Developing an integrated prediction model for daylighting, thermal comfort, and energy consumption in residential buildings based on the stacking ensemble learning algorithm

Abstract

Accurate and rapid predictions of residential building performance are crucial for both new building designs and existing building renovations. This study develops an integrated prediction model using a stacking ensemble learning algorithm to predict daylighting, thermal comfort, and energy consumption in residential buildings. The model incorporates multimodal residential building information as inputs, including image-based floorplans and vector-based building parameters. A comparative analysis is presented to evaluate the prediction performance of the proposed stacking ensemble learning algorithm against three base models: Resnet-50, Inception-V4, and Vision Transformer (ViT-32). The results indicated that the stacking ensemble learning algorithm outperforms the base models, reducing the mean absolute percentage error (MAPE) by 0.17%–1.94% and the coefficient of variation root mean square error (CV-RMSE) by 0.37%–2.06% for daylighting metrics; the MAPE by 0.63%–4.46% and the CV-RMSE by 0.62%–5.13% for thermal comfort metrics; the MAPE by 1.42%–6.43% and the CV-RMSE by 0.27%–5.09% for energy consumption metrics of the testing dataset. Further prediction error analyses also indicate that the stacking ensemble learning algorithm consistently yields smaller prediction errors across all performance metrics compared to the three base models. In addition, this study compares the stacking ensemble learning algorithm to traditional machine learning models in terms of prediction accuracy, robustness, and generalization ability, highlighting the advantages of the stacking ensemble learning algorithm with image-based inputs. The proposed stacking ensemble learning algorithm demonstrates superior accuracy, stability, and generalizability, offering valuable and practical design support for building design and renovation processes.