Deep multi-attribute spatial–temporal graph convolutional recurrent neural network-based multivariable spatial–temporal information fusion for short-term probabilistic forecast of multi-site photovoltaic power

Accurate solar photovoltaic (PV) power forecast is crucial to accommodating large-scale PV power into the electricity grid and realizing carbon neutrality. Current researches mainly focus on using temporal information in the historical PV power data for forecast. The spatial information in adjacent PV stations is often neglected. Meanwhile, multivariable historical information and future clear-sky physical prior knowledge are also often neglected in PV forecast. Aiming to solve above problems, this paper proposes a novel PV forecast method from the perspective of multivariable spatial–temporal information fusion for the first time. Physical analysis of PV power is used to for verifying the feasibility of multivariable spatial–temporal forecast. Multi-attribute spatial information including PV power, Global Horizontal Irradiation (GHI), temperature, historical clear-sky GHI and future clear-sky physical prior knowledge from the spatially adjacent PV stations are first used in PV forecast. Transfer entropy are innovatively introduced to verify the information gain of adding multi-attribute spatial information from the viewpoint of information theory for the first time. A novel method called Deep Multi-Attribute Spatial-Temporal Graph Convolutional Recurrent Neural Network (DMA-STGCRNN) is proposed to realize multi-site PV power forecast, where multi-attribute spatial graph is first proposed to extract the multi-attribute spatial information in PV forecast. Kernel Density Estimation (KDE) is used to give probabilistic confidence interval of PV power. Experiments in two-year (2021–2022) PV power generation data from 11 provinces of Belgium verify that the proposed DMA-STGCRNN method significantly outperforms conventional spatial–temporal single-variable forecast methods and 10 temporal forecast methods.