【摘要】 针对有序充电中电力负荷数据的复杂时间特征，提出一种基于时序图神经网络预测模型。首先，采用聚类算法对历史负荷数据进行自适应时段划分，自动识别出具有独特用电模式的时间段，在此基础上，不仅依据时间顺序构建边连接，还通过K近邻（Knearest neighbor,KNN）算法挖掘特征相似的时段间关联，构建了融合时序性与特征相似性的复合时序关系图，能精确表征不同时段间的复杂关联。其次，设计了一个融合长短期记忆（long short-term memory,LSTM）网络与图注意力网络（graph attention network,GAT）的双分支预测模型，分别用于学习负荷的纵向时间依赖和横向时段关联并通过融合层将两类特征进行集成以完成预测。最后，在真实电力负荷数据集上的实验表明，所提模型在均方根误差（root mean square error,RMSE）和平均绝对误差（mean absolute error,MAE）等指标均优于对比基线模型，验证了其有效性。更多还原

【Abstract】 To address the complex temporal characteristics of power load data in coordinated charging, a short-term load forecasting model based on a temporal graph neural network is proposed. First, a clustering algorithm is used to adaptively divide the historical load data into time periods, each representing distinct electricity consumption patterns. On this basis, we construct edge connections according to temporal order and use theK-nearest neighbors( KNN) algorithm to capture correlations between time periods with similar features, thereby forming a composite temporal-similarity graph that accurately captures complex inter-period correlations. Second, we design a dual-branch forecasting model that integrates long short-term memory(LSTM) network and the graph attention network( GAT) to learn temporal dependencies and inter-period correlations, respectively. Then, we fuse the two feature types for final forecasting. Experiments on real-world power load datasets demonstrate that the proposed model outperforms baseline models in terms of root mean square error(RMSE) and mean absolute error(MAE), confirming its effectiveness.更多还原