<![CDATA[The HVAC (Heating, Ventilation, and Air Conditioning) system plays a crucial role in maintaining thermal comfort and energy efficiency in commercial and industrial buildings. However, early detection of anomalies or failures in this system is often suboptimal, leading to increased energy consumption, reduced operational performance, and high maintenance costs. This study aims to develop and evaluate various machine learning models for real-time anomaly detection in HVAC systems, using a real-world dataset that includes 11 key operational variables. Several algorithms are used, including Logistic Regression, K-Nearest Neighbors (KNN), Support Vector Classifier (SVC), Decision Tree, Random Forest, AdaBoost, Gradient Boosting, and XGBoost. The dataset is labeled based on dynamic deviations between actual temperature and setpoint using the Exponential Moving Average (EMA) approach, which allows for adaptive anomaly labeling. The experimental results show that the XGBClassifier achieves an accuracy of 99.32%, with precision and recall of 0.98 each, and an F1-score of 0.98, making it the best model for detecting anomalies in a balanced manner. Logistic Regression (accuracy 99.54%, F1-score 0.99) and Random Forest (accuracy 98.70%, F1-score 0.96) also proved to be reliable and computationally efficient alternatives. Thus, this research not only provides a comprehensive comparison of models but also emphasizes the novelty of the adaptive labeling strategy to support real-time anomaly detection in HVAC systems, which can enhance energy efficiency while reducing maintenance costs.]]>
