基于CatBoost机器学习模型的风电机组机舱温度异常预警研究

RESEARCH ON EARLY WARNING OF ABNORMAL TEMPERATURE IN WIND TURBINE NACELLE BASED ON CatBoost MACHINE LEARNING MODEL

风电机组机舱内部的机械部件众多,以机舱温度为研究对象可以实现对风电机组故障的预警。首先提取风电机组正常运行状态下的机舱温度数据,综合Pearson相关系数和Spearman相关系数,以及轻型梯度增强学习器(LightGBM)和CatBoost算法的特征变量重要性,筛选出与机舱温度相关性较大的20个特征变量,作为风电机组机舱温度的特征变量集合;然后选择CatBoost、LightGBM、随机森林(Random Forest)3个算法分别建立模型,以均方误差(MSE)、均方根误差(RMSE)、平均绝对误差(MAE)、判定系数R2作为评价指标进行综合评价,最终选择评价指标最优的采用CatBoost算法建立的模型作为风电机组机舱温度异常预警模型,并采用实际的风电机组机舱温度异常的历史数据对模型的预警效果进行验证。该模型可在机舱温度预测值与真实值之间偏离程度较大时进行预警,专业检修人员可以根据模型输出的特征变量重要性排序,优先检修相关性较高的部件,实用性较强。

There are many mechanical components in the nacelle of wind turbine,and taking the nacelle temperature as the research object can realize the early warning of wind turbine failure.Firstly,this paper extracts the nacelle temperature data under the normal operation of the wind turbine,integrates Pearson correlation coefficient and Spearman correlation coefficient,as well as the importance of the characteristic variables of light gradient boosting machine(LightGBM)and CatBoost algorithm,and selects 20 characteristic variables that have a greater correlation with the nacelle temperature as the characteristic variable set of the nacelle temperature of the wind turbine.Then three algorithms,namely CatBoost,LightGBM and Random Forest,are selected to establish models respectively,and mean squared error(MSE),root mean square error(RMSE),mean absolute error(MAE)and decision coefficient R2 are used as evaluation indicators for comprehensive evaluation.Finally,the model established by CatBoost algorithm with the best evaluation indicators is selected as the early warning model of abnormal temperature in wind turbine nacelle.The warning effect of the model is verified by using the actual historical data of abnormal temperature in the wind turbine nacelle.This model can give an early warning when the deviation between the predicted value of nacelle temperature and the real value is large.Professional maintenance personnel can prioritize the maintenance of components with high relevance according to the importance ranking of the characteristic variables output from the model,which is more practical.