<![CDATA[Rainfall forecasting plays a crucial role in hydrology, agriculture, water resource management, and disaster mitigation. However, rainfall data typically exhibit fluctuating, seasonal, and nonlinear characteristics, which make the forecasting process quite complex. In this study, we propose a hybrid multi-model stacking ensemble to improve rainfall prediction accuracy in Kediri Regency. Our framework integrates statistical models—namely the Seasonal Autoregressive Integrated Moving Average (SARIMA) and Holt-Winters models—with machine learning and deep learning models, specifically Random Forest and Long Short-Term Memory (LSTM). We use Linear Regression as a meta-learner to combine predictions from all base models. The dataset contains monthly rainfall records from 2009 to 2022. Various preprocessing techniques are applied to the dataset, primarily normalization, lag feature construction, stationarity testing, and time-series data transformation, to enable deep learning. We use the Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) to evaluate each model's predictions. In the experiments, the ensemble stacking model outperformed the other models, with an MAE of 28.56, MSE of 1053.26, RMSE of 32.45, and MAPE of 13.05%. The results of the models used in the experiments, including the standalone SARIMA and Holt-Winters models, Random Forest, and LSTM, also showed inferior performance. Our model forecasted rainfall over the next 12 months while preserving historical seasons and data fluctuations, supporting the claim that the hybrid stacking ensemble method optimizes the accuracy, stability, and robustness of rainfall prediction for complex time-series data. Contribution to Sustainable Development Goals (SDGs):SDG 2 – Zero HungerSDG 6 – Clean Water and SanitationSDG 13: Climate Action]]>
