Explaining building energy efficiency prediction through architectural and engineering solutions considering environmental impacts using a hybrid model

Abstract

The aim of this study is to investigate building energy efficiency by integrating hybrid modelling approaches and interpretable models into architectural design processes and engineering solutions while considering environmental impacts. At the same time, it is aimed to achieve the most efficient building design possible. Separate analyses were carried out using a total of 6,913 data from Rutland (185) and Salford (6718) cities. The data were grouped into carbon emission information, architectural information, lighting information, personal heating information, and main heating system information to form a dataset. Light Gradient Boosting Machine (LightGBM) was preferred as the base model and Particle Swarm Optimisation (PSO) method was applied for hyperparameter optimisation. The hybrid model created in this way is called PSO-LightGBM. The optimization process was carried out using software in both R Studio and Python environments, utilizing seven different hyperparameters. Apart from the hybrid model used as a method, 2 different SHAP analyses, neural network based, and tree based, were performed to clearly explain the parameter relationships. The PSO-LightGBM hybrid model provided more successful predictions compared to the basic LightGBM model. While R² values improved between 0.82 and 0.90 in the Rutland dataset, this value increased from 0.8687 to 0.8901 for test data and from 0.8538 to 0.9091 for training data in the Salford dataset. R² values show an improvement of 7% in the Rutland dataset and maximum 6% in the Salford dataset. When the reduction in error rates is evaluated, it is found that the greatest improvement is in the Mean Squared Error (MSE) metric. MSE decreased by 17% in the Rutland dataset and by 4% in the Salford dataset. According to the SHAP Analysis results, CO₂ emissions have the largest impact on energy consumption, while primary fuel types, number heated rooms and individual heating systems are other important parameters. While the tree-based SHAP model is more sensitive to physical parameters, the neural network model is more sensitive to indirect relationships. In both analyses, the communal heating system type has the lowest impact. In order to improve building energy efficiency, high efficiency individual boiler systems should be preferred, architectural approaches that optimise the number of heated rooms and smart heating solutions should be used, and central heating systems should be modernised. The results highlight the effectiveness of hybrid modelling approaches with SHAP analyses based on different baselines to ensure the integration of environmental impacts, architectural design processes and engineering solutions in terms of building energy efficiency. Furthermore, the findings contribute to the importance of interdisciplinary work in buildings to improve energy efficiency. Future studies can focus on the development of building energy performance prediction models that can be explained by integrating various learning algorithms and different optimisation techniques, using regions with high cooling demand, large datasets and different building types.