ULTRA-SHORT-TERM PREDICTION METHOD OF DISTRIBUTED PHOTOVOLTAIC POWER OUTPUT BASED ON DIRECTED GRAPH CONVOLUTION RECURRENT NETWORK

Most distributed photovoltaic power forecasting methods focus on mining the temporal features of photovoltaic output，ignoring the spatial correlations between multiple adjacent PV stations’output，which leads to a large forecasting error. This paper proposes an ultra-short-term prediction method of distributed photovoltaic power method based on a directed graph convolution recurrent network，which can simultaneously extract the temporal features and spatial correlation of photovoltaic output so as to effectively reduce the forecasting error. Firstly，the temporal features and spatial correlations of photovoltaic output data are analyzed，and the temporal features are extracted by a gated recurrent unit，and the directed graph convolution network is constructed to extract the directed spatial correlations of photovoltaic output that traditional graph convolution network cannot capture. Then，the gated recurrent unit and the directed graph convolution network are fused to construct the directed graph convolution cyclic network to extract the spatio-temporal correlations of multiple photovoltaic stations’output，and the attention mechanism is used to assign weights to the spatio-temporal features at different timesteps. Finally，the prediction results are obtained through the fully connected layer. A case study is conducted with actual power data of 79 rooftop photovoltaics under different forecasting horizons. The results illustrate that compared with traditional gated recurrent unit，the MAE of the proposed method decreases by 16.3%，20.7% and 28.1% for 15-min，30-min and 60-min-ahead forecasting tasks. © 2024 Science Press. All rights reserved.