Residential Building Energy Usage Prediction via Optimized Ensemble Learning

With rapid increases in energy consumption in residential buildings, predicting its usage energy is essential task. Therefore, engineers and designers try to find a reliable tool to help them analyze and predict the energy consumption of buildings. The state-of-the-art eXtreme Gradient Boosting (XGB) algorithm was successfully used to estimate the Heating Load (HL) and Cooling Load (CL) energy usage based on building design characteristics. To boost its performance and develop more reliable tools, a Bayesian technique called Tree-Structured Parzen Estimator was used to optimize the XGBoost Model. The model Named as TPE-XGBoost. Also, a new standard testing method was proposed to have a fair evaluation. The evaluation results based on the proposed standard testing method showed that the proposed model had outperformed other state-of-the-art models and improved the prediction accuracy of HL and CL by 5.5% and 9.8%, respectively.

The available dataset contains only 768 records of different buildings which is small dataset. To overcome low data problems and improving TPE-XGBoost performance, a new Tabular framework based on Generative Adversarial Networks (TG) was proposed for data augmentation which has generated 3863 new samples. The TG-TPE-XGB result showed that the proposed TG-TPE-XGB model outperformed basic XGBoost and proposed TPE-XGBoost. The proposed TG-TPE-XGB improved HL and CL estimation by 9% and 16% and improved TPE-XGBoost model performance by 5.9% and 6.6% in HL and CL. Therefore, by using data augmentation techniques, it is possible to overcome the problem of low data availability. The result indicated that TPE and data augmentation module can enhance model exploration and improve its estimation performance. Therefore, the developed models could be used by engineers and designers at an early stage of residential building construction to predict its energy usage.