Intelligent optimization models based on hard-ridge penalty and RBF for forecasting global solar radiation

• CS-hard-ridge-RBF and DE-hard-ridge-RBF are proposed to forecast solar radiation.
• Pearson and Apriori algorithm are used to analyze correlations between the data.
• Hard-ridge penalty is added to reduce the number of nodes in the hidden layer.
• CS algorithm and DE algorithm are used to determine the optimal parameters.
• Proposed two models have higher forecasting accuracy than RBF and hard-ridge-RBF.

Due to the scarcity of equipment and the high costs of maintenance, far fewer observations of solar radiation are made than observations of temperature, precipitation and other weather factors. Therefore, it is increasingly important to study several relevant meteorological factors to accurately forecast solar radiation. For this research, monthly average global solar radiation and 12 meteorological parameters from 1998 to 2010 at four sites in the United States were collected. Pearson correlation coefficients and Apriori association rules were successfully used to analyze correlations between the data, which provided a basis for these relative parameters as input variables. Two effective and innovative methods were developed to forecast monthly average global solar radiation by converting a RBF neural network into a multiple linear regression problem, adding a hard-ridge penalty to reduce the number of nodes in the hidden layer, and applying intelligent optimization algorithms, such as the cuckoo search algorithm (CS) and differential evolution (DE), to determine the optimal center and scale parameters. The experimental results show that the proposed models produce much more accurate forecasts than other models.