基于时空特征聚类和双层动态图卷积网络建模的短期净负荷预测

Short-term Net Load Forecasting Based on Temporal-spatial Feature Clustering and Two-layer Dynamic Graph Convolutional Network Modeling

净负荷是实际负荷与光伏出力之差,针对净负荷中实际负荷强波动性与光伏出力强随机性相互耦合、表后光伏出力不可见等特点导致准确预测困难的问题,提出了一种基于时空特征聚类和双层动态图卷积网络建模的短期净负荷预测方法。首先,通过提取用户净负荷的日内时间特征、长期趋势特征和空间关联特征建立净负荷子集群聚类模型;其次,以子集群为图节点构建考虑“负荷-光伏”双维相关性的图结构,使其能够同时反映负荷和光伏出力特性;最后,引入净负荷总节点和动态邻接矩阵,构建通过长短期记忆神经网络连接的双层动态图卷积模型,得到净负荷预测结果。基于悉尼Ausgrid实际净负荷数据设计的消融实验结果表明,所提时空特征聚类方法和双层动态图结构分别使净负荷预测结果的均方根误差降低了13.44 kW和7.55 kW。未来将进一步拓展预测尺度,为电网保供决策提供更多信息支撑。

The net load is the difference between the actual load and the photovoltaic(PV)output.To address the chal-lenge of accurate forecasting which stems from the strong coupling between the fluctuating of actual load and the stochasticity of PV output,as well as the invisibility of PV output behind the meter,we put forward a short-term net load forecasting method based on temporal-spatial feature clustering and two-layer dynamic graph convolutional networks(GCN)modeling.Firstly,a net load subset clustering model is established by extracting daily temporal features,long-term trend features and spatial correlation features of user net loads.Secondly,a graph structure considering the“load-PV”bi-variate correlation is built with sub-clusters as graph nodes,to simultaneously reflect load and PV output characteristics.Finally,the total node of net load and the dynamic adjacency matrix are introduced,and a two-layer dynamic graph con-volution model connected by long short-term memory neural networks is constructed to obtain the net load forecasting results.The ablation experiments results based on the actual net load data from Ausgrid in Sydney indicate that the pro-posed spatio-temporal feature clustering method and the two-layer dynamic graph structure can reduce the root mean square error of net load forecasting results by 13.44 kW and 7.55 kW,respectively.Future work will further expand the forecasting scale and provide more information supports for power grid power supply decision-making.