Short-term solar radiation prediction in Huddersfield based on the feedforward neural network

Solar energy as the best renewable energy source has the potential in reducing carbon emissions and climate warming in which solar radiation (SR) prediction is of great significance for making power dispatch plans. A large number of works have been contributed to the SR forecasting at different regions based on meteorological data and artificial intelligence models. However, reliable SR prediction greatly depends on the selection of training features and the short-time forecasting is more conducive to decision-making in power distribution. In this paper, the feedforward neural network (FNN) is used to achieve the hourly SR prediction in Huddersfield, UK based on the historical observation data from MET office. The input features are specifically selected by considering the contribution of each meteorological feature. Besides, the month is included as the inputs to minimize the influence from the seasonal feature, thus ensuring prediction accuracy. Therefore, four important input variables including hour, humidity, temperature and month are finally selected for FNN model training. The prediction accuracy can be more than 85% in terms of R² and its MAPE is lower than 0.8%. The results reveal that FNN provides a more accurate SR prediction of Huddersfield than the least square support vector regression (LSSVR) method.