Data-driven optimization of auto-cascade refrigeration system via machine learning and evolutionary multitasking

This study proposes a data-driven intelligent decision-making framework for screening low-global-warming-potential (GWP) refrigerants and performing multi-objective optimization of auto-cascade refrigeration systems. Thermodynamic simulations are first conducted to evaluate the energy and exergy performance of five low-GWP refrigerant mixtures under variable operating conditions. A high-fidelity feedforward neural network (FNN) surrogate model is then established to accurately capture the nonlinear relationships between composition, operating parameters, and key performance indicators, including the coefficient of performance (COP), refrigeration capacity, exergy destruction, and exergy efficiency. Subsequently, the constraints separation based evolutionary multitasking (CSEMT) algorithm is employed to achieve parallel global optimization of refrigerant mixtures and operating variables. The entropy-weighted TOPSIS method is finally used to identify the best compromise solution from the Pareto-optimal set. Results reveal that the ternary mixture R13I1/R14/R600a provides the most balanced performance, achieving a COP of 0.5321. After FNN-CSEMT optimization, the COP increases to 0.743 and the refrigeration capacity reaches 101.2 W, while the exergy destruction is maintained at 86.25 W, substantially lower than that of R14/R1150/R600a. Correlation analysis reveals a near-perfect correlation between COP and exergy efficiency, and a strong positive correlation between pressure ratio and exergy destruction (r = 0.86–0.90). The framework not only quantifies critical trade-offs among refrigerant candidates but establishes a robust, efficient methodology for the multi-objective optimization of auto-cascade refrigeration systems.