THE USE OF ARTIFICIAL NEURAL NETWORKS FOR PERFORMANCE PREDICTION OF RETURN CHANNELS FOR INDUSTRIAL CENTRIFUGAL COMPRESSORS

摘要

An innovative procedure for the preliminary design and optimization of return channels for centrifugal compressors is explained. A typical configuration of bladed return channel for industrial centrifugal compressors is analyzed by means of a well-known commercial Navier-Stokes solver. A set of geometrical parameters groups is chosen in order to represent the most significant changes in geometry with respect to the base configuration. A series of new return channel configurations is obtained as the result of variations of one or more geometrical parameters. Each geometry obtained with this procedure is analyzed by the flow solver which returns a set of accurately chosen performance indices quantifying aerodynamic losses and distortions at the eye of the downstream impeller. The results thus obtained are used to train a simple one-layer Neural Network (NN) which is afterwards interrogated to compute some performance maps linking the performance indices to the three most relevant geometrical parameters. A further computation is carried out on some return channel configurations which have not been previously analyzed. The results confirm that the interpolator is able to predict the return channels performances with a good accuracy. The resulting performance maps, validated by some random tests, seem to be a valid tool for performance prediction of this kind of return channels.