新能源大规模接入电网，要求发电型燃气轮机频繁切换工作状态，更容易发生故障，异常检测对燃气轮机安全运行更加重要。针对燃气轮机异常检测问题，提出了一种基于NARX的基线建模方法。利用NARX模型对燃气轮机声压特征信号建立基线模型，利用CatBoost算法增强NARX拟合效果，利用贝叶斯优化对模型超参数优化，增强模型泛化能力，使用燃气轮机实机数据进行测试。结果表明：NARX-CatBoost方法明显优于NARX-FROLS和edRVFL方法。该方法对正常声压均方根的拟合RMSE值0.00850，拟合准确度明显优于其他方法；该方法对异常声压均方根的异常检测准确率为96.94％，说明通过正常声压特征数据建立基线模型进行异常检测的可行性与准确性。

The large-scale access of new energy to the power grid requires the frequent switching of the working state of the power generation gas turbine, which is more prone to failure. Anomaly detection is more important for the safe operation of the gas turbine. Aiming at the problem of gas turbine anomaly detection, a baseline modeling method based on NARX is proposed. The NARX model is used to establish a baseline model for the sound pressure characteristic signal of the gas turbine. The CatBoost algorithm is used to enhance the NARX fitting effect. Bayesian optimization is used to optimize the hyperparameters of the model and enhance the generalization ability of the model. The gas turbine real machine data is used for testing. The results show that the NARX-CatBoost method is significantly better than the NARX-FROLS and edRVFL methods. The fitting RMSE value of the normal sound pressure root mean square is 0.00018, and the fitting accuracy is obviously better than other methods. The accuracy of this method for anomaly detection of abnormal sound pressure root mean square is 94.18 %, which shows the feasibility and accuracy of establishing a baseline model for anomaly detection through normal sound pressure characteristic data.

TK221

黑龙江省重点研发计划，NO.GA23A907.