The prediction of chiller performance parameters is crucial for optimal control and fault diagnosis. Numerous efficient and accurate data-driven models have been developed and implemented. These models are normally trained on historical operational data of chiller units. However, the distribution of operational data may shift due to accumulated operating hours or changes in control strategies. Under new operating conditions, models trained on historical data often generalize poorly, leading to prediction deviations. To address this issue, this study integrates a one-dimensional convolutional neural network with a domain adaptation method that extracts features from both the source and target domains and aligns their inverse Gram matrices in terms of angle and scale. A predictive model applicable to multiple chiller performance parameters is established using limited historical data, enhancing the model’s generalization ability. Compared to the baseline model (MLP), the proposed method achieves an average reduction of 74.3% in mean absolute error (MAE) and 76.1% in root mean square error (RMSE), while the R2 values exceed 0.96 (for certain scenarios). Additionally, this paper analyzes the data distribution between the source and target domains, investigates key factors affecting the model’s generalization capability, and provides insights for evaluating the quality of modeling data.
Liu et al. (Fri,) studied this question.