Deep learning model architecture performance comparison in photovoltaic systems with irradiance and temperature data

Microgrids are one of the most reliable energy types of production in rural and isolated places. The studies on predicting the power produced by campuses, hospitals, and rural microgrids have recently shown a growing interest in electrical power decision-makers and researchers. The accuracy of predicted power is utmost when modeling a reliable power delivery. In this study, model performance and configuration are investigated for predictive power modeling with the application of deep learning techniques. Photovoltaic systems utilizing irradiance and temperature data are employed to investigate the analysis of the various aspects of model configuration such as learning rate adjustment, activation function variation, batch size variation, alteration of the number of hidden units, and modification of the optimization algorithm across combinations of two, three, and four of the following deep learning methods: Convolutional Neural Network, Recurrent Neural Network, Feedforward Neural Network, Long Short-Term Memory Networks, and Gated Recurrent Unit. Evaluation metrics such as Mean Squared Error, Mean Absolute Error, Root Mean Squared Error, Mean Absolute Percentage Error, and R-squared are used to verify the performance of different model combinations. The results demonstrated different impacts on model configurations with grid search tuning, with certain combinations showing better accuracy than others, or ones showing potential issues with stability and convergence.