一种针对事件级风电故障的高精度预测模型的探究

A Study on High Precision Prediction Model For Event Level Wind Power Stoppage

本文针对常见的风电故障事件进行类别和性质上的分类,并通过分析事件的发生频率、故障率和差异概率等属性对事件类别和事件性质进行深入分析,结合了属性分析、业务逻辑理解和风电故障相关的经验总结,用以进行后续特征工程中的特征构建。在特征工程过程中,我们主要从线性趋势、波变换、分布描述、异常点和检验统计量等几个角度出发,结合了风电故障的特点以及传统时序分析的技巧,并利用主成分分析去除特征的多重共线性并进行数据降维,随后本文利用LightGBM算法构建了针对故障事件的预测,并实现了针对事件级的时间跨度为7天的高精度预测。最终针对整理的23类风电事件,共有12类事件得到有效预测(预测精度超过70%),其中可以精准预测(预测精度超过90%)的有4件,而综合故障预测精度为85.24%。最后本文将LighGBM与传统机器学习算法进行了对比,证明了算法的高可用性,随后本文对模型的优缺点进行分析,肯定了模型在实际应用中的积极意义,最后本文对模型进行了综合的评价并对改进方向提供了探索思路。

Based on common fault events of wind generators,we classified wind events by their categories and properties,and by analyzing the event frequency,failure rate and differential probability and so on,we drilled deeply down in the event category and property,combining with attribute analysis,business logic and fault-related experience for feature construction in subsequent feature engineering.In feature engineering,we started study mainly from the perspectives of linear trend,wave transformation,distribution description,outliers and test statistics,combining the characteristics of wind generators’fault experience as well as the traditional time series analysis technique,and realizing data dimension reduction by using principal component analysis to remove multicollinearity of features.By the time we used LightGBM algorithm for fault prediction of events,we had built the high precision of prediction for event in seven days in future.Finally,12 types of events were effectively predicted(with a prediction accuracy of over 70%)in the 23 types of wind generator events,among which 4 types of events could be accurately predicted(with a prediction accuracy of over 90%),while the comprehensive fault prediction accuracy was 85.24%.In the end,we compared LightGBM with traditional machine learning algorithms and proved the high availability of the algorithm.Then,we analyzed the advantages and disadvantages of the model and affirmed the positive significance of the model in practical application.Finally,we made a comprehensive evaluation of the model and provided an exploration idea for the improvement direction in future.