An Integrated Sparse Gated Graph Density Network Based on Transfer Learning for Multi-Site Probabilistic Forecasting of Renewable Energy

Large-scale new energy grid-connected poses significant challenges to the safe and efficient operation of smart grids. Renewable energy probabilistic forecasting (REPF) technology can analyze uncertainties in power generation, quantitatively balance risks, and prevent the breakdown of the grid. However, current REPF methods reliant on spatio-temporal maps fail to accurately estimate the probability density function (PDF) of renewable energy, resulting lacking comprehensive uncertainty analysis for distributed power generation systems (DPGS). To fill this gap, in this study, an integrated sparse gated graph density network (ISGGDN) that incorporates transfer learning to tackle the REPF challenge. A sparse gated graph dynamic convolutional network based on cross attention and residual connection is developed, which can effectively extract spatial features and spatio-temporal interactions between sites and improve the accuracy of probabilistic prediction. Furthermore, to effectively identify the types of features lost during the transfer process and to enhance the transfer learning (TL) capability, we developed an integrated approach involving multiple fine-tuning strategies based on TL. We evaluated the proposed model using wind and photovoltaic (PV) power generation data from two neighboring multi-sites, and the experimental results demonstrate that ISGGDN outperforms other existing solutions in terms of accuracy and effectiveness in REPF.