Physics-guided soft sensing and inflow forecasting for urban drainage pump stations

Accurate inflow forecasting is critical for urban drainage pump-station control, yet it is hindered by limited direct flow monitoring and strong hydraulic non-stationarity caused by frequent pump switching. This study aims to develop a physics-guided forecasting framework that improves both inflow reconstruction and short-horizon prediction under mixed stationary and transient operating regimes. A mass-conservation-based soft-sensing method with hierarchical denoising was first developed to reconstruct high-quality inflow labels from noisy Supervisory Control and Data Acquisition variables. On this basis, a hybrid forecasting architecture, termed dynamic dilated temporal convolutional network–long short-term memory–attention, was constructed. The model decouples the input into trend and event features, uses a long short-term memory branch to encode slowly varying hydraulic trends, employs a dynamic dilated temporal convolutional network to capture abrupt transient disturbances and integrates both branches through a temporal attention-gated fusion mechanism. Experiments on minute-resolution Supervisory Control and Data Acquisition data from a large-scale urban drainage pump station showed that the proposed framework achieved the best overall and transient prediction performance among the compared models. Relative to baseline methods, the proposed model reduced transient mean absolute error by 18.3% compared with standalone long short-term memory and by 35.0% compared with DLinear, while also achieving the best global accuracy. The proposed framework provides a physically grounded and operationally relevant solution for inflow forecasting in urban drainage systems. By combining physics-guided soft sensing with adaptive multi-scale sequence modelling, it improves both average trajectory tracking and response fidelity under pump-switching disturbances.