Water Depth Prediction in Stormwater Sewers Using Deep Learning Method

Accurate prediction of water depth in urban drainage systems is essen-tial for effective flood risk mitigation and sustainable urban water management. This paper presents a deep learning model based on Long Short-Term Memory (LSTM) architecture for forecasting water depth in a stormwater sewer, tested on a part of Belgrade’s stormwater system in Serbia. The model uses high-resolution rainfall time series and hydraulic simulation data from SWMM model to predict water levels with 5-minute temporal resolution. A comprehensive hyperparame-ter sensitivity analysis was conducted using the One-at-a-time (OAT) method to systematically optimize the model architecture and training parameters. The re-sults demonstrate that the optimized LSTM model achieved a coefficient of de-termination (R²) of 0.95 for 5-minute-ahead predictions, showing strong perfor-mance in short-term forecasting while maintaining computational efficiency. The study identifies the optimal network configuration with 128 neurons, ReLU acti-vation function and learning rate of 0.001 after testing and validation. These find-ings highlight LSTM's superior capability to model complex rainfall-runoff rela-tionships in urban environments, providing valuable tools for real-time storm-water management and enhanced flood early warning systems.