A Neural Network-based Time-Series Model for Predicting Global Solar Radiations

Abstract

Global solar radiation is variable in nature, hence its forecasting is very important. This paper aims to predict the global solar radiations using real experimental data collected over time by developing time series based neural networks in MATLAB. In this research work three models are developed, one for the daily prediction of global solar radiations and two for the hourly (with and without night times) prediction of the global solar radiations. The results obtained are evaluated using statistical analysis which confirmed that nonlinear autoregressive models are a good choice for predicting the global solar radiations. The regression coefficient was found high in the case of hourly prediction (including night hours) followed by hourly prediction (excluding night hours) and daily prediction (after averaging) respectively. The value of the best validation performance for hourly prediction (including night hours) was found 0.060887 with a regression coefficient of 0.96481. To further validate the applicability of the proposed algorithm, the model was compared with four time-series models developed in Waikato Environment for Knowledge Analysis. The results confirmed the high efficiency of our proposed models among others, based on the performance parameters root mean square error and mean Absolute error. The root mean square error for our proposed model came 1.28 which is the lowest as compared to other time series models. Also, the accuracy of the hourly prediction model was found higher as compared to daily prediction models. Further, the critical analysis of our method with various available methods in the literature is also discussed.