MONTHLY RAINFALL QUANTITIES FORCASTING USING NARX NETWORK

Abstract

An accurate precipitation forecast can reflect a positive impact in several areas. It provides helpful data in hydrological projects designs, such as constructing dams, reservoirs, and rainfall networks, as well as taking some precautionary measures that can overcome the flooding problems. This paper proposes a monthly quantitative precipitation forecasting model that covers the total land area of Iraq. The model is based on the use of Nonlinear AutoRegressive with eXogenous input neural network (NARX). This type of network is considered one of the most important dynamic networks that can deal with time-series data. It is a type of recurrent network with feedback connections between its layers and tapped delay lines. The data used to train and test the network are real data obtained by NASA GES DISC which represents monthly quantitative precipitation of more than 1350 sites uniformly distributed to cover the land of Iraq for a historical period of ten years. The designed forecasting network model showed good performance, in which the total calculated MSE for the testing data set is about (2.8×10^-3), and its correlation coefficient R is about (0.95). The correlation of the predicted error with time has been checked also; it showed that almost all the autocorrelation function values fall within the bounds of the confidence interval.