Neural network approach to compressor modelling with surge margin consideration

Artificial neural networks are gaining popularity thank to their fast and accurate response paired with low computing power requirements. They have been proven as a method for compressor performance prediction with satisfactory results. In this paper a new approach of artificial neural networks modelling is evaluated. The auxiliary parameter of 'relative stability margin Z' was introduced and used in learning process. This approach connects two methods of compressor modelling such as neural networks and auxiliary parameter utilization. Two models were created, one with utilization of the `relative stability margin Z' as a direct indication of surge margin of any estimated condition, and other with standard compressor parameters. The results were compared by determination of fitting, interpolation and extrapolation capabilities of both approaches. The artificial neural networks used during the process was a two-layer feed-forward neural-network with Levenb erg-Marquardt algorithm with Bayesian regularization. The experimental data was interpolated to increase the amount of learning data for the neural network. With the two models created, capabilities of this relatively simple type of neural-network to approximate compressor map was also assessed.