Machine learning based hydrological model: deep feed forward neural network (DFFNN)

Abstract

This study applies the Deep Feed Forward Neural Network (DFFNN) model to forecast the number of rainy days in Indonesia, a tropical country characterized by significant seasonal rainfall variations. The data used includes daily rainfall data from the Juwata Tarakan Meteorological Station over the period from 2014 to 2024. DFFNN was chosen for its ability to handle non-linear relationships and temporal dependencies within complex time series data. Based on performance evaluation metrics such as MAPE, MSE, and RMSE, the model with a 4-hidden-layer architecture and a 4-5-3-5 neuron configuration, standardized using Z-Score, showed the best performance compared to other model combinations. The forecasting results reveal a seasonal pattern consistent with tropical rainfall cycles, with peaks in rainfall occurring from January to March and a decline observed between May and August. The model proved capable of providing accurate predictions for the number of rainy days, making it suitable for supporting early warning systems in mitigating natural disaster risks such as floods and landslides. However, the study also identified several limitations, particularly in the selection of input variables, hyperparameter tuning, and network architecture choices, which influence model performance. Therefore, further research is recommended to optimize these aspects to enhance prediction accuracy and stability in weather forecasting and disaster management.