Spatial-Temporal Graph Neural Network for Regional Photovoltaic Power Forecasting Based on Weather Condition Recognition
M. Zhang, P. Tao, Peiming Ren, Zongwei Zhen, F. Wang, Guang Wang · IET conference proceedings. · 2021
Photovoltaic (PV) power generation can effectively utilize solar energy resources. The regional PV power forecasting is helpful for dispatchers to formulate dispatching plans in a more scientific and reasonable way. At present, the existing methods don't sufficiently consider the temporal correlation of power output and the physical connection caused by clouds movement between PV plants. Aiming at these shortcomings, this paper proposes a spatial-temporal graph neural network method for regional photovoltaic power forecasting based on weather condition recognition. Firstly, we select the spatial-temporal graph convolutional neural network as the forecasting model, which can capture spatial feature through the Graph Convolution Neural model and capture temporal feature through the Gated Convolution Neural unit. The input is the graph structure data consists of the feature matrix and the adjacency matrix representing the connection relationship between plants, the output is the regional power. Secondly, due to the correlations between PV plants varies with weather conditions, we divide the weather into three categories: cloudless, partial cloud coverage and total cloud coverage. Different adjacency matrices will set for different weather conditions to train the forecasting model. Finally, the actual operation data is used for simulation, and the results verify the effectiveness of classification forecasting.