Comparative Analysis of LSTM and XGBoost ML Models for Short-term Rainfall Forecasting

Abstract

Floods are the most widespread and frequent of all weather-related natural disasters. Accurate short-term rainfall forecasting is essential for flood management and disaster preparedness. Traditional models often struggle to capture the variability and nonlinear patterns of rainfall, and they can be computationally intensive. Machine learning offers a promising alternative due to its ability to learn temporal dependencies in time series data. This study investigates deep learning and tree-based approaches to rainfall prediction by comparing LSTM neural networks and XGBoost models, along with a combined ensemble model. Using data from the IAD Airport station in Houston, Texas, models were trained to predict the next hour’s rainfall using the previous 24 hours. The dataset used spans from 2012 to the present, with roughly the first 10 years used for training and the rest for evaluation. Evaluation was based on MAE, RMSE, and R². Results show that while all three models produced similar accuracy, with over 96% of predictions within ±1 mm of actual, the LSTM model achieved MAE/RMSE of 0.2287/1.5389 mm; XGBoost scored 0.2174/1.4622 mm; the ensemble outperformed both with 0.1944/1.3761 mm and an R² of 0.5222, an error reduction of about 12–15%. These findings suggest that ensemble learning improves short-term rainfall prediction and offers a promising real-time rainfall forecasting method. Future work could explore incorporating additional environmental variables such as humidity or wind to further enhance predictive performance under diverse meteorological conditions.