Intra-hour solar irradiance forecasting: An end-to-end Transformer-based network

The exponential growth of photovoltaic (PV) power generation prompts accurate PV forecasting due to the nature of PV power fluctuations. Cloud movements have a significant impact on solar irradiance, which further affects PV power. Although there are many models available to predict intra-hour irradiance using ground-based sky images, most of them are based on long short term memory (LSTM) and convolution nerual network (CNN) models. In this paper, we propose an end-to-end Transformer-based model, where Informer extracts historical time-series data features and ViViT extracts sequential ground-based sky image features. Finally, the multi-modal features are concatenated in the predictive network and then go through the fully connected layer to forecast intra-hour direct normal irradiance (DNI). Experiments demonstrate that the proposed model has a forecast skill of 35.814%, 30.407%, and 31.492% on 10-minute ahead (MA), 20 MA, and 30 MA, respectively, which outperforms the LSTM model and the CNN model. Moreover, it proves the feasibility and effectiveness of combining time series data with ground-based sky images.