Machine learning strategies for multiannual rainfall prediction and drought early warning: insights from Ceará, Brazil

Accurate rainfall prediction is essential for drought mitigation and water resource management, particularly in regions prone to climatic variability. This study evaluates various machine learning models for predicting monthly rainfall in Ceará, Brazil, utilizing chaos theory and phase-space reconstruction to capture rainfall’s chaotic dynamics. By mapping time series data to a higher-dimensional space, models better align with underlying chaotic patterns. Data from 20 hydrological stations (1962 to 2006) were used to assess models such as Decision Tree, Random Forest (RF), Support Vector Machine, Long Short-Term Memory Artificial Neural Network, and a Stacked ensemble model. Performance metrics included Mean Absolute Error, Root Mean Square Error, Nash–Sutcliffe Efficiency (NSE), and Pearson Correlation. RF and the Stacked model achieved the highest predictive accuracy, with average NSE values of 0.91 and 0.93. Optimal embedding dimensions varied across stations, generally ranging from 5 to 15, highlighting the need for dimension-specific tuning. The study’s machine learning-based rainfall predictions hold practical significance for real-world applications. By providing early forecasts, these models can serve as early warning tools, enabling proactive drought mitigation efforts and improving water resource allocation. Accurate, location-specific predictions can support water managers and policymakers in planning interventions.