Penerapan Metode SARIMA untuk Peramalan Curah Hujan Dasarian di Stasiun Klimatologi Banten

Abstract

Rainfall plays an important role in various aspects of life, particularly in water resource management and agriculture. This study aims to identify the temporal characteristics and seasonal patterns of dasarian (10-day) rainfall, to formulate and select an appropriate SARIMA model, and to evaluate its forecasting performance on out-of-sample data at the Banten Climatology Station. The data analyzed cover the period from Dasarian I of January 1976 to Dasarian I of January 2025, totaling 1,765 observations, divided into training and testing data. The analysis stages include data exploration, stationarity testing, model identification via ACF and PACF, estimation of several candidate models, model selection based on AIC and BIC, residual diagnostics, and forecast evaluation using RMSE and MAE. The results show that rainfall has an annual seasonal pattern with a period of 36 dasarians and is stationary in mean without either nonseasonal or seasonal differencing. The SARIMA(1,0,1)(1,0,1)36 model was selected because it produced the lowest AIC and BIC and had significant parameters. Residual diagnostics showed no autocorrelation. The residuals exhibited non-uniform variance (heteroskedasticity) across year groups; however, the ARCH effect test was not significant, so there was no evidence of conditional heteroskedasticity influenced by past squared residuals. The residuals were not normally distributed because of extreme values. Evaluation on the testing data produced an RMSE of 59.61 mm and an MAE of 44.94 mm. The model can represent the seasonal pattern but produced smoother predictions, so it did not fully capture extreme rainfall events. The selected model is better suited for forecasting the average seasonal pattern of dasarian rainfall than for capturing the magnitude of extreme events.