Utilizing Artificial Intelligence for Efficient Defrosting Operation of Fin-tube Heat Exchanger

다국어 초록 (Multilingual Abstract)

The heat exchanger is a crucial component in various industries and environmental fields, and its importance continues to grow. However, frost formation occurs in the heat exchangers of air source heat pumps exposed to cold storage or outdoor air during winter. This reduces the heat transfer capacity of heat exchangers, leading to decreased thermal performance and compromised system stability. Therefore, defrost operations are necessary to remove frost. Defrost operations involve temporarily suspending regular operation and using heat to melt frost, which interrupts normal functionality and consumes a significant amount of energy. Further, defrost methods commonly used in practical settings, such as periodic defrosting, are inefficient and may lead to improper defrosting phenomena. In this study, the goal is to develop an algorithm that overcomes the limitations of periodic defrosting by utilizing temperature information from the heat exchanger, specifically the temperature of the fluid passing through the fin spacing, to detect frost formation and determine the defrosting timing. To gather training data for the algorithm, a wind tunnel experiment device is constructed to simulate frost formation on the heat exchanger. Eighteen environmental conditions are derived through the experimental design, considering actual frost formation conditions. Experiments are conducted under these environmental conditions, and data on the temperature of the heat exchanger fin spacing and heat transfer rate during the frost formation process are obtained. The collected data have a time-series nature, and preprocessing is performed to use these data as suitable training data for artificial intelligence algorithms. Therefore, data preprocessing techniques such as maximum-minimum normalization using the maximum and minimum values of the data, preprocessing of time-series data using sliding window techniques, and classification of the collected data into training and validation datasets for artificial intelligence training are conducted. To realize the construction of the algorithm structure, discussions are made regarding recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) models. To regress the heat transfer rate from the characteristics of the time- series data, an FC-Layer is used. Further, dropout, batch normalization, and bidirectional layers are added to improve the performance of the artificial intelligence model. A comparison of the performances of the RNN, LSTM, and GRU models is conducted to derive the appropriate algorithm for achieving the research objectives, and the LSTM model is found to be suitable. Additionally, it is deduced that the appropriate time information for identifying temperature change patterns owing to frost formation in the heat exchanger is approximately 30 min. After deriving the structure of the algorithm that aligns with the research objectives, the performance of the algorithm is evaluated using data acquired from environments not used as the training data. The artificial intelligence algorithm developed in this study takes the temperature of the heat exchanger fin spacing temperature as the input and regresses it to the heat transfer rate. It then determines the defrosting timing when the inferred heat transfer rate value decreases below half of the maximum heat transfer rate of the heat exchanger. The performance evaluation of the algorithm shows that it accurately derives the optimal defrosting timing even in environments with variations in the refrigerant temperature, airflow velocity, and humidity. With a maximum error of approximately 11% in deriving the optimal defrosting timing, it is confirmed that the developed algorithm is effective in detecting frost formation and determining the defrosting timing.

목차 (Table of Contents)

• Table of contents i
• Nomenclature iii
• List of Tables v
• List of Figures vi
• Abstract viii
• 1. Introduction 1
• 1.1. Background 1
• 1.2. Motivations 6
• 1.2.1. Side effects of defrost operation 6
• 1.2.2. Previous research to detect frost formation 10
• 1.3. Objective of research 12
• 2. Experimental section 14
• 2.1. Experimental setup and data acquisition environment 14
• 2.1.1. Experimental setup 14
• 2.1.2. Experimental procedures 24
• 2.1.3. Experimental conditions 25
• 2.2. Data processing for artificial intelligence learning 28
• 2.2.1. Feature selection: Temperature & time 28
• 2.2.2. Label selection: Heat transfer rate 28
• 2.2.3. Sliding window method 29
• 2.2.4. Data normalization and splitting 32
• 2.3. Artificial intelligence architecture 34
• 2.3.1. Deep neural network models 34
• 2.3.2. Activation function 41
• 2.3.3. Additional layer 42
• 2.3.4. Model architecture for estimating heat transfer rate and detecting frost formation 44
• 2.3.5. Model training and hyperparameter configuration 48
• 3. Results and discussion 50
• 3.1. Analyzing model performance based on algorithm type 50
• 3.2. Analyzing model performance based on input width (time-step) size 54
• 3.3. Performance of AI frost detection algorithm 57
• 4. Conclusion and future studies 64
• 4.1. Conclusion 64
• 4.2. Future studies 65
• References 67
• 이력 및 연구 실적 71
• 국문 요지 73
• 감사의 글 75