Deep learning for single-site solar irradiance forecasting using multi-station data

This study examines the integration of data from multiple stations for solar irradiance forecasting at a single site using advanced deep learning models, such as long-term memory (LSTM), deep modular attention (DeepMap), and graph convolutional networks (GC-LSTM). The research addresses an important gap: the statistical evaluation of the contribution of neighboring data to improving forecast accuracy in solar PV applications. Using a large dataset from 12 Colombian locations, representing diverse climatic conditions, we rigorously evaluate the ability of these models to take advantage of spatio-temporal information. The results reveal slight improvements in short-term forecasting, but these improvements are statistically insignificant, as validated by chi-square tests. The results highlight the need for more advanced methods to effectively exploit spatial data, which will guide the future development of solar irradiance prediction models.