Reducing Uncertainty in Monthly Rainfall Forecasting Using Machine Learning and Ensemble Models in Semiarid Regions

Reliable rainfall forecasting is a critical component of water resources management, particularly in semiarid regions characterized by strong climatic variability and recurrent extremes. This study evaluates the performance of classical, machine learning, and ensemble time-series models for monthly rainfall forecasting at a pluviometric station located in Crato, Ceará State, northeastern Brazil. A dataset comprising 328 monthly observations (1994–2022) was analyzed using Autoregressive Integrated Moving Average (ARIMA), Support Vector Regression (SVR), Multi-Layer Perceptron (MLP), and Long Short-Term Memory (LSTM) neural networks. In addition, forecast combination strategies based on arithmetic mean, median, and minimum-variance weighting were applied to assess uncertainty reduction. The models were evaluated via Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). Results indicate that deep learning models, particularly LSTM, outperformed classical approaches by better capturing non-linear dynamics and autoregressive dependencies. The minimum-variance ensemble further enhanced forecast stability by reducing error dispersion under highly variable rainfall conditions. These findings highlight the relevance of advanced forecasting and ensemble techniques for supporting water resources planning, infrastructure design, and risk management in semiarid environments.