Real-time performance prediction of air source heat pumps using deep learning-based model

Frost easily forms on the surface of the outdoor heat exchanger during winter [1].Frosting seriously affects the operating performance of an air source heat pump (ASHP), reducing its heating capacity and efficiency [2].Thus, defrosting is inevitable to prevent efficiency degradation in ASHPs.However, an inappropriate defrosting initiation may increase the energy consumption [3].To determine the optimal defrosting initiation, it is important to consider heating capacity, power consumption, and frost growth [4].Performance prediction based on these variables can be achieved using deep learningbased models.While deep learning-based performance prediction models have been studied [5], additional research is required to enhance their accuracy under varied operating conditions.Therefore, a real-time prediction model should be developed to update the varied data continuously.In this study, experiments were conducted under frosting conditions on the evaporator in a psychrometric chamber.During the operation, the heating capacity was calculated using the measured temperature and pressure from sensors in the indoor unit, and the power consumption was measured using a power meter in the outdoor unit.The blockage ratio was determined by processing images on the evaporator surface to measure the degree of frost accumulation.The deep learningbased prediction model received inputs from 15 variables, including outdoor conditions, refrigerant properties, heating capacity, power consumption, and blockage ratio.The prediction model was trained to learn the relationships among these variables.To predict the heating capacity and power consumption in real-time, the model was trained on 10 min of historical data to predict the next 5 min, and it was continuously updated every 2 min.Evaluation metrics used to assess the predicted values were mean absolute error (MAE) and root mean square error (RMSE).Prediction results in the period before the first defrosting showed an MAE of 0.25 and an RMSE of 0.46 for the heating capacity, and an MAE of 0.07 and an RMSE of 0.15 for the power consumption, at an outdoor dry-bulb temperature of 2 and a wet-bulb temperature of 1 .In the period after the second defrosting, the MAE and RMSE decreased by approximately 20% and 30%, respectively.Additionally, prediction results maintained high accuracy with changes in the outdoor conditions, demonstrating their adaptability.Therefore, the real-time prediction model can be used to determine the optimal defrosting point under various operating conditions.