Interpretable deep learning for sewer network water level forecasting in a Northern Chinese City: Implications for enhancing real-time assessment of system operational conditions

Changchun City in North China. Water level prediction within sewer networks ‌plays a crucial role in‌ urban water system real-time control. With‌ advanced automated monitoring technologies, the generation of vast amounts of observational data ‌makes it possible‌ to predict water levels using deep learning models. However, a knowledge gap remains regarding how input feature dimensionality and temporal sequence length influence performances of such models. To address this issue, we developed an ‌explainable deep learning framework for predicting the influent water level at the terminal wastewater treatment plant of a typical sewer system by integrating Long Short-Term Memory (LSTM) with SHapley Additive exPlanations (SHAP) for interpretable analysis. Utilizing monitored precipitation data and upstream pipeline water level time-series data as input yields satisfactory one-step-ahead prediction performance. Contrary to expectations, incorporating additional upstream sensors did not improve model accuracy. Model performance improved progressively with longer input sequences, reaching optimal performance at 48-time steps before deteriorating with further temporal extension. SHAP analysis revealed that manual control operations exerted stronger influence on model outputs than the spatial proximity of monitoring sites to prediction points. Furthermore, sensor data anomalies and deficiencies in monitoring network configurations could be identified, confirming its utility for diagnosing behavioral patterns in modeled sewer network. These findings provided new hydrological insights and practical guidance for sewer system real-time control. • An interpretable deep learning framework for predicting sewer water level proposed. • Upstream sewer water level observations enable accurate one-step-ahead prediction. • Manual sewer operations outweigh monitoring site proximity in model output. • Longer input temporal sequences boost model performance until optimal length. • Data anomalies and monitoring network deficiencies detected by interpretable analysis.