Perbandingan Model Deep Learning LSTM, GRU, dan Bi-LSTM untuk Prediksi Hujan Harian Australia Menggunakan Teknik SMOTE

Abstract

Climate change causing increasingly erratic rainfall patterns, triggering an increase in hydrometeorological disasters such as floods, droughts, and declining agricultural productivity. Therefore, accurate rainfall prediction is crucial for mitigation and decision-making. However, previous research often focuses solely on accuracy metrics without evaluating the model's computational burden, and often ignores the problem of class imbalance in weather datasets. This study evaluates the performance and computational efficiency of LSTM, GRU, and Bi-LSTM deep learning models for daily rainfall prediction using the historical Australian meteorological dataset weatherAus. The novelty of this study lies in the comprehensive mapping between predictive quality and resource efficiency after dataset balancing. The preprocessing stage includes handling missing values, categorical data transformation, data leakage prevention, data sharing, and the application of SMOTE oversampling. The results of the area under the curve (AUC-ROC) evaluation show that the GRU model is superior with a value of 0.85, surpassing LSTM and Bi-LSTM, respectively, at 0.84. In the rain class recall metric, GRU again leads (0.70), compared to LSTM (0.67), and Bi-LSTM (0.57). Computational evaluation, GRU is significantly more efficient with the fastest training time (1,306.26 seconds), followed by LSTM (2,259.12 seconds), and Bi-LSTM (13,348.43 seconds). Peak RAM usage relatively comparable, GRU (2,053.77 MB), LSTM (1,971.47 MB), and the highest Bi-LSTM (2,242.60 MB). These findings conclude that GRU is recommended as the most optimal model that balances accuracy and efficiency, LSTM as an alternative, while Bi-LSTM is considered less effective. Future research recommended to explore hybrid architectures or ensemble learning to capture more complex spatiotemporal patterns.