基于残差分组卷积神经网络和多级注意力机制的源荷极端场景辨识方法  

Source-load Extreme Scenarios Recognition Based on GCNN-resnet Network and Multi-level Attention Mechanism

作　　者：郭红霞[1] 李渊[1] 陈凌轩 王建学[2] 马骞[3] GUO Hongxia;LI Yuan;CHEN Lingxuan;WANG Jianxue;MA Qian(School of Electric Power,South China University of Technology,Guangzhou 510640,Guangdong Province,China;School of Electrical Engineering,Xi’an Jiaotong University,Xi’an 710049,Shaanxi Province,China;Power Dispatching and Control Center of China Southern Power Grid,Guangzhou 510663,Guangdong Province,China)

机构地区：[1]华南理工大学电力学院,广东省广州市510640 [2]西安交通大学电气工程学院,陕西省西安市710049 [3]中国南方电网电力调度控制中心,广东省广州市510663

出　　处：《电网技术》2025年第2期459-469,I0019-I0024,共17页Power System Technology

基　　金：国家重点研发计划项目(2022YFB2403500)。

摘　　要：为应对极端天气事件给新型电力系统安全稳定运行带来的影响,在电网的生产模拟中需要考虑极端场景。然而极端场景历史样本数量少,传统场景生成方法无法直接生成极端场景,需要对场景进行辨识。为此,提出一种计及源荷双侧的极端场景辨识方法。首先,将风电、光伏和负荷序列进行重塑,并在通道维度上拼接;然后,基于分组卷积和深度残差网络,提取场景的时序特征和源荷场景之间的耦合特征;其次,模型内部嵌入通道注意力机制和多头注意力机制,以赋予重要特征更大的权重,并对场景进行分类;此外,采用改进损失函数解决训练样本中数据集不均衡的问题;最后,基于历史数据集进行验证。验证结果表明,所提方法能够对场景进行有效的分类,可以从历史场景中识别出具有高保供或高消纳风险的源荷极端场景。To address the challenges posed by extreme weather events to the safe and stable operation of new power systems,it is imperative to consider extreme scenarios in the production simulation of the power grid.Traditional scenario generation methods cannot directly produce extreme scenarios due to the scarcity of historical extreme samples,necessitating the recognition of scenarios.To address this,the paper proposes an extreme scenario identification method that incorporates the bilateral aspects of source-load dynamics.Wind,photovoltaic(PV),and load sequences are initially reshaped and concatenated along the channel dimension.Subsequently,the temporal features of the scenario and the coupling features between the source-load scenarios are extracted using grouping convolution and deep residual networks.Furthermore,the model integrates a channel attention mechanism and a multi-head attention mechanism to assign greater significance to critical features and categorize the scenarios.Additionally,the issue of imbalanced datasets in the training samples is addressed by employing an improved loss function.Finally,validation is conducted using historical datasets.The results demonstrate the effectiveness of the proposed method in accurately classifying scenarios,particularly in identifying source-load extreme scenarios with risk of power supply guarantee or renewable energy consumption from the historical dataset.

关 键 词：极端场景辨识 残差神经网络 分组卷积 注意力机制 源荷不确定性 

分 类 号：TM721[电气工程—电力系统及自动化]