Optimal Neural Network for Predicting Solar Energy in Sensor Units Based on a Cascaded Input/Structure Direct Optimization

Abstract

The sensor units are considered one of the significant technologies that use solar energy as an assistant power source to the batteries. Despite their advantages over the other forms of renewables, solar energy has an intermittent nature which negatively affects the operation of these units. Reaching an effective operation ensuring sustainable units requires a prior prediction of the harvested solar energy. Artificial neural networks (ANNs) appeared recently as a promising prediction approach with those units. This is attributed to the high accuracy compared to the conventional stochastic and statistical ones. Till now, the optimal neural network that fits with sensor units has not been precisely determined. This paper is aimed at finding the optimal neural network that would be applied with solar-supplied sensor units. This is performed by applying a cascaded input/structure direct optimization. The optimization process handles the aspects of accuracy, computational efforts, and complexity. It mainly identifies the type and number of parameters that would be utilized as inputs in the first stage. Then, it optimizes the structure by addressing the number of hidden layers and hidden neurons. The corresponding analysis has been implemented for premeasured real data over five-year time period. The results showed that the optimal neural network can be achieved by using three input parameters which are the air temperature (AT), the relative humidity (RH), and the zenith angle ( ${\theta }_z$ ). For the structure, it has been concluded that the proposed optimal ANN should have two hidden layers with ten neurons in each of them. Lastly, the proposed optimal ANN was verified against the associated prediction error which is minimized to less than 2%.