Komparasi Model LSTM dan GRU untuk Prediksi Curah Hujan dan Suhu Udara Harian di Kabupaten Garut

Abstract

Perubahan iklim dan tingginya variabilitas cuaca menyebabkan prediksi curah hujan dan suhu udara harian menjadi semakin penting, khususnya di Kabupaten Garut yang memiliki kondisi topografi beragam dan rawan bencana hidrometeorologi. Penelitian ini bertujuan membangun serta membandingkan kinerja model Long Short-Term Memory (LSTM) dan Gated Recurrent Unit (GRU) dalam memprediksi curah hujan dan suhu udara harian di Kabupaten Garut. Penelitian menggunakan data harian curah hujan dan suhu udara NASA POWER periode 2016–2025. Data dibagi menjadi 80% data latih dan 20% data uji, kemudian melalui tahap prapemrosesan berupa transformasi logaritmik, normalisasi MinMaxScaler, serta pengujian tambahan menggunakan smoothing simple moving average. Kombinasi hyperparameter terbaik diperoleh menggunakan metode random search, sedangkan performa model dievaluasi menggunakan Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), dan Mean Absolute Scaled Error (MASE). Hasil penelitian menunjukkan bahwa penerapan smoothing meningkatkan akurasi prediksi pada kedua model dengan menurunkan nilai RMSE, MAE, dan MASE. Model LSTM menghasilkan nilai kesalahan yang lebih rendah dibandingkan GRU, untuk data tanpa maupun dengan smoothing, sehingga dipilih sebagai model terbaik untuk prediksi curah hujan dan suhu udara harian di Kabupaten Garut.

Climate change and increasing weather variability have made daily rainfall and air temperature prediction increasingly important, particularly in Garut Regency, which has diverse topographic conditions and is prone to hydrometeorological disasters. This study aimed to develop and compare the performance of the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models for daily rainfall and air temperature prediction in Garut Regency. The study utilized daily rainfall and air temperature data obtained from NASA POWER for the 2016–2025 period. The dataset was divided into 80% training data and 20% testing data, followed by preprocessing, including logarithmic transformation, MinMaxScaler normalization, and additional experiments using simple moving average smoothing. The optimal hyperparameter combination was determined using the random search method, while model performance was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Scaled Error (MASE). The results showed that the application of smoothing improved the predictive performance of both models by reducing RMSE, MAE, and MASE values. The LSTM model produced lower prediction errors than the GRU model for both the original and smoothed datasets, making it the best-performing model for daily rainfall and air temperature prediction in Garut Regency.