Data-driven enhanced first-principles modeling for centrifugal compressor performance prediction in compressed air energy storage systems

Centrifugal compressors are critical to compressed air energy storage (CAES) systems, where accurate and rapid performance prediction is essential for charge–discharge cycle design, off-design analysis, and control strategy optimization. Existing first-principles one-dimensional (1D) models still have some issues, such as accuracy degradation at high rotational speeds and the absence of efficiency prediction. To improve pressure ratio prediction, this paper proposes a methodology that couples an improved pressure ratio model with a machine-learning-based efficiency surrogate. Meanwhile, four physical improvements are introduced into the pressure ratio model: outlet blade angle, slip factor, partitioned friction parameters, and incidence losses. To better efficiency prediction, a surrogate modeling strategy employing temperature-difference as the meta-variable is proposed, and three regression methods—Gaussian process regression (GPR), support vector machine (SVM), and neural network (NN)—are evaluated. Validations of the improved model were carried out on two compressor test setups. The results reveal that the improved pressure ratio model achieves a root mean square error (RMSE) of 0.0424 and a coefficient of determination (R 2 ) of 0.9758, outperforming the original model (RMSE = 0.0537, R 2 = 0.9613). The SVM-based efficiency model yields the best performance (RMSE = 0.0487, R 2 = 0.9125). Requiring only limited experimental calibration, the proposed method enables complete performance maps to be predicted within seconds, offering a computationally efficient tool for CAES system design, compressor–turbine matching, and operational optimization.