Understanding and improving building energy prediction via data characteristic analysis and adaptive training data reconstruction

Accurate and reliable prediction of building energy consumption remains challenging due to the variability of operational conditions and the limited interpretability of data-driven models. Existing approaches often focus on algorithm optimization while overlooking the role of training data characteristics and their impact on model performance. This study proposes a data characteristic-aware modeling approach that integrates multi-dimensional data analysis with similarity-based training sample selection to enhance both prediction accuracy and interpretability. The method systematically characterizes energy data using pattern recognition, dispersion metrics, and periodicity analysis, and reconstructs task-specific training datasets through similarity evaluation. An adaptive modeling process is further implemented to automatically select features and algorithms while providing post-hoc interpretability using SHAP analysis. The proposed approach is validated using datasets from four types of public buildings under different operational conditions. Results show that the method significantly improves prediction performance, particularly during transitional periods, with R 2 increasing by up to 22% compared with conventional approaches. The analysis also reveals that reduced data dispersion and improved similarity between training and target conditions are key factors driving performance gains. Overall, this work highlights the importance of training data characteristics in building energy prediction and provides a practical solution for developing accurate and interpretable models in real-world energy management applications.