COMPARATIVE EVALUATION OF LSTM AND TRANSFORMER MODELS FOR SURFACE METHANE FORECASTING IN JAKARTA

Abstract

Short-horizon forecasting of near-surface methane (CH?) in tropical urban environments remains underexplored. This study compares Long Short-Term Memory (LSTM) and Transformer-based models for three-hour-ahead surface CH? forecasting in Jakarta using CAMS EAC4 reanalysis and BMKG Kemayoran observations. Historical CH?, relative humidity, 2-m air temperature, surface pressure, and wind speed were used as predictors. Both datasets underwent time synchronization, cleaning, training-only normalization, and sequence generation using a 56-step lookback and one-step forecast horizon. Data were chronologically divided into 70% training, 10% validation, and 20% testing subsets. Performance was evaluated using MAE, RMSE, and R². LSTM consistently outperformed the Transformer model. For CAMS, LSTM achieved MAE of 0.7067 ppb, RMSE of 0.9138 ppb, and R² of 0.7817, while surface observations yielded MAE of 301.8316 ppb, RMSE of 701.2368 ppb, and R² of 0.1815. The novelty lies in applying an equivalent leakage-resistant protocol across datasets with different statistical characteristics. Results establish LSTM as a robust baseline while highlighting local variability and extreme concentrations as challenges for urban methane forecasting.