Deep learning-based efficiency forecasting of automotive air conditioning system with variable-capacity compressors under different conditions

Abstract

This study uses experimental data to demonstrate the accuracy of deep learning (DL) modeling for automotive air conditioning (AAC) systems. The experimental AAC system uses a variable capacity compressor and R1234yf as the refrigerant. The experimental system is equipped with different control and data acquisition systems to obtain the best data. The experimental data is obtained at different compressor speeds and condenser and evaporator inlet air flow rates, temperatures, and relative humidity values. Data obtained according to various conditions were collected according to the system’s steady state. The AAC system with a variable capacity compressor and R1234yf was tested 108 times for DL application. The obtained data were evaluated according to the compressor discharge temperature, evaporator outlet airflow temperature, refrigerant mass flow rate, compressor power, cooling capacity, and coefficient of performance. To construct a dependable DL model for the AAC system utilizing R1234yf, the dataset was partitioned into training (72.22%) and testing (27.78%) subsets. Linear Interpolation-based Data Augmentation was employed to address data shortage and enhance generalization, augmenting the training dataset to 1008 patterns. The optimal model performance was achieved with HLNN set to 35 and NHL set to 4. Exceptional prediction accuracy was attained in all instances, with R² values surpassing 0.9998 and minimal error metrics (e.g. MSE = 0.000027 for refrigerant mass flow rate, MSE = 0.00113 for coefficient of performance). The results indicate that the DL model, improved by data augmentation, yields precise, generalizable predictions, and diminishes the necessity for substantial physical experimentation in AAC system analysis.