基于隐含狄利克雷分布主题模型和特征级异构数据融合的电力故障主动性预警研究

Proactive Warning System Based on Electronic Power User Interactive Complaint Text and Multi-Source Heterogeneous Big Data Analysis

智能化客服系统是国家电网在传统人工客服上转型的重要方向。相对于传统客服,智能客服可以并发处理电力用户的诉求事件,掌握覆盖辖区内配电网准确、可靠、全面、及时的状态信息,并通过分析历史诉求文本数据和电力用户相关的多源异构大数据来积极应付突发事件,对诉求热点进行预测并进行主动性预警。文中首先通过隐含狄利克雷分布概率(LDA)主题模型对电力用户的交互式诉求文本进行主题挖掘,获得诉求用户的诉求主题标签。根据电力公司所收集到的多源异构大数据,文中设计多种特征提取算法,搭建基于卷积神经网络(CNN)和特征级数据融合的分类模型,来实现对未来时间内诉求热点的预测。实验证明LDA模型可以很好地提取出诉求文本中的主题,多源异构数据分类模型最终得到高达94%的分类准确率,相对于传统分类器平均提升12.6%,最终可以实现电力公司对电力故障和用户诉求的主动性预警功能。

The intelligent customer service system is an important direction for the transformation of the state grid in traditional manual customer service.Compared with traditional customer service,intelligent customer service can concurrently process the demand events of power users,grasp the accurate,reliable,comprehensive and timely status information of the power distribution network in the jurisdiction,and analyze the historical demand text and multi-source heterogeneity related to power users for supporting proactive warnings.The latent dirichlet Allocation theme(LDA)model is applied to perform topic mining on the interactive complaint texts of power users and the corresponding appeal topic labels are obtained.Combined with the multi-source heterogeneous data collected by the power company,our work applies different feature extraction algorithm for each data source and build a multi-source data classifier based on the convolutional neural network(CNN)to predict the hot spots of appeals.The experiment proves that the LDA model can extract the topics in the complaint text well,and the multi-source heterogeneous data classifier designed can obtain up to 94%classification accuracy and 12.6%improvement compared to the traditional classifiers.Finally,it can realize the power company’s proactive warning system for electricity interruption and user demands.