Optimal Control of Combined Sewer Overflow (CSO) Using Gradient-Based Neural Network Inversion with Projected Precipitation Inputs

Combined sewer overflow (CSO) modeling using future rainfall is challenging due to the projected variability and limited temporal data resolution. Precipitation variability is crucial for storm sewer system optimization, as gate and pumping operations are highly sensitive to precipitation patterns. Existing methodologies often struggle to capture these complexities, primarily due to the lack of fine temporal resolution rainfall data from contemporary climate models. This study integrates high-resolution, 15-min climate model precipitation data with the recently proposed neural inversion optimization technique, a gradient-based method. Our findings demonstrate the effectiveness of the approach in handling diverse precipitation patterns. The neural inversion optimization technique reduced CSO volume by up to 49.1% compared to the unoptimized baseline. Mass balance analysis also showed a significant improvement in storage utility, increasing from 39.8% to 94.5%. This study highlights the potential of integrating complex precipitation data with machine learning optimization, offering a promising prototype for CSO management.