基于有向图卷积循环网络的分布式光伏出力超短期预测方法

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

提出一种基于有向图卷积循环网络的分布式光伏超短期功率预测方法,该方法可同时提取光伏出力的时序特征和空间相关性,有效减小预测误差。首先,分析光伏出力数据兼具时序性和空间相关性,利用门控循环网络提取时序特征,构建有向图卷积网络提取传统图卷积无法捕捉的光伏出力有向空间相关性;然后,融合门控循环单元和有向图卷积网络,构建有向图卷积循环网络以提取多光伏站点出力的时空相关性,并利用注意力机制为不同时刻的时空特征分配权重;最后,通过全连接层输出最终的预测结果。采用某地区屋顶光伏实际出力数据在不同预测时间尺度下比较所提方法与其他方法的预测性能,结果表明,预测时间尺度为15、30和60 min时,相对于传统循环网络,所提方法的MAE分别减少16.3%、20.7%和28.1%。

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.