基于极端不平衡学习的泛化低压异常箱表关系识别研究与应用

Generalized Identification for Low-Pressure Abnormal Box-Table Relationship Based on Extreme Unbalance Classification Learning

针对低压配电网箱表关系存在人工核查成本高、异常案例少、难以实现异常规律捕获的问题,采用极端不平衡分类学习方法实现低压异常箱表关系识别的泛化应用推广.通过电压原理识别出部分异常箱表关系样本集,随后构建CNN(卷积神经网络)异常箱表关系识别模型,通过样本三分类赋权值实现类别均衡处理;并在模型推广应用过程中,采用强化学习实现离线模型的在线泛化学习,并以分组模型交互学习和竞争优化的方式筛选出最优泛化识别模型.实验证明,通过人工核查和数据反馈,该方法可实现模型对异常样本数据分布规律的自拟合学习,提高模型对不同应用环境的泛化性,进一步降低人工现场核查工作量,保障低压台区用户拓扑网络关系的准确性.

Due to high labor cost and few abnormal cases of power box-table relations inspection,which difficulty to obtain the law.The extreme unbalanced classification learning method was used to capture the generalization.Through the principle of voltage,abnormal box-table relationship sample sets were identified.And by three-class weighting balance,the CNN(convolutional neural network)abnormal box-table relationship recognition model was constructed.In addition,the grouped parallel generalization learning of recognition model was realized by reinforcement learning.The experiment proves that,through self-learning the distribution of newly identified abnormal sample data,which improve the generalization to different environments.This reduces the workload of manual on-site verification and ensures the accuracy of the topology network relationship in the low-voltage station area.