Enhancing Flood Prediction Using Hybrid LSTM-Transformer Deep Learning Approach

Abstract

Flood prediction is crucial for effective disaster management, yet it remains a complex challenge due to the nonlinear nature of meteorological processes. This study develops and evaluates a novel hybrid model that integrates Long Short-Term Memory (LSTM) networks and Transformer attention mechanisms to enhance predictive accuracy for rainfall-based flood forecasting. Using extensive Australian weather data collected from 49 stations over a decade (2007-2017), the model incorporates comprehensive feature engineering, including derived meteorological indicators, rolling statistical measures, and temporal lag features. The hybrid LSTM-Transformer architecture achieved superior precision (77.69%) and high accuracy (84.57%) compared to a Random Forest baseline model. Confusion matrix analysis illustrated the hybrid model’s strength in reducing false alarms, indicating a conservative yet highly reliable predictive performance. Feature correlation analysis revealed important relationships among temperature, humidity, pressure, and rainfall, highlighting the complexity of meteorological interactions. The findings demonstrate the effectiveness of integrating sequential and global temporal modeling for flood prediction, providing valuable guidance for operational forecasting systems and disaster preparedness strategies. This research contributes significantly to existing flood forecasting methodologies and suggests promising directions for future enhancements.