Deep learning-based prediction method on performance change of air source heat pump system under frosting conditions

Since frost on an outdoor heat exchanger in winter reduces the performance of an air source heat pump (ASHP), a defrosting process is necessary to restore the degraded performance. Therefore, frosting and defrosting are crucial challenges. For a more efficient defrosting process, many researchers have conducted studies on demand-based defrosting control so far. Recently, various researches on frost growth prediction using neural networks have been conducted. Here, we propose a novel method to quantitatively predict changes in the performance (heating capacity, power consumption, and COP) of ASHPs due to frost growth using a single model based on deep learning. Based on prediction results, this method can be utilized to optimize the defrosting start control strategy. With multiple outputs regression models, we can predict three performance parameters simultaneously. They are models trained with only the initially installed sensors without additional sensors. Besides, we compared the prediction accuracy differences depending on three deep learning structures, such as a fully-connected deep neural network (FCDNN), convolutional neural network (CNN), and long short-term memory (LSTM). Summarizing the results, the optimal FCDNN-based model achieved a root-mean-square (RMS) error of 2.8% for the prediction of heating capacity, 2.4% for power consumption, and 3.4% for COP of ASHPs. (c) 2021 Published by Elsevier Ltd.

Automated summary

Researchers have proposed a deep learning-based model to quantitatively predict performance changes of ASHPs, aiming to optimize defrosting start control strategy to address challenges of frosting and defrosting. Training with multiple outputs regression models using only initially installed sensors, the model achieved accurate predictions for heating capacity, power consumption, and COP of ASHPs.