Abstract: Aiming at the problems of operation errors and work delays in the wind turbine overhauls business process,the predictive monitoring method of business process was applied to predict the next event,execution time and remaining time of the business for reminding the staff to prevent and avoid risks.According to different prediction tasks,a priority-based feature self-selection strategy was proposed,and the input features that had positive impact on the prediction results was obtained by using Light Gradient Boosting Machine(LightGBM)algorithm as the prediction model of feature selection strategy.Then,for different prediction tasks,LightGBM algorithm and Long Short Term Memory(LSTM)neural network were used to build prediction models.Through experimental evaluation and analysis,the feature selection strategy could provide effective features for prediction tasks,ensure the accuracy of prediction,and has practical application value in the wind turbine overhauls business process.For different prediction tasks,LightGBM algorithm was more suitable for next event task prediction,and LSTM model was more suitable for time task prediction.

Key words: process prediction monitoring, wind turbine overhauls, feature selection, next event, remaining time

摘要： 针对风机检修业务流程中存在的操作失误和工作延期等问题,应用业务流程预测性监控方法,预测业务的下一事件、下一事件执行时间和剩余时间,以提醒工作人员预防和避免风险的发生。首先,针对不同预测任务,提出一种基于优先级的特征自选取策略,并使用LightGBM(Light Gradient Boosting Machine)算法作为特征选择策略的依托预测模型,得到对预测结果有积极影响的输入特征;然后,针对不同预测任务分别采用LightGBM算法和LSTM(Long Short Term Memory)神经网络构建预测模型;最后,经实验评估和分析,在风机检修业务流程中,特征选择策略能够为不同的预测任务提供有效特征,确保预测的准确率,具有实际应用价值。对于不同预测任务而言,LightGBM算法更适用于下一事件任务预测,LSTM模型更适用于时间方面的任务预测。

关键词: 流程预测性监控, 风机检修, 特征选择, 下一事件, 剩余时间