Enhanced LSTM Model for Flood Forecasting Systems: A Case Study of the Piracicaba River Basin in Brazil

The heavy rains and floods that struck southern Brazil in May 2024 highlighted the vulnerability of the population to extreme hydrological disasters, resulting in 176 confirmed fatalities, around 40 people missing, and over 422,000 individuals displaced. Flood Early Warning Systems (FEWS) are crucial tools for reducing flood-related damage and fatalities. The accurate prediction of flood peaks and their timing is essential for effective evacuation planning. This study proposes an enhanced Long Short-Term Memory (LSTM) model for runoff prediction, incorporating novel loss functions that prioritize flood periods (e.g., peak flow and peak timing) , while reducing the importance of normal and low flow periods. Additionally, the study evaluates the model’s performance across various forecast horizons (0 – 24 hours), aiming to understand how forecast accuracy varies with increasing forecast horizons. Using Piracicaba City in Brazil as a case study, 10-minute flood stage and rainfall data from 18 upstream stations (2018–2023) were utilized to predict flood stages at the target station using the enhanced LSTM model. The results compare LSTM predictions with traditional loss functions (e.g., mean absolute error) to those using the newly designed loss functions, evaluated through several metrics to assess improvements in flood prediction accuracy. By analyzing the model’s accuracy across different forecast horizons, the study provides valuable insights into the optimal lead time for issuing warnings in Flood Early Warning Systems.