Exploring the Potential for Machine Learning-Based Flow Predictions in Sewer Systems

Abstract

This study explores the potential of machine learning (ML) for predicting flow in sewer systems, using data generated by the Storm Water Management Model (SWMM). A Long Short-Term Memory (LSTM) network, chosen for its effective handling of time series data, was trained using both hypothetical and real rainfall data. The final model achieved a mean error of 4.6 % in predicting peak flows and demonstrated prediction times that were at least 3 times faster for individual events and up to 600 times faster in mass simulations, compared to an equivalent hydrodynamic model. Tested with 5-fold cross-validation, the model exhibited significant improvements in accuracy and speed compared to its initial version, largely due to enhanced complexity in the model architecture. However, when applied to a more complex sewer system, a decrease in accuracy was observed, underscoring the need for further validation with real-world data. These findings illustrate the promising potential of ML models to boost real-time prediction efficiency but also highlight the need for model adaptation in more complex scenarios.