Leveraging Sequential and Attention-Based Deep Learning Architectures for Accurate Daily Rainfall Prediction in Jakarta, Indonesia, Using Atmospheric Predictors

In this study, we developed and evaluated daily rainfall prediction models utilizing deep learning architectures, specifically comparing Long Short-Term Memory (LSTM) and Transformer models integrated with various atmospheric predictors. Our results indicate that the LSTM achieved superior accuracy at short-term lags—reaching an R² of 0.94 and an RMSE as low as 4.81 at lag-3—whereas the Transformer demonstrated higher consistency across all lags, maintaining stable R² values between 0.87 and 0.88. The application of a 5-day smoothing pre-processing step significantly enhanced prediction quality for both architectures by mitigating high-frequency noise, a benefit particularly pronounced in the LSTM due to its sensitivity to data fluctuations. Notably, the inclusion of tropical wave variables did not substantially improve model performance and, in some instances, reduced LSTM accuracy at longer lags by increasing input complexity; conversely, the Transformer remained robust to these additional variables. Among the predictors evaluated, Vertically Integrated Moisture Flux Divergence (VIMD) emerged as the most critical feature, underscoring its physical relevance to precipitation processes in convective and monsoonal regions. These findings suggest that while LSTMs excel at capturing immediate temporal dynamics, Transformers provide a more stable framework for longer-range rainfall forecasting in Jakarta.