Development of a Dynamic Neural Network Model for Multistep ahead Prediction of Exhaust Gas Temperature in Heavy Duty Gas Turbines (original) (raw)

Several studies has reported the use of neural networks in the dynamic modelling and simulation of heavy duty turbines. However; focus on exhaust gas temperature a key indicator of turbine thermal health is yet to be made. In this paper the modelling of exhaust gas temperature using the non-linear autoregressive network and subsequent multi step prediction with data collected from GT13E2 turbine was embarked upon. Features which were statistically significant for EGT prediction were selected through stepwise regression. One hidden layer was sufficient to approximate the function and The optimal architecture for training was achieved by training the network with a fixed hidden neuron and varying time delay at the inputs and output. It is observed that the optimal performance is realized when the prediction is regressed at tapped input delay of 1 in open loop. 7 hidden neuron and 1 tapped delay is selected for function approximation after series of neurons ranging from 4-15 was tested. The appropriate model was carefully selected by utilizing the method of holdout cross validation, corrected Akaike Informationon Criterion and Schwartz Bayesian information criterion. The final architecture was trained, and converted to close loop NARX network where 100 time steps ahead prediction of EGT was made. Although it was observed that accuracy diminishes as prediction horizon increases, the chosen optimised architecture successfully predicted EGT 100 steps ahead with MAE of 2.9665 and RMSE of 3.9675. Therefore; the dynamic NARX model can be utilized for multistep ahead prediction in incidence of sensor malfunction at the turbine outlet of heavy duty gas turbines.