摘要：

空间负荷预测为合理建设和使用变电站、馈线等提供了重要的指导，成为配电网规划中不可或缺的一部分。配电网规划的精细化产生了大量高分辨率的负荷数据，社会的快速发展使得地块的用电特征日趋复杂。当前的空间负荷预测没有充分考虑负荷数据之间的时间特性，且在预测过程中也未考虑到不同类型地块间可能存在的负荷峰值出现时间不一致问题。为此，提出一种空间负荷预测方法，通过基于多尺度限制对齐路径长度（LDTW）的谱聚类分析用户的负荷曲线在形状上的相似性，并提取不同地块的典型用电行为，以进一步分类确定同类型地块对应的同时率。多尺度LDTW通过限制序列之间匹配步长的上限来抑制病态匹配的产生，提高曲线相似性的综合评估能力。根据聚类结果筛选适合待预测区域的训练样本并构建基于时间卷积网络(TCN)的回归预测模型，将预测结果基于地块各自的同时率进行聚合，实现空间负荷预测。实验结果表明：该方法加强了对负荷曲线形状的分析和对不同类型地块同时率的区分，在聚类方面，DBI指数达到0.57，Ⅵ指数达到0.31；在预测方面，相对误差达到1.93%，决定系数达到0.941，相比其他典型方法均取得了较大改善。

关键词: 空间负荷预测, 动态时间规整, 谱聚类, 同时率, 时间卷积网络

Abstract:

Spatial Load Forecasting(SLF) provides necessary guidance for the rational construction and use of substations and feeders, and it has become an indispensable aspect of distribution network planning. The refinement of distribution network planning has generated a large number of high-resolution load data, and the rapid development of society has made the electricity characteristics of land plots increasingly complex. The current spatial load-forecasting method does not fully consider the time characteristics between load data and ignores the possible inconsistent time of the peak load between different types of blocks during the forecasting process. The proposed spatial load-forecasting method analyzes the similarity of users'load profiles in shape by spectral clustering based on multiscale LDTW and extracts the typical electricity consumption behaviors of different plots. Based on further classification, the corresponding simultaneity rate of plots of the same type is determined. Multiscale LDTW can inhibit pathological alignment by limiting the upper limit of matching steps between sequences and improve the comprehensive evaluation ability of curve similarity. Based on the clustering results, the training samples suitable for the region to be predicted are screened, and the regression forecasting model based on the Temporal Convolutional Network(TCN) is established. The forecasting results are aggregated based on the simultaneity rates of the blocks to achieve spatial load forecasting. The experimental results show that the proposed method strengthens the analysis of the shape of the load curve and distinguishes the simultaneity rate of different types of blocks. In terms of clustering, the DBI index reaches 0.57, and the Ⅵ index reaches 0.31. In terms of forecasting, the relative error reaches up to 1.93%, and the coefficient of determination is up to 0.941, which indicates a significant improvement compared with other methods.

Key words: Spatial Load Forecasting(SLF), Dynamic Time Warping(DTW), spectral clustering, simultaneous rate, Temporal Convolutional Network(TCN)