Stacking Ensemble Deep Neural Networks with Regressor Chains for Building Energy Performance Prediction

Abstract

The energy performance of buildings (EPB) is a critical factor in reducing global energy consumption, mitigating greenhouse gas emissions, and achieving sustainability goals. Predictive modelling of EPB constitutes a complex, non-linear multi-target learning problem, where multiple continuous outputs must be estimated simultaneously from a common set of input variables. Multi-Target Regression (MTR) presents significant challenges due to complex output dependencies, high output dimensionality, imbalanced and noisy targets, and distributional shifts, which collectively degrade predictive performance. To address these challenges, this study proposes a novel ensemble regressor-chain framework integrated with a stacking ensemble deep neural network architecture for MTR modelling. The proposed approach is evaluated using five benchmark multi-target regression datasets related to building energy performance. Experimental results demonstrate that the proposed model consistently outperforms classical regression methods (linear regression, generalized linear models, and CART) as well as recent state-of-the-art approaches, including regression forests and sparse regression techniques. Performance gains of up to 12% reduction in RMSE and a 9% improvement in R² are achieved. Robustness is further validated through statistical testing using the Friedman test with Finner’s post-hoc correction, supported by visual analyses such as scatter plots and error distributions. Overall, the results indicate that ensemble deep learning architectures combined with regressor chains provide a more effective and scalable solution for multi-target EPB prediction than traditional regression models, offering practical value for real-world energy efficiency assessment and sustainability-oriented decision making.