Solar Irradiance Forecasting in Tropical Weather using an Evolutionary Lean Neural Network

Clean electricity system based on solar energy is rapidly growing. However, the intermittency of solar power remains an issue. An accurate solar irradiance forecast can mitigate the impact of variable solar power generation. In this paper, we present the Evolutionary Lean Neural Network (EVLNN) for time-series solar irradiance forecasting. The key novelty of EVLNN lies in incorporating a feedback structure to a partially connected neural network while using an improved genetic algorithm to optimize its architecture. We train the model with tropical weather data to provide a days-ahead forecast of solar irradiance at four different time-steps: 1-min, 15-min, 30-min, and hourly. We investigate the effects of fewer input features in model training and assess the model accuracy using a combination of Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Adjusted Coefficient of Determination (Adjusted R-2) metrics. We then compare the results with those obtained from neural networks modeled using other evolutionary algorithms (EA), namely the Particle Swarm Optimization, the Differential Evolution, and the classic Genetic Algorithm. We include the fully connected nonlinear time-delay backpropagation neural network as a benchmark against which to evaluate the EA-based models. The results demonstrated EVLNN's good generalization capability, specifically in the presence of a sparse dataset. Moreover, our proposed model trained with a single input variable achieved improved performance for the hourly, 30-min, and 15-min time-step predictions compared with those modeled using other learning algorithms.