针对燃气轮机气路性能退化的周期性与非线性特点，提出一种基于长短期记忆（Long-Short Term Memory,LSTM）神经网络的燃气轮机排气温度趋势预测方法。利用标准化与归一化数据预处理方法提取排气温度数据中的退化特征，减小了环境和工况变化对退化特征的影响；通过滑窗法提取一定长度的历史数据，增强LSTM模型的泛化能力；利用LSTM网络的内部循环节点，通过时序相似性搜索，实现退化周期燃气轮机排气温度预测。最后，采用船用燃气轮机水洗周期历史数据进行排气温度趋势预测试验。研究表明：基于LSTM排气温度预测方法的预测精度大于87.4%，且预测结果的波动性和滞后性较小。

Aiming at the periodic and nonlinear characteristics of gas path performance degradation of gas turbine,a kind of prediction method of gas exhaust temperature trend based on longshort term memory (LSTM) neural network was proposed.The standardized and normalized data preprocessing method was used to extract the degradation features of exhaust temperature data,which reduced the influence of environment and working condition changes on the degradation characteristics.The sliding window method was used to extract a certain length of historical data to enhance the generalization ability of LSTM model.The internal cycle nodes of LSTM network were used to realize the gas turbine exhaust temperature prediction in degradation cycle by means of the sequential similarity search.Finally,the historical data of marine gas turbine in washing cycle was used for the exhaust temperature trend prediction test.The research results show that the prediction accuracy of exhaust temperature prediction method based on LSTM is greater than 87.4%,and the fluctuation and hysteresis of the prediction results are small.

TK478

国家科技重大专项（2017-Ⅰ-0007-0008)