Multi-grained adaptive informer for multi-step solar irradiance forecasting

Solar energy is a crucial element of contemporary renewable energy systems . The accurate forecasting of solar irradiance is crucial for grid stabilization and carbon emission management. We develop a Multi-Granularity Adaptive Informer (MGAI) model for the multi-scale forecasting of solar irradiance. To resolve the issue of attribute entanglement in solar irradiance data, we offer a multi-granularity disentangler that enables the model to discern connections within the data at various granularities, hence improving data representation. And we employ relative position embedding to prevent the loss of positional information. Furthermore, we employ the advanced Informer as the foundational model to enhance computational efficiency while maintaining superior modeling capacity. We enhance ProbSparse self-attention by incorporating adaptive features and optimizing it using the Spider Bee Optimization (SWO) algorithm. We conducted comparison experiments using datasets from two distinct climate zones, across four temporal resolutions and four forecasting scales. The suggested model has superior performance relative to leading solar irradiance forecast and time series forecasting models, with mean absolute error (MAE) and root mean square error (RMSE) decreased by 9.79% and 11.12%, respectively. • A multi-granularity adaptive Informer was applied to solar irradiance forecasting. • A multi-scale convolution and a positional embedding method were designed. • The original ProbSparse self-attention was improved by an adaptive mechanism. • Multiple dimensional comparative experiments were conducted on two datasets.