新型电力系统背景下，各行业用电行为呈现复杂动态特性。准确的区域月电量预测对电力部门进行电网规划和运行调度有重要意义。为准确把握各行业用电量的影响关系进而提升预测精度，提出了基于细分行业多尺度时序图特征挖掘的新型电力系统月电量预测方法。文中建立了多行业时序特征独立编码的多时序注意力并行神经网络以及多时间尺度特征挖掘模块，实现各细分行业在周、月、季度等不同周期用电特性的深度刻画，提出了多行业用户时序耦合挖掘的图卷积神经网络，通过图模型建立不同行业间的相互影响关系，并通过卷积网络解码实现区域用电量预测。基于京津唐电网2020-2022年历史运行数据的算例分析结果表明，预测方法可平均降低RMSE（root mean square error）误差6.58亿千瓦时，平均降低MAPE（mean absolute percentage error）误差1.53%。实验证明了文中月电量预测方法的有效性、准确性和可行性。
英文摘要:
Under the background of the novel power system, the electricity consumption behavior of various industries presents complex dynamic characteristics. Accurate regional monthly electricity prediction is of great significance for the power sector to carry out grid planning and operation scheduling. To accurately grasp the impact relationship of electricity consumption in various industries and improve prediction accuracy, a new monthly electricity prediction method for power systems based on multi-scale time-series graph feature mining of segmented industries is proposed. A multi-industry temporal feature independent encoding multi-temporal-attention parallel network and multi-scale convolutional network are established to achieve deep characterization of the electricity consumption characteristics of various segmented industries in different cycles such as weeks, months, and quarters. A graph convolutional neural network for multi-industry user temporal coupling mining is proposed, which establishes the mutual influence relationship between different industries through graph models, and implements regional electricity consumption prediction through convolutional network decoding. The analysis results based on the historical operation data of the Beijing-Tianjin-Tangshan power grid from 2020 to 2022 demonstrate that the prediction method in this paper can reduce the RMSE error by an average of 658 million kWh and the MAPE error by an average of 1.53%. The experiment has verified the effectiveness, accuracy, and feasibility of the monthly electricity forecasting method proposed in this paper.