Artificial and convolutional neural networks applications to predict heating and cooling loads for residential buildings

Mirindi, Derrick; Bayongwa, Ahana; Adedoyin, Abiodun; Kandel, Manisha; Sinkhonde, David; Cecilia Osuagwu, Eberechi; Bezabih, Tajebe; Mirindi, Frederic

doi:10.1016/j.ssaho.2026.102647

Journal Article

FZJ-2026-02107

Artificial and convolutional neural networks applications to predict heating and cooling loads for residential buildings

Mirindi, D. (Corresponding author) ; Kandel, M. ; Adedoyin, A. ; Sinkhonde, D. ; Bezabih, T. ; Cecilia Osuagwu, E. ; Mirindi, F. ; Bayongwa, A.FZJ*

2026
Elsevier Amsterdam

Social sciences & humanities open 13, 102647 (2026) [10.1016/j.ssaho.2026.102647]

This record in other databases:

Please use a persistent id in citations: doi:10.1016/j.ssaho.2026.102647 doi:10.34734/FZJ-2026-02107

Abstract: In the face of the rapid evolution of artificial intelligence (AI) and the increasing cost of energy in residential buildings, accurately predicting thermal loads has become crucial for sustainable construction practices. We present the use of artificial neural network (ANN) and convolutional neural network (CNN) models to predict the energy efficiency of residential buildings. The dataset comprises eight input parameters, namely surface area, relative compactness, wall area, overall height, roof area, glazing area, orientation, and glazing area distribution, with two output thermal loads (the heating load (HL) and the cooling load (CL)). Additionally, we split the data, comprising 768 observations, into training (70%), testing (15%), and validation sets (15%). Results based on the Pearson correlation matrix indicated that all input variables exhibit a positive correlation with the thermal loads, except the surface and roof areas of the building. In addition, the feature importance and Shapley Additive exPlanation (SHAP) analysis demonstrated that building geometry parameters, such as relative compactness, wall, surface, and glazing areas, dominate thermal load predictions. Furthermore, the ANN models showed high performance, with R2 values ranging from 0.9618 to 0.9783 for HL and CL. However, the CNN models significantly outperformed ANN models. When comparing training, testing, and validation, CNN models achieve exceptional R2 values exceeding 0.99 for all dataset splits, even in the presence of outliers. K-fold cross-validation analysis demonstrated the outstanding reliability of the CNN models, with coefficient of variation (CV) values of 0.26% for HL and 0.65% for CL, suitable for engineering applications and real-world deployment. However, the ablation study results identified the non-regularized CNN configuration as optimal for production deployment, having low gap metric values between training and validation HL (−0.0001) and CL (0.0040) models. Beyond technical achievement, this research demonstrates that building energy prediction serves as a tool for advancing household energy consumption. Community engagement and five ethical considerations are proposed for citizen science programs and scientific education initiatives focused on sustainable energy consumption.

Classification: