SEASONAL RAINFALL PREDICTION USING ARTIFICIAL NEURALNETWORK

Abstract

Considering the significance of rainfall in many decision making processes such as water resource management and agriculture, the present study aims at investigating the capability of non-linear techniques on seasonal rainfall prediction. One of the nonlinear techniqueswidely used is the Artificial Neural Networks (ANN) approach which has the ability of mapping between input and output patterns without a priori knowledge of the system being modeled. The complexity of the atmospheric processes that generate rainfall makes quantitative forecasting of rainfall an extremely difficult task (Hung et al, 2008). The main aim of this study is to train/develop Artificial Neural Network (ANN) model to predict seasonal rainfall (June-October) using some meteorological variables like, Sea Surface Temperature (SST), Equivalent Potential Temperature(Theta-E) and U-wind at (700, 850 and 1000) hpa levels. The study used monthly total of June-October (JJASO) rainfall dataand January-May (JFMAM) monthly data of SST, Theta-E and U-wind at (700, 850 and 1000) hpa levels for a period of 30 years (1983-2012) over Enugu, Ibadan, Kano, Maiduguri and Port Harcourt Stations. A feedforward ANN model was developed to investigate its potentials in predicting seasonal rainfall in five different stations in different ecological zones (rainforest, sudan savanna, sahel savanna and swampy forest zones).The models were trained based on Modified Levenberg-Marquardt algorithm (MLMA) with tansigmoid activation function for the hidden layer and purelin activation function for the output layer. The models were developed using JJASO rainfall data and JFMAM ThetaE, SST and U-wind data of 1983-2006 (24years). The models were tested using JJASO rainfall data of 2007-2012 (6years). The error metrics that was used for model's performance evaluation was Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The index of agreement (d) was chosen as a criterion to compare models. The simulation results from the five stations considered in this study showed that ANN model can perform well in simulating seasonal rainfall over the study areas.Also, the results showed that the selected meteorological variables (predictors) used in training the model are significant in explaining the JJASO seasonal rainfall variations in Nigeria.Thus, ANN model can be adopted in predicting seasonal rainfall in Nigeria.