基于长短时记忆网络融合SCADA数据的风电齿轮箱状态监测

针对不具有时间记忆能力的机器学习方法融合风电机组数据采集与监控系统(SCADA)的时序数据而导致风电齿轮箱状态预测精度不高的问题,提出基于长短时记忆(LSTM)网络融合SCADA数据的风电齿轮箱状态预测模型。选择能表征风电齿轮箱运行状态的某个监测量作为模型的输出量,基于灰色关联度选择与该监测量关联密切的SCADA参数作为预测模型的输入量;使用正常状态下的SCADA数据训练LSTM预测模型,得出预测值和残差,通过3σ准则计算出上下预警阈值,用于风电齿轮箱状态监测和故障预警。某风电场风电齿轮箱的SCADA数据验证表明所提出的方法能有效预警风电齿轮箱故障。

CONDITION MONITORING OF WIND TURBINE GEARBOX BASED ON LSTM NEURAL NETWORK FUSING SCADA DATA

Due to existing machine learning methods that fusing SCADA time series data without time memory capability may lead to low accuracy for wind turbine gearbox condition prediction,a model based on a LSTM neural network fusing SCADA data is proposed to solve this problem. Firstly,selecting one monitoring parameter which can reveal the working operation condition of gearbox as the model output,and grey correlation analysis method is used to select SCADA parameters closely related to the monitoring parameter as the model inputs. Then,the LSTM prediction model is established using healthy working condition data to calculate the prediction values and residuals. Finally,the upper and lower thresholds to monitor the condition of wind turbine are calculated based on the three sigma rule. The results of experiment adopted the measured SCADA data of a wind farm show that the proposed model can effectively realize the wind turbine gearbox fault warning.