Optimal Control of Combine Sewer Overflows (CSOs) Using Gradient-Based Neural Network Inversion with Projected Precipitation Inputs

Abstract Submission: Combined sewer overflows (CSOs), often caused by heavy rainfall, pose significant risks to public health and the environment, including drinking water contamination. Various strategies, such as Green Infrastructure, Integrated Planning, and Smart Control, have been developed to reduce the impact of CSOs. Smart active control strategies (e.g., efficient gate operation) are among the most cost-effective solutions that are gaining popularity. Traditional control methods rely on either expert knowledge (rule based etc.,) of the sewer system or optimization techniques (Evolutionary Algorithms etc.,) to determine the best operation schedules for the active control structures (e.g., gates and pumps) to minimize CSOs. These evolutionary optimization techniques, however, are non-gradient-based, making them less efficient. This limitation arises because the flow models used in these systems cannot track changes in the system (differentiations), particularly during the flow calculation process. One promising solution to this problem is the use of neural networks. Neural networks can be applied as predictive tools to estimate CSO volumes or rates, and their differentiability allows for the use of faster, gradient-based optimizers instead of non-gradient algorithms. This process is known as neural network inversion, enables more efficient and real-time optimization. While the use of neural network inversion in CSO optimization is still relatively new and limited, early implementations show promising results. Existing studies have relied on synthetic or limited historical data. Our work aims to replace synthetic data with projected future precipitation, providing a diverse range of precipitation patterns for training. That could potentially improve optimization in real-time control (RTC).

Learning Objectives/Expected Outcome (Optional) :