Urban drainage systems (UDSs) are critical built assets increasingly challenged by short-duration extreme rainfall, aging infrastructure, and rising surcharge risk. Physics-based hydrodynamic models are widely used for system assessment, but their high computational cost limits real-time operational prediction. Existing data-driven prediction approaches improve computational efficiency, but often rely mainly on sensor inputs and provide limited asset-level interpretation. This study develops an explainable digital twin for real-time prediction of storm-driven water level response in a separate sewer network in the Yangtze River Delta, China. The framework integrates 5 min monitoring and SCADA data, including water level, flow, pump status, and rainfall, with GIS and as-built asset information, including pipe geometry, hydraulic capacity, catchment characteristics, and network connectivity. A hybrid TCN-LSTM model was developed to predict water level and surcharge risk probability at 15–60 min lead times. A surrogate-based SHAP module was used to explain model predictions at the node and subcatchment scales. Multi-source fusion reduced the RMSE by approximately 18% compared with sensor-only baselines. The SHAP results showed that the pipe capacity-related variables and upstream contributing area were the main drivers of surcharge onset. The framework provides interpretable, operationally relevant predictions to support the resilience-oriented management of urban drainage systems.
Wang et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: