Applying Evolutionary Algorithms to Optimize Hyperparameters for Prediction Model of Solar Power Generation

Because of climate change and global warming, the demand for renewable energy grows continually. Among the renewable energy sources, solar power is the most common type due to its low construction cost and easy parallel connection with existing power grids. The power company can not only dispatch power but obtain better electricity price contracts by forecasting the power generation of photovoltaic panels. In the past, many studies have focused on the research of solar power generation, from statistical regression to mathematical planning models to heuristic meta methods and evolutionary algorithms. Recently, there are more and more literatures using machine learning to establish power generation forecasting models and even the deep learning model of artificial intelligence. However, research on hyperparameter optimization to make ensemble learning algorithms perform better is still scarce. This paper attempts to optimize the hyperparameters in the modeling process of ensemble learning with evolutionary algorithms and construct more accurate solar power prediction models. Gradient boosting regressor is employed as ensemble learning models where the hyperparameters are optimized by differential evolution, Jaya algorithm, particle swarm optimization and genetic algorithm for comparison. The data is based on practical data and weather forecasting data of solar power plants in central Taiwan. The computational results reveal that differential evolution outperforms to explore the optimal hyperparameter combination of the prediction model for solar power generation.