Digital twin-oriented decision support for central air conditioning systems: Integrated load prediction and fault diagnosis via hybrid machine learning

Central air conditioning systems represent a dominant source of energy consumption in commercial buildings, accounting for up to 40% of total building energy use. Accurate load forecasting and reliable fault diagnosis are therefore important for supporting efficient operation and maintenance of chiller systems. This study presents a digital twin-oriented decision support framework for central air conditioning chiller units that integrate data acquisition, predictive modeling, fault diagnosis, and visualization to support building operation and maintenance. To address the inherent nonlinearity and multivariable complexity of chiller load dynamics, a hybrid CNN-WOA-XGBoost model is proposed, wherein a Convolutional Neural Network (CNN) extracts temporal features from multivariate operational data, and the Whale Optimization Algorithm (WOA) adaptively optimizes XGBoost hyperparameters to enhance regression accuracy. Validated using operational data from a commercial office building and the large-scale Buildings Bench benchmark dataset, the proposed model achieved an R2 of 0.995, a Mean Absolute Percentage Error (MAPE) of 0.81%, and a Root Mean Square Error (RMSE) of 2.21, outperforming established baselines including SVR, standalone XGBoost, and KNN. Complementarily, an enhanced one-dimensional CNN integrated with Wavelet Packet Decomposition is developed for compressor-bearing fault diagnosis, attaining a 100% recognition rate under clean conditions and 96.4% macro-average accuracy under severe noise interference (−10 dB SNR). Embedded within a visualization and decision support platform, the proposed framework provides load prediction and fault diagnosis information for maintenance inspection and operation support. These findings indicate the potential of the proposed framework to support data driven operation and maintenance of central air conditioning systems.