Analysis of energy efficiency of commercial refrigerator systems with machine learning and its contribution to sustainability goals
Journal of Thermal Analysis and Calorimetry, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s10973-026-16496-y
- Dergi Adı: Journal of Thermal Analysis and Calorimetry
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Chemical Abstracts Core, Chimica, Compendex, Index Islamicus, INSPEC, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
- Anahtar Kelimeler: Commercial refrigeration, Energy efficiency, Machine learning, RDC, Sustainability
- Gazi Üniversitesi Adresli: Evet
Özet
The energy efficiency of commercial refrigeration systems plays an important role in achieving sustainability goals. However, comprehensive studies using machine learning techniques in this field are limited. Commercial refrigeration systems are influenced by interactions among multiple variables, including ambient temperature, relative humidity, shelf arrangement, and product packaging. Repeated experiments in standard climate classes are required to evaluate the performance, but these processes are costly and time-consuming. This study aims to develop an experimental procedure to evaluate the performance of commercial refrigeration equipment in a specific climate class and to make accurate predictions under other M-package temperature conditions. A closed, semi-vertical, multi-deck display refrigerated cabinet (CSVRDC) was designed, constructed, and tested as an experiment in accordance with ISO 23953-2:2023. 300 g of natural R290 was used as the refrigerant for the twin-circuit cooling system. The tests were carried out under the M0, M1, and M2 M-package conditions, while the test-room climate remained Class 3, using 500 g test packs equipped with temperature sensors. Test data were recorded and analysed in real time. A comparison of six regression algorithms was conducted, with inputs ranging from raw data to z-score-standardised versions. To limit temporal leakage, scaling was fitted exclusively within each training fold, and performance was estimated using tenfold contiguous, compressor-cycle-blocked cross-validation with adjacent cycles purged. The CatBoost model, with raw inputs, was selected exclusively from the M0 development results and then held fixed for both external evaluations. The M0 cross-validation performance was characterised by RMSE = 21.94 ± 12.26 W, MAE = 5.02 ± 2.91 W, and R2= 0.9870 ± 0.0115 (mean ± SD across folds). Gap-aware cycle-level bootstrap analysis yielded R2= 0.9519 (95% CI 0.9436–0.9593) for M1 and 0.9015 (0.8936–0.9093) for M2. It was evident that larger errors were concentrated in threshold-defined state-change samples and the low-power region near 100 W, with a stronger covariate shift in M2. The findings indicate that cross-condition screening of electrical-power behaviour for this cabinet may support preliminary evaluation; however, they do not establish that standard qualification tests or independent product validation can be replaced.