A Hybrid Random Forest-LSTM Framework with SMOTE, Temporal Feature Engineering, and Deep Sequence Learning for Enhanced Weather Prediction

Abstract

Accurate weather prediction is essential for numerous applications, including agriculture, transportation and disaster management. This study presents a comparative analysis of RF –LSTM framework that combines the strengths of classical machine learning and deep sequence learning to improve weather classification. To address class imbalance, SMOTE was applied, and temporal feature engineering, including month, day-of-week, and lag features, was used to capture sequential patterns in weather data. The Random Forest model effectively extracts structured feature patterns, whereas the LSTM model learns temporal dependencies across sequences of days. The Experimental results demonstrate that the hybrid approach outperforms the individual models in terms of accuracy and reliability. Furthermore, strategies such as feature scaling, sequence tuning, and hybrid weighting have been shown to enhance predictive performance. This study highlights the potential of integrating machine and deep learning techniques for robust weather prediction and lays the groundwork for future improvements using advanced sequence modeling and ensemble strategies.