| 引用本文: | 谢从珍,周晓静,余 松,等.基于先验知识与孪生网络监督的输电线路山火跳闸预测模型[J].电力系统保护与控制,2025,53(21):72-83.[点击复制] |
| XIE Congzhen,ZHOU Xiaojing,YU Song,et al.Wildfire-induced transmission line trip prediction model based on prior knowledge and Siamese network supervision[J].Power System Protection and Control,2025,53(21):72-83[点击复制] |
|
| 摘要: |
| 准确有效地预测输电线路山火跳闸事件对电网安全运行至关重要。但历史跳闸数据存在小样本不平衡问题，使得机器学习模型易将跳闸事件误分类为正常，影响预测准确性。为缓解传统样本不平衡方法中信息噪声引发的模型坍塌风险，提出了一种基于先验知识与孪生网络监督的山火跳闸预测模型。首先，在原始跳闸数据集的基础上，基于多元概率统计方法确定生成虚拟样本的数量，缓解小样本不平衡问题。其次，基于先验知识约束的生成式过采样法生成虚拟正样本，以修正数据集正样本分布。然后，采用孪生网络模型筛选虚拟样本，使虚拟正样本具备与真实数据相匹配的特征。最后，将支持向量机(support vector machine, SVM)作为二分类模型对山火条件下线路跳闸进行预测。通过高质量、低需求的数据生成，预测模型召回率相比于常规方法最大可提升31.94%，有效提高了实际工程环境下山火跳闸事件的预测效果。 |
| 关键词: 山火 跳闸 样本生成 不平衡数据 数据驱动 |
| DOI：10.19783/j.cnki.pspc.250046 |
| 投稿时间：2025-01-14修订日期：2025-03-26 |
| 基金项目:国家自然科学基金项目资助(51977084)；国家电网有限责任公司科技项目资助(SGNC0000DKJS2210093) |
|
| Wildfire-induced transmission line trip prediction model based on prior knowledge and Siamese network supervision |
| XIE Congzhen,ZHOU Xiaojing,YU Song,MO Ziyang,HUANG Mengcheng,LAN Ziyi |
| (School of Electrical Power, South China University of Technology, Guangzhou 510641, China) |
| Abstract: |
| Accurately and effectively predicting wildfire-induced transmission line trip events is crucial for the safe operation of power grids. However, historical tripping data suffer from small sample imbalance, which makes machine learning models prone to misclassifying trip events as normal, thereby reducing prediction accuracy. To mitigate the risk of model collapse caused by informational noise in traditional sample imbalance handling methods, this paper proposes a wildfire-induced transmission line trip prediction model based on prior knowledge and Siamese network supervision. First, based on the original trip dataset, a multivariate probability statistical method is used to determine the number of virtual samples to be generated, thereby alleviating the small sample imbalance issue. Second, a generative oversampling method constrained by prior knowledge is applied to generate virtual positive samples and correct the distribution of positive samples in the dataset. Then, a Siamese network model filters the virtual samples, ensuring that generated positive samples match the characteristics of real data. Finally, a support vector machine (SVM) is employed as a binary classifier to predict line trips under wildfire conditions. Through high-quality and low-demand data generation, the proposed model improves the recall rate by up to 31.94% compared to conventional methods, effectively enhancing the prediction performance of wildfire-induced trip events in practical engineering environments. |
| Key words: wildfire trip sample generation imbalanced data data-driven |
