Machine learning-based performance and optimal refrigerant charge prediction for a split air conditioning system

Gomathi, B (2025) Machine learning-based performance and optimal refrigerant charge prediction for a split air conditioning system. Applied Thermal Engineering, 274: 126764. ISSN 13594311

[thumbnail of Machine learning-based performance and optimal refrigerant charge prediction for a split air conditioning system.pdf] Text
Machine learning-based performance and optimal refrigerant charge prediction for a split air conditioning system.pdf - Published Version

Download (15MB)

Abstract

Maintaining the optimal refrigerant charge and accurately predicting the performance of split-type air conditioning (STAC) systems are essential for enhancing energy efficiency and reducing carbon emissions. The adoption of the environmentally friendly refrigerant R290 as a replacement for R22 has increased due to updated regulations. However, optimizing refrigerant charge and system performance typically demands substantial research, cost, and time. To address this challenge, six machine learning (ML) models—artificial neural network (ANN), multi-layer perceptron (MLP), extreme gradient boosting (XGB), support vector regression (SVR), random forest (RF), and K-nearest neighbor (KNN)—were employed to predict STAC performance and determine the optimal refrigerant charge. For model development, the considered input features were indoor and outdoor temperatures—specifically, wet-bulb temperature (WBT), dry-bulb temperature (DBT)—and refrigerant charge. The predicted outputs were system pressure, power consumption, coefficient of performance (COP), refrigerant mass flow rate, and evaporator capacity. Correlation analysis showed a strong negative correlation between outdoor DBT and COP (−0.96), followed by indoor DBT (−0.81) and indoor WBT (−0.67). However, no perfect relationship was observed between COP and refrigerant charge, likely due to nonlinearity. Among the models, SVR, ANN, and RF performed best on smaller datasets (<20 samples), while XGB demonstrated superior performance on larger datasets (>40 samples). XGB outperformed MLP and RF, achieving deviations within ±5 %, a mean absolute error of 2.539, and an R2 value of 0.957. The optimal refrigerant charge amounts predicted for R290 were 460 and 320 g for the existing and modified systems, respectively, 2.3 % and 3.2 % higher than the experimental values. This study underscores the potential of ML models, particularly XGB, in optimizing STAC systems using environmentally friendly refrigerants.

Item Type: Article
Subjects: C Computer Science and Engineering > Machine Learning
F Mechanical Engineering > Refrigeration
Divisions: Computer Science and Engineering
Depositing User: Dr Krishnamurthy V
Date Deposited: 24 May 2025 03:34
Last Modified: 24 May 2025 03:34
URI: https://ir.psgitech.ac.in/id/eprint/1436

Actions (login required)

View Item
View Item