Deep learning-based hybrid short-term solar forecast using sky images and meteorological data

The global growth of solar power generation is rapid, yet the complex nature of cloud movement introduces significant uncertainty to short-term solar irradiance, posing challenges for intelligent power systems. Accurate short-term solar irradiance and photovoltaic power generation predictions under cloudy skies are critical for sub-hourly electricity markets. Ground-based image (GSI) analysis using convolutional neural network (CNN) algorithms has emerged as a promising method due to advancements in machine vision models based on deep learning networks. In this work, a novel deep network, "ViT-E," based on an attention mechanism Transformer architecture for short-term solar irradiance forecasting has been proposed. This innovative model enables cross-modality data parsing by establishing mapping relationships within GSI and between GSI, meteorological data, historical irradiation, clear sky irradiation, and solar angles. The feasibility of the ViT-E network was assessed the Folsom dataset from California, USA. Quantitative analysis showed that the ViT-E network achieved RMSE values of 81.45 W/m2 , 98.68 W/m2 , and 104.91 W/m2 for 2, 6, and 10-minute forecasts, respectively, outperforming the persistence model by 4.87%, 16.06%, and 19.09% and displaying performance comparable to CNN-based models. Qualitative analysis revealed that the ViT-E network successfully predicted 20.21%, 33.26%, and 36.87% of solar slope events at 2, 6, and 10 minutes in advance, respectively, significantly surpassing the persistence model and currently prevalent CNN-based model by 9.43%, 3.91%, and -0.55% for 2, 6, and 10-minute forecasts, respectively. Transfer learning experiments were conducted to test the ViT-E model's generalisation under different climatic conditions and its performance on smaller datasets. We discovered that the weights learned from the three-year Folsom dataset in the United States could be transferred to a half-year local dataset in Nottingham, UK. Training with a dataset one-fifth the size of the original dataset achieved baseline accuracy standards and reduced training time by 80.2%. Additionally, using a dataset equivalent to only 4.5% of the original size yielded a model with less than 2% accuracy below the baseline. These findings validated the generalisation and robustness of the model's trained weights. Finally, the ViT-E model architecture and hyperparameters were optimised and searched. Our investigation revealed that directly applying migrated deep vision models leads to redundancy in solar forecasting. We identified the best hyperparameters for ViT-E through manual hyperparameter space exploration. As a result, the model's computational efficiency improved by 60%, and prediction performance increased by 2.7%.