多模型分层融合的配用电系统用户数据识别

User data identification of power distribution system based on multi-model hierarchical fusion

针对配用电系统用户数据识别中特征选择困难和单模型分类精度不高的问题,本文提出多模型分层融合的识别方法。首先,设计多尺度联结的递归差分卷积网络对用户数据进行特征提取,使浅层融合的有效信息不会随着层数的增长而消失;其次,改进自适应学习率优化算法训练模型,增加模型分类性能;最后对6种基模型分层加权融合,以阈值划分层级,新定义混淆矩阵的错误样本数确定权值,有效克服了基模型适应度缺陷。融合方法在用户特征库的识别上获得99.43%的准确率,比传统卷积神经网络与加权融合方法的准确率分别提高0.53%、0.47%,实现了对用户数据的特征提取和高准确率识别,有助于提高配用电系统信息处理和智能决策水平,为电力需求侧用户服务和经营管理提供支撑。

Aiming at the difficulty of feature selection and the low accuracy of single-model classification in the identification of user data in power distribution systems,a multi-model hierarchical fusion identification method is proposed.First,a multi-scale connected recursive differential convolutional network is designed to extract features from user data,so that the effective information of shallow fusion will not disappear with the increase of the number of layers.Secondly the adaptive learning rate optimization algorithm training model is improved to increase the model classification performance.Finally,the six basic models are layered and weighed and merged,and the levels are divided by thresholds.The number of incorrect samples of the newly defined confusion matrix is determined to determine the weight,which effectively overcomes the fitness defect of the basic model.In the recognition of user feature database,the accuracy of the optimal basic model is 0.35%higher than that of the traditional convolutional neural network,and the accuracy of the fusion method is 0.47%higher than that of the traditional weighed fusion method,which realizes the feature extraction and high accuracy of the user data.Identification helps to improve the information processing and intelligent decision-making level of the power distribution system,and provides support for the service and operation management of power demand-side users.