Monthly Maximum Accumulated Precipitation Forecasting Using Local Precipitation Data and Global Climate Modes

Abstract

Successive days of precipitation are known to cause flooding in monsoon-susceptible countries. Forecasting of daily precipitation facilitates the prediction of the occurrences of rainfall and number of wet days. Using the maximum five-day accumulated precipitation (MX5d), we can predict the magnitude of precipitation in a specific period as it may indicate the extreme precipitation. In this study, a method to forecast monthly extreme precipitation using artificial neural networks (ANNs) is assessed using past MX5d data and global climate indices such as Southern Oscillation Index (SOI), Madden Julian Oscillation (MJO), and Dipole Mode Index (DMI) in Kuantan and Kota Bharu, Malaysia. The use of combined inputs (MX5d with SOI, MJO, and DMI) is proposed to investigate the concurrent effect of lagged values of local precipitation data and global climate indices on seasonal extreme precipitation. Four cases of data are sampled representing two major seasonal variations in Malaysia. The analysis of extreme precipitation trends is important for the prediction of high precipitation events. ANNs are widely applied in the hydrology field because of their nonlinear ability in predicting nonstationary and seasonal data. In this paper, we have compared ANNs with seasonal autoregressive integrated moving average (ARIMA) and regression analysis using out-of-sample test data. The results for Kuantan indicate that seasonal ARIMA is the best method to forecast extreme precipitation when MX5d lags are used as input. For Kota Bharu, ANN exhibits better generalization ability than ARIMA and regression analysis when dual inputs (lagged MX5d and lagged global climate indices) are utilized in the model.