Self-supervised learning based multi-modal intra-hour irradiance forecast

Multi-modal short-term irradiance forecast models usually fuse multi-source heterogeneous variables to predict future irrdiance and ground-based cloud images is one of the important modal. The dynamic information of clouds is shown to improve the models performance. However, such methods usually use original images directly for prediction, which decreases the efficiency and practicality of the training process. Therefore, a self-supervised learning method is proposed to learn the features of ground-based cloud images and meteorological factors. The obtained representations of exogenous variables are then used as the input for multi-modal short-term irradiance forecast. This method is validated on an open access dataset, and the results demonstrate that the self-supervised based approach outperforms other classical forecast models. Meanwhile, the images are compressed into vectors, which not only saves storage space but also reduces the training time of the prediction model significantly. In addition, the model achieves good transferability for ground-based cloud images with different qualities.