Short-Term Probabilistic Forecasting for Regional PV Power Based on Convolutional Graph Neural Network and Parameter Transferring

Abstract

This paper proposes a novel end-to-end deep learning model for short-term probabilistic regional PV power forecasting. This model is of two-tier local-global structure. In the local tier, a dynamic spatial convolutional graph neural network utilizing directed-graph model is built to learn high-level representations for PV plants. In the global tier, a dynamic graph pooling method is proposed, through which local representations of PV plants are aggregated into global representations and then mapped to probabilistic regional PV power forecasts. To avoid overfitting, this paper also proposes a new training strategy based on the parameter-based transfer learning. Experimental results on the public realistic data verify that the proposed end-to-end model can provide high-quality and reliable short-term probabilistic regional PV power forecasts.