A data informativeness evaluation method for grey-box modeling of building thermal dynamics

Grey-box modeling has been widely used in building thermal modeling due to its adaptability and interpretability. The identification of model parameters mainly depends on the measured dataset, and its optimal construction is critical for ensuring model accuracy. Existing studies commonly discuss the influence of training data quantity on the model accuracy. However, the training data informativeness is always ignored, which reflects the quality and richness of information within the data samples and informs the estimates of model parameter values. Notably, the informativeness level may vary among samples, and the quantity of data does not necessarily correlate with its informativeness. Here, we propose a data informativeness evaluation method that can well select informative training data for grey-box models under different scenarios. The method establishes two evaluation criteria based on the characteristics of grey-box model: one describes the consistency between training and forecasting data distributions, and the other outlines the distribution variations within the training data. The effectiveness of the proposed method is demonstrated using data from experiment case. The results indicate that the proposed data informativeness index reflects the quality of the dataset well and has a high correlation with prediction accuracy (The Pearson correlation coefficient varies from −0.6 to −0.8). This evaluation method will be of great significance for optimizing the dataset construction of grey-box model of building thermal dynamics.