Machine Learning Enhances Green Building Efficiency
A recent article published in Scientific Reports proposed integrating machine learning (ML) and deep learning (DL) techniques to design green buildings (GBs) to enhance their energy efficiency and environmental sustainability.
Background
The building and construction industry uses excessive natural resources, harming the environment. Moreover, it consumes the maximum energy (36%) and emits the maximum CO2 (37%) globally. Accordingly, the GB concept is endorsed as a method to enhance the building industry’s sustainability throughout a building’s lifecycle.
However, GB design is a longer process than conventional building design due to several green requirements and performances requiring integration into the building design. Digitization, automation, and integration enabled by technological advancements in construction can improve decision-making and productivity in GB initiatives.
Artificial intelligence methods such as ML and DL assist designers in completing their work more quickly and precisely. ML enables a system to grow and learn from its own experiences without programming. Thus, this study aimed to create a design prediction model for GBs using ML and DL methods.
Methods
The predictive model creation process involved multiple steps, such as obtaining datasets, preparing data, making model predictions, and analyzing the predicted results.
The model design was based on the ASHRAE dataset acquired through field investigations of 160 global building sites. This dataset was meant to create a model of preferred thermal convenience. Overall, it forms a component of the RP-884 accessible repository.
A total of 56 features and 12595 records in the ASHRAE RP-884 dataset were used in this research to create an adaptive classifier. This comprised over 20000 consumer convenience scores from 52 polls conducted across 10 climatic zones.
Data preprocessing was performed using exploratory data analysis (EDA). It helped clean, sort, and convert the category data into numerical values through label encoding. Specifically, a Z-score normalization was used for data cleaning.
Various ML and DL models were explored for GB’s sustainable environment. The former types were Random Forest (RF), Decision trees (DT), Extreme Gradient Boosting (XGB), and Stacking, while the latter types were Graph Neural Networks (GNN), long-short-term memory (LSTM), and recurrent neural network (RNN).
All ML and DL classifiers were trained using the ASHRAE dataset and made to forecast energy use in GBs. Notably, different models are combined for increased accuracy. Their performance was measured using the F1-score, recall, accuracy, and precision metrics. F1-score identified the most balanced performance as it combined memory (or recall) and precision metrics as the harmonic mean.
Results and Discussion
The ML models (XGB, RF, and Stacking) based on the ASHRAE-884 dataset exhibited a good accuracy of 0.84. However, the model had an accuracy of 0.76. In addition, the XGB model had the highest precision (0.83), while the RF model had the lowest (0.82).
Furthermore, the RF and XGB models demonstrated the maximum recall (0.84), while the DT model had the poorest recall (0.76). The XGB, RF, and Stacking models also had the maximum F1 Score of 0.80, while the DT had an F1 Score of 0.77.
Similar to the ML models, the DL models employed various GB energy consumption parameters from the ASHRAE dataset to forecast a building’s energy utilization. The evaluation of these models revealed whether they were trained on the test or training dataset. Notably, the training dataset yielded better results from all three models than the test dataset.
The GNN model performed exceptionally well, with an accuracy of 0.83 on the test and 0.85 on the training datasets. Alternatively, the corresponding accuracy for the LSTM and RNN models was approximately 0.79 and 0.81, respectively. Thus, the GNN model could effectively capture the relationships between various features in the data.
The accuracy was higher than the precision and recall in all DL models. Therefore, false positive predictions were more common than false negatives in these models. Additionally, the F1-score of the three models declined from the test to the training dataset.
Conclusion and Future Prospects
Overall, the researchers successfully demonstrated various ML and DL-based predictive models for GB design. These models can maximize resource usage, improve occupant comfort, and reduce the building industry’s environmental impact throughout the GB design process.
The performance of the proposed models was exhibited using the ASHRAE-884 dataset. DL models such as GNN and LSTM performed more accurately and efficiently than conventional DL techniques for environmental sustainability in green buildings.
However, this research has limitations regarding the dataset as the resolution of the climate statistics was not precise enough to capture localized temperature fluctuations, which are critical for accurately predicting energy usage.
Thus, the researchers suggest developing more robust and accurate ML and DL models validated with climatic and historical energy consumption data. Moreover, including longer periods and better temporal and spatial resolution of climate information in the ASHRAE-884 dataset will enhance the models’ ability to capture long-term climatic trends as well as localized temperature variations.
Journal Reference
Mahmood, S., Sun, H., El-kenawy, El-S. M., Iqbal, A., Alharbi, A. H., & Khafaga, D. S. (2024). Integrating machine and deep learning technologies in green buildings for enhanced energy efficiency and environmental sustainability. Scientific Reports, 14(1). DOI: 10.1038/s41598-024-70519-y.
link