Assessment of neural networks performance in modeling rainfall amounts

Abstract

This paper presents the evaluation of performance of Neural Network (NN) model in predicting the behavioral pattern of rainfall depths of some locations in the North Central zones of Nigeria. The input to the model is the consecutive rainfall depths data obtained from the Nigerian Meteorological (NiMET) Agency. The neural networks were trained using neural network toolbox in MATLAB with fifty years (1964–2014) total monthly historical data of five locations while two other locations, Abuja and Lafia with twenty-nine years (1986-2014) and eleven years (2004-2014) total monthly data respectively. Analysis showed the variation in the values of correlation coefficients (R) for each location of the study area in response to change in number of hidden neurons. The average R values of 0.80, 0.62, 0.65, 0.67, 0.79, 0.76 and 0.81 with corresponding mean square errors of 2.12, 0.23, 0.26, 0.36, 2.61, 1.18 and 1.03 were obtained for Abuja, Makurdi, Ilorin, Lokoja, Lafia, Minna and Jos respectively. The results showed some slight variability in the performances of NN due to changes in the number of hidden neurons during the network training. These values of R indicated that the networks are fit to be used for the subsequent quantitative prediction of rainfall depths in each location which is useful for safeguarding against future flood and drought occurrence in the North Central zone, Nigeria.

Keywords: Rainfall depths, NN, coefficient of correlation, mean square errors