The Use of Machine Learning and Ensemble Stacking for Classifying Rainfall Intensity in Central Jakarta in 2025

Abstract

High rainfall intensity in Central Jakarta can trigger inundation and flooding that disrupt infrastructure, mobility, and public activities. This condition highlights the need for a reliable rainfall classification model. This study develops a next-day rainfall classification approach using a Stacking Ensemble that combines Random Forest, Gradient Boosting, XGBoost, LightGBM, and CatBoost as base models, with LassoCV as the meta-model. The dataset was obtained from BMKG meteorological records for Central Jakarta in 2025. The research process follows the CRISP-DM framework, covering data understanding, data preparation, feature engineering, modeling, and evaluation. From 365 daily observations, the feature engineering stage produced 34 derived features. The dataset was then split into 80% training data and 20% testing data according to chronological order, while model validation was conducted using 5-fold Time Series Cross Validation. The evaluation results indicate that Stacking Lasso achieved the best performance, with an F1-Score of 0.774 and a recall of 0.878. During the weighting process, LassoCV retained only LightGBM as the active contributor, with a coefficient of 0.2671. SHAP interpretation showed that average relative humidity was the most influential variable in predicting rainfall. These findings suggest that the Stacking Ensemble approach has potential as a foundation for developing an urban hydrometeorological disaster mitigation support system.