Rainfall model for short term forecasting / Amir Khomeiny Ruslan

Prediction of flash flood begins with forecasting of heavy rainfall and it is mutually dependable. Rainfall forecasting and warning system is considered as effective nonstructural measure to minimize the losses of properties and human life. Forecasting of rainfall event can be described based on characteristic of proposed rainfall forecasting model which has been devoted since past decade. This study focus on rainfall forecasting based on historical rainfall data from the local Drainage and Irrigation Department Malaysia. The approach of this study is based on event based rainfall forecasting. The application of low-order Autoregressive Moving Average (ARMA), processes to model short-term precipitation is considered following the modeling framework based on Box and Jenkins procedures. The modelling procedures involved several important steps i.e. through model identification stage, parameter estimation stage and diagnostic checking stage. The analysis of the short term ARMA model is carried out on three rainfall stations from the state of Kelantan, Malaysia. Two distinct samples sets from each rainfall station are analyzed, i.e. one set consists of ten rainfall events for model building strategies and the second set consists of five rainfall events to evaluate the performance of the mode. The model building strategies includes a procedure of determining the best ARMA model based on lowest AICc values and observation on ACF and PACF plots. The final analysis included the diagnostic checking on the model residuals. The performance of the ARMA model is evaluated based on MAPE, MAD and MSD analysis. This method analyses the model accuracy and its capability to preserve the original statistical properties. The performance of the model is evaluated on the hourly rainfall intensity (mm/hr) and cumulative rainfall intensity (mm/hr). The best ARMA model identified for Kg Aring rainfall station is ARMA (2,0,1), Gunung Gagau rainfall station, ARMA (1,0,0) and Tok Ajam rainfall station ARMA (1,0,1). The identified forecasting rainfall model should fitted enough with statistical properties of the previous historical data, minimized the root means square errors, minimized the mean absolute percentage errors, simple model with less set of parameter, the errors distribution are white noise with low ACF values and minimized the AIC value.