基于随机森林和LSTM-自编码算法的风机高温降容状态智能检测方法  

Intelligent Detection Method of High-temperature Capacity Reduction of Wind Turbine Based on Random Forest and LSTM-Autoencoder

作　　者：张国珍[1] 王其乐 叶天泽 杨锡运[3] ZHANG Guozhen;WANG Qile;YE Tianze;YANG Xiyun(China Longyuan Power Group Co.,Ltd.,Beijing 100034,China;Zhongneng Power-Tech Development Co.,Ltd.,Beijing 100034,China;School of Control and Computer Engineering,North China Electric Power University,Beijing 102206,China)

机构地区：[1]龙源电力集团股份有限公司,北京100034 [2]中能电力科技开发有限公司,北京100034 [3]华北电力大学控制与计算机工程学院,北京102206

出　　处：《华北电力大学学报（自然科学版）》2021年第3期81-88,共8页Journal of North China Electric Power University：Natural Science Edition

基　　金：国家自然科学基金资助项目(51677067)。

摘　　要：发电机、齿轮箱等风机部件的高温降容状态是表征风电机组亚健康状态的良好指标,其评估的准确性直接影响了后期人员、设备、资金等多种资源的投入的多少,以及运维方案的最终效果。为了尽可能真实客观反映风机高温降容状态,提出了一种基于随机森林和长短时记忆网络自编码(Long Short Term Memory-Autoencoder,LSTM-Aec)算法相结合的智能评估方法,该方法首先采用随机森林算法对SCADA系统采集的数据进行特征约简,再利用LSTM-Aec算法对风机高温降容状态进行评估检测。测试结果表明,基于该方法的风机高温降容状态评估的精确率和准确率分别达0.9809和0.9558,整体优于未经过特征约简的RNN-Aec算法和未经过特征约简的LSTM-Aec算法的评估检测方法以及传统分类算法。The high temperature capacity reduction state of wind turbine components such as generator and gearbox well indicates the sub-health state of wind turbines.Its evaluation directly determines the following investment of personnel,equipment,cost and other resources,as well as operation and maintenance plans.In order to reflect the wind turbine′s high temperature capacity reduction state as truthfully and objectively as possible,this paper proposes an intelligent evaluation method based on the combination of random forest and Long short term memory-Autoencoder(LSTM-Aec)algorithm.This method first uses the random forest algorithm to perform feature reduction on the data collected by the SCADA system.And then it uses the LSTM-Aec algorithm to evaluate and detect the high temperature capacity reduction state of the wind turbine.The test results show that the precision and the accuracy of the high temperature capacity reduction state evaluation of the wind turbine based on this method are 0.9809 and 0.9558 respectively,which are higher than those of the RNN-Aec algorithm without feature reduction and the LSTM-Aec algorithm without feature reduction and traditional classification algorithms.

关 键 词：随机森林 长短时记忆网络 高温降容 状态评估 特征约简 

分 类 号：TM614[电气工程—电力系统及自动化]