Weather Forecasting Using Stacked-LSTM

Abstract

This study proposes a Stacked Long Short-Term Memory (Stacked LSTM) model for multivariate weather forecasting using historical meteorological data from Denpasar City. The dataset consists of 264,924 records collected between 1990 and 2020, including four key weather variables: temperature, humidity, pressure, and wind speed. The model is designed to capture temporal dependencies in time-series weather data through multiple LSTM layers. A sliding window technique is used to construct input sequences, and the model is trained for 50 epochs with a batch size of 64, incorporating dropout regularization to improve generalization. The dataset is divided using a train–test split, where 20% of the data is reserved for performance evaluation. Experimental results demonstrate that the proposed model achieves strong predictive performance across all weather variables. The evaluation on the test dataset yields an average Mean Absolute Error (MAE) of 1.08, Mean Absolute Percentage Error (MAPE) of 10.22%, Root Mean Squared Error (RMSE) of 1.93, and a Coefficient of Determination (R²) of 0.86. Among the predicted variables, humidity and temperature show the highest accuracy with R² values of 0.9537 and 0.9031, respectively. The findings indicate that the Stacked LSTM architecture successfully captures both short-term and long-term temporal relationships within multivariate weather datasets. The proposed approach demonstrates strong potential for improving automated weather forecasting systems, particularly in tropical urban environments characterized by complex climatic dynamics. Future work may focus on integrating real-time weather data sources and adaptive retraining mechanisms to further enhance prediction accuracy and operational applicability.