at 1–3 h. At 4–5 h, it reduces false alarms and late detections. Event-based results show more on-time hits and shorter timing lags. Feature analysis shows that rainfall intensity and duration are the main drivers. Seasonal effects and antecedent occurrence also provide added value. Residual errors concentrate at longer horizons e.g. 5 h later where transitions between adjacent states are intrinsically difficult.