Selecting the Best Design for Residential Buildings from an Energy Usage Perspective with Advanced Machine Learning Techniques

Accurate prediction of heating load (HL) and cooling load (CL) in residential buildings is crucial for energy optimization and cost savings. This study combines the strengths of two innovative approaches to address the challenges posed by small datasets and complex dependencies in energy load prediction. The first approach combined the state-of-the-art XGBoost model with Generative Adversarial Network (GAN)-based data augmentation techniques, and its parameter tuned with grid search algorithm. The result indicated that proposed model outperforms other models.

The second approach introduces the DAPM-LLM model, which integrates Large Language Models (LLMs) with a prompt generation module that used new synthetic dataset generated using GAN algorithm. Tabular building data is transformed into linguistic prompts using a prompt generation module, and the pre-trained BART-base model is fine-tuned to predict HL and CL. The inclusion of a GAN-based data augmentation module significantly improves LLM model performance, increasing HL prediction accuracy by 600% and CL prediction accuracy by 300%. Comparison with other methods reveals that DAPM-LLM surpasses most existing models and holds promise for achieving further improvements with larger pre-trained models and more extensive datasets.

The third approach focuses on TabNet, an interpretable deep learning framework for tabular data, combined with a hybrid tabular data augmentation module based on GAN and Conditional GAN (CGAN) Named TDAM module. The augmentation module expands the training dataset five times, addressing the issue of limited data and enabling TabNet to leverage its advanced feature selection and interpretability capabilities. Results show that, the TabNet-TDAM model achieved superior performance, producing results comparable to classical machine learning methods for HL and CL prediction.

All proposed models effectively address the challenges of limited data and complex dependencies in energy prediction tasks. They offer practical tools for engineers and designers to select optimal building plans and designs from an energy efficiency perspective. The findings underscore the potential of combining advanced deep learning architectures with robust data augmentation techniques to improve the accuracy, interpretability, and scalability of energy consumption prediction models.