基于并行的F-LSTM模型及其在电力通信设备故障预测中的应用  

Parallel-Based F-LSTM Model and Its Application in Power Communication Fault Prediction

作　　者：杨济海 刘洋[2] 刘杰 余伟[2] 李石君[2] YANG Jihai;LIU Yang;LIU Jie;YU Wei;LI Shijun(Information & Telecommunication Branch of State Grid Jiangxi Electric Power Company, Nanchang 330077, Jiangxi, China;School of Computer Science, Wuhan University, Wuhan 430072, Hubei, China)

机构地区：[1]国网江西省电力有限公司信息通信分公司,江西南昌330077 [2]武汉大学计算机学院,湖北武汉430072

出　　处：《武汉大学学报（理学版）》2019年第3期263-268,共6页Journal of Wuhan University:Natural Science Edition

基　　金：国家自然科学基金青年基金(61502350)

摘　　要：电力通信网设备时序故障预测的目标是通过过去设备告警数据,预测设备在下一个时间段是否发生故障,这对设备的管理和维护起着重要作用。为了预测电力设备未来的状态,提出一种Forward-LSTM(F-LSTM)学习模型,对设备故障的时序特征和非时序特征(静态信息)进行并行训练,探索出一种新的对静态-时序数据的训练方法,将其应用在电力通信网故障预测中。F-LSTM结合了两个组件,一个学习时序特征的长短期记忆神经网络(LSTM)与一个处理静态数据的前向全连接神经网络(forward full connection neural networks,FC),数据的静态/时序属性被自动判断并传递给FC或LSTM来并行训练。对于具有同时产生动态数据与静态数据的电力通信网络,Forward-LSTM(F-LSTM)模型能以较高速度与精度预测其故障发生的位置。此外,本文采用一种加权的损失函数,可以更好地捕捉设备故障的时序规律。选取某电力通信网络系统中2016—2017年设备故障数据,对本方法进行测试。实验结果显示,与Xgboost模型相比,F-LSTM模型对故障预测的召回率提高5%,同时F-LSTM模型较LSTM模型缩减了计算量,加快了模型的训练速度。The goal of the equipment timing failure prediction of power supply network is to predict whether the equipment will fail in the next time period through past equipment alarm data, which plays an important role in equipment management and maintenance. In order to predict the future state of power equipment, this study proposes a Forward-LSTM(F-LSTM) learning algorithm to train the timing and non-timing features(static information)of equipment faults in parallel, and explore a new training method for static-time series data. It is used in power communication network fault prediction. F-LSTM combines two components, a long short-term memory(LSTM) neural network that learns timing characteristics and a forward full connection neural networks(FC) that processes static data. The static/timing properties of the data are automatically judge and pass to the FC or LSTM for parallel training. For power communication networks with both dynamic and static data, the Forward-LSTM model predicts where faults occur at higher speeds and accuracy. In addition, this paper uses a weighted loss function to better capture the timing of device failures. This method was tested by selecting the equipment failure data of 2016-2017 in a power communication network system. The experimental results show that the F-LSTM model improves the recall rate of fault prediction by 5% compared with the Xgboost model, while the calculation amount is reduced, and the training speed of the model is accelerated compared with the LSTM model.

关 键 词：电力通信网 故障预测 LSTM Xgboost F-LSTM 

分 类 号：TN95.853[电子电信—信号与信息处理]