风机持续健康稳定运行是电站机组安全性与经济性的重要保障，故障预警技术对于提高风机运行可靠性和降低维护成本尤为重要。为此，本文提出一种基于长短期记忆(Long short-term memory,LSTM)神经网络与贝叶斯优化算法的早期故障预警方法，充分挖掘电站风机正常运行数据，采用LSTM网络挖掘多种参数的关联特性及历史数据的时序特性，建立风机运行状态预测模型。为了提高预测模型的精确度，利用贝叶斯优化算法优化并设定LSTM网络的最佳超参数组合。考虑模型预测偏离度的非平稳性和多极值特点，引入广义极值理论从正常运行工况中确定报警阈值，以实现设备的早期故障预警。最后，将所提出的算法应用于某燃煤电站引风机故障预警中。结果表明：贝叶斯优化算法优化后的LSTM神经网络不仅可以精确表征风机在正常状态下运行行为，同时能够准确地获取风机的故障信息，从而能够在故障发生前4 h发现异常，实现故障预警。

The continuous healthy and stable operation of fans is an important guarantee for the safety and economy of power station units. Fault warning technology is particularly important to enhance the operating reliability of power plant fans and reduce maintenance costs. Thus, an early fault warning method based on long shortterm memory (LSTM) neural network and Bayesian optimization (BO) algorithm is proposed in this paper. By making full use of the normal operation data of fans, LSTM network is used to mine the correlation characteristics of various parameters and the time series characteristics of historical data, and the prediction model of fan operation state is established. In order to improve the accuracy of the prediction model, BO algorithm is used to optimize and set the optimal hyperparameter combination of LSTM network. Considering the nonstationarity and multipleextremum characteristics of the model prediction deviations, generalized extreme value theory is introduced to determine the alarm threshold from normal operating conditions and to realize the early fault warning of equipment. Finally, the proposed algorithm is applied to detect the early fault alert of an induced draft fan in a coalfired power plant. The results show that the LSTM neural network optimized by Bayesian optimization algorithm can not only describe the normal operation behavior of the fan precisely, but also obtain the fault information accurately. Thus, the anomaly can be found 4 hours before the fault occurs,so as to realize fault warning.

TP277

国家重点研发计划课题（2021YFB2601405）