Forecasting and optimization for minimizing combined sewer overflows using Machine learning frameworks and its inversion techniques

Abstract

Combined sewer overflows (CSOs), which typically occur during heavy rainfall events, pose significant threats to both public health and the environment. These threats encompass various concerns, including contamination of drinking water. Numerous studies have developed strategies aimed at mitigating the adverse effects of CSOs. These strategies include Green Infrastructure, Integrated Planning, and Smart Control Strategies. Among these, Smart Control Strategies have gained the most traction due to their exceptional cost-effectiveness. Nevertheless, the existing control methods face a challenge in striking the right balance between precision and computational efficiency. While employing full numerical methods as predictive models can provide high accuracy, they often prove inefficient in terms of runtime, especially when applied to real-world complex combined sewer systems. Conversely, reduced-order models tend to offer quicker results but may sacrifice accuracy. To address this issue, we propose an exploration of various mainstream machine learning models for CSO predictions. Additionally, we introduce a novel approach known as “inversion of neural networks” to bridge the gap between prediction and optimization. This innovative method enables us to use a single neural network for both CSO prediction and optimization tasks, resulting in a significant enhancement in terms of computational efficiency. The accuracy of our predictive approach has been validated through simulation results. In terms of optimization performance, it provides similar outcomes to the genetic algorithm, while significantly improving computational speed.