Spatial heterogeneity identification for rainfall-derived inflow and infiltration in urban sewer systems based on water level sensor networks: Insights from an interpretable deep learning method

Due to the spatial heterogeneity of rainfall-derived inflow and infiltration (RDII) in urban sewer systems, accurately identifying the subcatchment with the most severe RDII is crucial for the subsequent management of the sewer system. With the increase in monitoring sensors in sewer systems, it is an urgent problem to mine information from the rich water level data to identify subcatchments with severe RDII. Traditional modeling methods either rely on large amounts of data related to water quality and flow, or struggle to capture the topological patterns and complex nonlinear relationships in sewer systems. Thus, this study proposes a severe RDII subcatchment identification method based on the water level sensors and interpretable deep learning algorithm. First, the sewer system is divided into several subcatchments based on the locations of the sensors. Deep learning prediction models for each subcatchment in dry and wet weather are developed using low-cost water level monitoring data. The deep learning models are subsequently analyzed using explainable artificial intelligence (XAI) method to identify RDII severity. After the method was tested in the case study, the developed deep learning models were proven to capture the response relationships between different monitoring sites and external rainfall. Moreover, the results show that the utilization of simple water level sensors, combined with advanced interpretable deep learning algorithms, can effectively identify different degrees of RDII in each subcatchment. This work provides a feasible method for spatial heterogeneity identification for RDII based on low-cost measurements that reduces monitoring and modeling costs, and has the potential for widespread application.