Machine Learning-Based Multivariate Weather Prediction for Agricultural Decision Support in Aceh Besar

Abstract

Accurate localized weather prediction plays a crucial role in supporting agricultural decision-making, particularly in climate-sensitive regions. This study proposes a machine learning-based multivariate weather prediction framework for agricultural decision support in Aceh Besar Regency, Indonesia, using a 2024 meteorological dataset. The dataset comprises five key variables: air temperature, relative humidity, solar radiation, cloud cover, and precipitation, collected from regional meteorological monitoring stations and validated secondary sources. A systematic preprocessing pipeline was applied, including missing value imputation, outlier detection, and Min–Max normalisation to ensure data quality and model stability. Three predictive models were implemented and evaluated, namely Random Forest, Transformer, and a hybrid CNN–LSTM architecture. Experimental results indicate that the CNN–LSTM model achieved the best performance for temperature prediction, with RMSE of 0.82 °C, MAE of 0.65 °C, and R2 of 0.92. Correlation analysis reveals a strong negative relationship between temperature and relative humidity (r = −0.687), along with moderate associations among other climatic variables. Residual diagnostics demonstrate near-normal error distribution under typical conditions, although deviations are observed during extreme weather events. The predicted weather outputs are further interpreted to support agricultural decision-making, including the identification of potential planting windows, irrigation planning, and early indications of drought risk. These findings demonstrate that machine learning models can effectively capture localised weather dynamics and provide a reliable decision-support basis for climate-adaptive agricultural management in Aceh Besar.