Deep Learning for Downstream Water Level Prediction in Complex Hydrology Systems: An LSTM Approach

Flood forecasting is an essential tool for mitigating flood risks and enhancing disaster preparedness, particularly in vulnerable urban and agricultural areas where timely and accurate predictions can significantly reduce damage and ensure public safety. This paper presents a comprehensive approach to forecasting the water level of downstream network systems, including reservoirs and irrigation districts, during the flood season. The proposed methodology utilizes the Long Short-Term Memory (LSTM) deep learning technique to accurately predict flood water level patterns. Historical water level and rainfall data from a network comprising four rainfall gauging stations with long-term records and one water level gauging station within the case study area are processed and analyzed to improve prediction accuracy. The results demonstrate that the proposed LSTM model with SGDM technique effectively captures regime dynamics, providing valuable insights for decision-makers in flood risk management. The model achieves high accuracy in predicting flood peaks, both in magnitude and timing, with determination coefficient ？？2 values of 0.92 for 24-hour forecasts and 0.81 for 48-hour forecasts. The Root Mean Square Errors (RMSEs) for the entire flood season are 0.23 m for 24-hour forecasts and 0.35 m for 48-hour forecasts, within a water level range of approximately 2.0 m to 8.0 m during the intentional flood period at the control stations. The forecasted flow accuracy P are 84.85% and 87.60% for 24-hour and 48-hour forecasts, respectively. These findings highlight the potential of the LSTM model to enhance forecast performance, contributing to more efficient and sustainable water distribution systems while improving flood risk management practices.