Decade Rainfall Prediction Using Prophet Algorithm and LSTM (Case Study in Banjarnegara Regency)

Abstract

Hydrometeorological disasters such as floods and landslides in Banjarnegara Regency are closely related to fluctuating rainfall variability. This study aims to predict decadal (10-day) rainfall by comparing the performance of the Prophet algorithm and the Long Short-Term Memory (LSTM) model. The dataset comprises daily rainfall records from 14 observation stations spanning the period 2005–2024. The research stages included preprocessing, modelling, hyperparameter optimization using Optuna, and evaluation with Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). The results indicate that the Prophet model outperformed LSTM in most locations, with an average RMSE of 69.55 and MAE of 53.05, lower than LSTM, which recorded 73.03 and 55.72, respectively. The ensemble averaging model produced competitive results at several locations, although it was less responsive to sharp fluctuations in rainfall. These findings confirm that Prophet is more effective in capturing seasonal patterns and long-term trends, thus providing significant potential to support climate-based disaster mitigation systems in vulnerable areas such as Banjarnegara