Optimization of Mineral Fuel Export Forecasting Using Attention-based Long Short-Term Memory

Abstrak

Purpose: This study aims to optimize the forecasting of the Net Value of Indonesia's mineral fuel exports using the Attention-based Long Short-Term Memory (LSTM) model, supported by Dropout and Recurrent Dropout techniques that are combined to produce an optimal model. Methods: Modeling uses an LSTM architecture equipped with an Attention mechanism, as well as Dropout and Recurrent Dropout. The research procedure uses the CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology. The research material used is the Indonesian mineral fuel export dataset with HS code 27 from 2014 to 2025. Model was built using the Random Search method to optimize hyperparameters such as the number of neurons (units), activation functions (Tanh, ReLu), and optimizers (Adam, Nadam, RMSprop). Result: The Attention-based LSTM model with Dropout and Recurrent Dropout techniques achieved a MAPE of 7.76%, which was better than the other models tested. Attention analysis shows that lag 12 has the greatest dominance, while lags 11 to 10 also contribute significantly, indicating an annual seasonal pattern. Projections for the next 12 months show a moderate decline in Net Value, in line with seasonal trends and historical data. Novelty: The main contribution of this research is the optimization of an Attention-based LSTM model using a combination of Dropout and Recurrent Dropout techniques, which is effective in forecasting Indonesia's mineral fuel export values because it is able to capture annual seasonal patterns, thereby improving the accuracy and stability of the forecast results.

Kata kunci

• Deep Learning
• LSTM
• Attention mechanism
• Dropout
• Recurrent Dropout