Application of nonlinear autoregressive neural network to estimation of global solar radiation over Nigeria

Abstract

In this paper, surface data meteorological were used as input variables to create, train and validate the network in which global solar radiation serves as a target. These surface data were obtained from the archives of the European centre for Medium-Range weather forecast for a span of 36 years (1980-2015) over Nigeria. The research aims to evaluate the predictive ability of the nonlinear autoregressive neural network with exogenous input (NARX) model compared with the multivariate linear regression (MLR) model using the statistical metrics. Model selection analysis using the index of agreement (dr) metric showed that the MLR and NARX models have values of 0.710 and 0.853 in the Sahel, 0.748 and 0.849 in the Guinea Savannah, 0.664 and 0.791 in the Derived Savannah, 0.634 and 0.824 in the Coastal regions, and 0.771 and 0.806 in entire Nigeria respectively. Meanwhile, error analyses of the models using root mean square errors (RMSE) showed the values of 1.720 W/m2 and 1.417 in the Sahel region, 2.329 W/m2 and 1.985 W/m2 in the Guinea Savannah region, 2.459 W/m2 and 2.272 W/m2 in the Derived Savannah region, 2.397 W/m2 and 2.261 W/m2 in the Coastal region and 1.691 W/m2 and 1.600 W/m2 in entire Nigeria for MLR and NARX models respectively. These showed that the NARX model has higher dr values and lower RMSE values over all the climatic regions and entire Nigeria than the MLR model. Finally, it can be inferred from these metrics that the NARX model gives a better prediction of global solar radiation than the traditional common MLR models in all the zones in Nigeria.