Objective In order to overcome the influence of the nonlinear time-varying characteristics of gas turbines on dynamic control and performance monitoring, this paper combines the time series memory and nonlinear relation expression ability of a long short-term memory neural network (LSTM) with the interval probability estimation ability of Gaussian process regression (GPR) to propose an online parameter identification algorithm for the key dynamic parameters of gas turbines based on an LSTM and GPR-based hybrid deep learning model (LSTM-GPR).
Methods First, the dynamic mechanism model of a gas turbine is established, and a large amount of training data is generated by taking fuel calorific value, compressor efficiency and load power moment as the parameters to be identified. Next, the parameter identification network model of LSTM-GPR is constructed, and the training data is input for network training and weight coefficient learning. Finally, the trained LSTM-GPR hybrid deep learning model is used to identify the dynamic operating parameters of the gas turbine model online, and the identification results are analyzed to verify the effectiveness of the proposed algorithm.
Results The simulation results show that the online identification results of the proposed LSTM-GPR hybrid model algorithm are accurate, with a recognition error of less than 1% and good real-time performance. Compared with the LSTM single model, the proposed algorithm can obtain a better mean estimation effect and provide a reliable confidence interval range.
Conclusions The LSTM-GPR hybrid algorithm can be effectively applied to the online parameter identification of a gas turbine model, laying a foundation for its further application to the dynamic operation parameter identification of practical units.
DownLoad: