Unveiling the impact of base model selection in heterogeneous ensemble learning for building energy prediction

Heterogeneous ensemble learning has shown promise for building energy prediction even under complex occupancy scenarios. The base model, acting as the foundational element of the heterogeneous ensemble model, significantly influences the overall predictive performance. However, previous studies have predominantly focused on validating its advantages without adequately addressing the process of base model selection, thus failing to maximize the advantages of different algorithms. This study develops a principled framework for selecting and combining base models to optimize ensemble performance. Through systematic evaluation of 63 possible combinations derived from six diverse base models, we demonstrate that integrating 3–4 base models achieves optimal performance, yielding MAPE reduction of 3.05 %–7.97 % compared to individual best-performing base models. Our analysis reveals that the heterogeneous ensemble architecture effectively amplifies the distinct characteristics of constituent algorithms, with predictive accuracy being significantly influenced by individual model performance, prediction stability, and the complementarity among base models. These findings contribute both methodological and practical insights for optimizing heterogeneous ensemble construction in building energy prediction, offering researchers and practitioners a systematic approach to enhance prediction accuracy in energy management applications.