Wide Range Performance Measurement and Low-Fidelity Modelling of Turbocharger Turbines for Optimising Powertrain Efficiency

Abstract

Turbochargers have been widely used for many years and have made remarkable contributions in reducing the vehicle emissions by reducing the engine size while maintaining the same power output. They are also needed in designs using alternative fuels, for example, hydrogen engines, hydrogen fuel cell systems and so forth. The requirements of designing turbochargers with higher performance over wider operating conditions for different applications are thus increasing rapidly. This thesis aims to achieve wider range turbine performance measurements and more accurate low fidelity modelling of turbocharger turbines to enable better and more efficient overall powertrain system design and optimization procedures. An instability issue identified in the viscous dynamometer established at Queen?s University Belfast was numerically investigated by CFD simulation using Ansys CFX with R-P cavitation model. The knowledge obtained was used to propose new sleeves designs which feature the addition of grooves onto the inner side of the sleeve. The cavitation modelling results indicated that using spiral grooves at the bottom of the feedholes decreased the cavitation area effectively and increased the critical speed where the cavitation became significant. The modified dynamometer with the manufactured new grooved sleeve has been tested and verified the stability improvement. The testing results showed that the stable operating range was expanded to 130k rev/min at low torque, corresponding to more than 198.8k rev/min at a turbine inlet temperature of 600?. A new low fidelity model for radial turbine wide range performance prediction based on a hybrid combination of conventional meanline method and prevailing machine learning techniques was then developed. The ANN-meanline model with trained ANN loss and blockage models was formulated using results of around 2500 CFD simulations and Python-coded meanline calculations. The novel steps included the division of the losses into those occurring before and after the throat location and the utilization of a variable blockage factor defining the ratio of the effective flow area over the geometrical area at the turbine throat. The hybrid model was evaluated on four numerical test cases spanning the design space and two experimental test cases. The results showed that the hybrid model, independent of empirical coefficients and loss calibration, improved the predictive capability for wide range turbine mass flow rate and efficiency predictions with an average error of 3.59% and 1.42%pts. The hybrid model was evaluated and enhanced by implementing advanced Bayesian Optimization into hyperparameter optimization and adaptive sampling process. By adding 540 CFD simulations into the training database, the enhanced hybrid model showed great capability in wide ranges of turbine mass flow rate parameters and efficiency prediction with average discrepancies of 2.02% and 1.174%pts. It just took three to five minutes to calculate a turbine performance map, which is magnitude faster than CFD modelling, without any significant coding optimization.