Perbandingan Performa Sliding Window dan Expanding Window pada Long Short-Term Memory (LSTM) dalam Peramalan Tinggi Muka Air Bendungan Katulampa

Abstract

Tren saat ini menunjukkan bahwa semakin banyak data deret waktu bervolatilitas tinggi, sehingga memerlukan metode peramalan yang adaptif. Salah satu metode analisis yang unggul dalam menangani pola kompleks, nonlinier, dan ketergantungan jangka panjang adalah Long Short-Term Memory (LSTM). Proses validasi silang pada data deret waktu berperan penting dalam mengukur performa model. Metode walk forward validation umum digunakan karena mampu memberikan performa yang baik. Data Tinggi Muka Air Bendungan Katulampa (TMA-Katulampa) termasuk jenis data dengan volatilitas tinggi. Hasil peramalan ini berperan penting dalam mendukung sistem peringatan dini potensi banjir di wilayah hilir seperti Jakarta. Penelitian ini bertujuan meramalkan TMA-Katulampa menggunakan metode LSTM serta membandingkan performa sliding window dan expanding window pada walk forward validation. Penelitian ini menggunakan data harian TMA-Katulampa periode Januari 2022 hingga April 2025. Hasil kombinasi hyperparameter terbaik learning rate, batch size, dan kernel initializer dengan sliding window yaitu 0,005, 1, dan glorot uniform, sementara expanding window yaitu 0,005, 16, dan he uniform. Akurasi nilai Mean Absolute Percentage Error (MAPE) pada sliding window dan expanding window secara berturut-turut terhadap data latih 20,47% dan 18,97%, serta data uji 21,85% dan 24,88%. Hasil peramalan 60 hari ke depan model LSTM dengan expanding window menunjukkan pola TMA-Katulampa yang lebih fluktuatif dan realistis.

Current trends indicate that there is an increasing amount of highly volatile time series data, which requires adaptive forecasting methods. One of the most effective methods for handling complex, nonlinear patterns and long-term dependencies is Long Short-Term Memory (LSTM). Cross-validation of time series data plays an important role in measuring model performance. The walk-forward validation method is commonly used because it provides good performance. The Katulampa Dam Water Level Data (TMA-Katulampa) is a type of data with high volatility. The results of this forecasting play an important role in supporting early warning systems for potential floods in downstream areas such as Jakarta. This study aims to forecast TMA-Katulampa using the LSTM method and compare the performance of sliding window and expanding window in walk forward validation. This study uses daily TMA-Katulampa data from January 2022 to April 2025. The best combination of hyperparameters for learning rate, batch size, and kernel initializer with the sliding window is 0.005, 1, and glorot uniform, while for the expanding window it is 0.005, 16, and he uniform. The accuracy of the Mean Absolute Percentage Error (MAPE) for the sliding window and expanding window respectively on the training data was 20.47% and 18.97%, and on the test data was 21.85% and 24.88%. The 60 day forecast results of the LSTM model with an expanding window show a more fluctuating and realistic TMA-Katulampa pattern.