Comparison of Machine Learning and Deep Learning Algorithms for Daily Weather Forecasting

Abstract

Global climate change has increased the complexity of weather patterns, particularly in tropical regions such as Merauke Regency. This study evaluates five machine learning models, including RF, SVM, Prophet, LSTM, and GRU, for daily weather forecasting. We aim to assess the effectiveness of deep learning methods in capturing temporal dependencies in tropical weather systems and compare their performance against conventional machine learning approaches. We apply multiple evaluation metrics, including Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the coefficient of determination (R²), to ensure robust model assessment. The results show that deep learning models consistently outperform traditional methods. GRU achieves the best performance with RMSE = 1.37, MSE = 1.88, and R² = 0.87, followed closely by LSTM with RMSE = 1.39 and R² = 0.86. In contrast, RF, SVM, and Prophet exhibit higher error rates and lower predictive accuracy. Correlation analysis reveals strong relationships between key meteorological variables, particularly rainfall and humidity, indicating that multi-variable inputs improve forecasting performance. Overall, the findings confirm that GRU is the most effective model for this dataset, while LSTM serves as a strong alternative. This study highlights the superiority of deep learning approaches for modeling complex, nonlinear, and time-dependent weather patterns in tropical regions.