Sky image-based solar forecasting using deep learning with heterogeneous multi-location data: Dataset fusion versus transfer learning

Solar forecasting from ground-based sky images has shown great promise in reducing the uncertainty in solar power generation. With more and more sky image datasets available in recent years, the development of accurate and reliable deep learning-based solar forecasting methods using more diverse multi-location data has seen a huge growth in potential. From that perspective, the joint utilization of these heterogeneous data – such as images captured with different camera setups, sensor measurements (i.e., irradiance versus photovoltaic power output) varying in scale and distribution – is both a unique opportunity and a critical challenge. This study explores ways to cope with such data heterogeneity and compares three different strategies for training solar forecasting models based on sky image datasets collected across three continents by three research groups. Specifically, for a location of interest, we compare the performance of (1) local models trained individually on a single local dataset (i.e., the standard methodology in the literature); (2) global models jointly trained on the fusion of multiple heterogeneous datasets; and (3) locally fine-tuned models trained via transfer learning from a pre-trained model. The results suggest that, with the current modeling strategy, local models work well when deployed locally, but significant errors are observed when applied offsite. The global model, with proper normalization of the prediction targets, can adapt well to individual locations at the cost of a potential increase in training efforts. Pre-training models on a large and diversified source dataset and transferring to a target location generally achieves superior performance over the other two strategies. With 80% less local training data, a fine-tuned model performs similarly to the baseline trained on the entire local dataset. Overall, algorithms built on heterogeneous multi-location sky image datasets have the potential to be more accurate, more robust, and adapt faster to new locations than local models based on a single location.