Simulation modelling and comparison of different training algorithms for multistep prediction

This study investigates nonlinear autoregressive neural network (NARNET) and nonlinear autoregressive neural network with exogenous input (NARXNET)-based artificial neural network (ANN) models for multistep prediction of specific enthalpy of steam. Real-time experimental data on specific enthalpy of steam has been collected and used for training of proposed models. The machine learning models have been trained using different training algorithms namely Levenberg-Marquardt (LM), Bayesian-regularisation (BR), scaled-conjugate gradient (SCG), one step secant (OSS) and resilient back-propagation (RB). The prediction performance of these algorithms has been analysed in terms of root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and bivariate correlation coefficient (COR) for a maximum step size of 30 multistep predictions. The results highlight superior performance of NARXNET model designed using BR-algorithm compared to prediction models designed using other training algorithms.