Abstract Submission: Flood management and Combined Sewer Overflow (CSO) predictions are critical for safeguarding communities and the environment. This study introduces two revolutionary surrogate models powered by advanced machine learning for Innovyze InfoDrainage and InfoWorks ICM software.
The first model, targeting flood management, employs a Bayesian Convolutional Neural Network (B-CNN) trained on extensive InfoDrainage simulations. It predicts the spatial distribution of water at given points in time within milliseconds, offering speed and accuracy. By integrating physical data, it identifies flow patterns and forecasts water depth during the construction phase, understanding the impact of surface structures like roads and airports in real-time. This technology serves as an optimal flood management tool and a real-time forecasting tool, providing advanced warnings and enabling proactive measures. It achieves global applicability through restricted network & boundary parameterization.
The second model predicts CSOs, essential for environmental protection and regulatory compliance. Using the InfoWorks ICM simulator and differentiable physics approach as a physics-informed neural network (PINN) , this machine learning workflow builds a surrogate model that can quickly and accurately forecast the temporal distribution of water at given points in space for various rainfall profiles. Achieving the same accuracy as traditional simulators but in under a minute, it allows for timely interventions, mitigating untreated sewage release and preventing water pollution, health hazards, and ecosystem damage. It can represent any degree of network / boundary complexity but requires site-specific training.
Both models highlight the potential of machine learning-based surrogate models to enhance decision-making, reduce risks, and protect communities and the environment from the impacts of flooding and CSOs.
Learning Objectives/Expected Outcome (Optional) : Learn about the development and application of Bayesian Convolutional Neural Networks (B-CNN) and differentiable physics as a physics-informed neural network (PINN) in flood and CSO management. Explore the integration of physical data with machine learning models to forecast water distribution and flow patterns in real-time. Ability to explain the benefits of using advanced machine learning surrogate models for flood and CSO management. Awareness of the potential for machine learning-based surrogate models to improve decision-making and proactive measures in environmental protection and regulatory compliance.
