Data Driven Modelling of Centrifugal Compressor Maps for Control and Optimization Applications

This repository hosts the source code of the paper: "Data Driven Modelling of Centrifugal Compressor Maps for Control and Optimization Applications".

How to use?

As explained in the paper, the proposed model approximates compressor efficiency and pressure ratio using Gaussian Process with Bayesian Optimisation, and Polynomial Regression. GP2D_scaled.py produces optimised kernels for each defined kernel type in sklearn with their hyperparameters and error metrics comparing to Polynomial regression. It predicts efficiency or pressure ratio over validation data with the best-performed model and saves the results. Different sample sizes and sampling methods may be assigned to analyse the performance of distinct models.

Comparison of different polynomial degrees and GPR performance is also possible using handcrafted_models.py. Line plots with uncertainty quantification can be obtained using this function by setting the "plot" parameter to True.

Contour plots for representing approximated efficiency map, scatter plot to see speed lines produced by pressure ratio approximation of GP and their respective shape to true ones, and surface plots of both GP and Polynomial regression can be obtained by plot.py

Available datasets

Currently, following datasets are ready to use in pressure ratio and efficiency regression.

Compressor name	Data source name
EFR 91S74	etas_full
Garrett 3076R	garrett_full1
Garrett 1544	garrett_full2
Garrett TO4B	garrett_to4b

Example estimated compressor efficiency maps

EFR 91S74:

GPR	Polynomial regression

Garrett 1544:

GPR	Polynomial regression

Garrett 3076R:

GPR	Polynomial regression

Garrett TO4B:

GPR	Polynomial regression

Example estimated speed lines

Garrett 1544:

Garrett 3076R:

Garrett TO4B: