Parallelized hybrid ensemble machine learning framework for scalable and accurate rainfall prediction

Accurate rainfall forecasting is essential for agriculture, hydrological planning, and disaster preparedness, while balancing prediction quality with computational constraints. This study establishes a stacking-based ensemble framework for binary rainfall prediction using the Australian RainTomorrow dataset (over 140,000 daily observations), showing learned meta-feature combination outperforms probability averaging for imbalanced weather prediction. This study compares voting architectures against two-layer stacking ensembles where an XGBoost meta-learner optimizes weights from leakage-safe out-of-fold probabilities. Two ensemble families are proposed: RXL (Random Forest + XGBoost + LightGBM) and TriBoost (CatBoost + XGBoost + HistGradientBoosting), evaluated as voting and stacking configurations with tuned single learners. The methodology ensures rigor through unified preprocessing, leakage control, and evaluation under three protocols: stratified holdout (Protocol A), 5-fold cross-validation (Protocol B), and temporal split (Protocol C). Class imbalance is addressed via class weighting, fold-internal SMOTE, and threshold tuning, while temporal features mitigate distribution drift. Results confirm advantages for stacking architectures. TriBoost Stack achieves state-of-the-art performance under Protocol A (Accuracy 0.864, Macro F1 0.776, PR AUC 0.751, ROC AUC 0.893) with strong minority-event detection (Recall Yes 0.565, Precision Yes 0.752), exceeding voting ensembles and benchmarks. RXL Stack follows (Accuracy 0.861, Macro F1 0.772), while voting variants underperform on minority-sensitive metrics. Cross-validation confirms stability (TriBoost Stack: Macro F1 0.772 ± 0.004), and temporal evaluation validates deployment readiness (Macro F1 0.754). Analysis shows practical viability: parallel CPU execution and GPU acceleration deliver training speedups without sacrificing quality. The framework provides a validated solution for operational weather forecasting systems, balancing predictive performance and computational constraints.