Space cooling energy usage prediction based on utility data for residential buildings using machine learning methods. (1st June 2021)

• Record Type: Journal Article
• Authors: Feng, Yanxiao; Duan, Qiuhua; Chen, Xi; Yakkali, Sai Santosh; Wang, Julian
• Is Part Of: Applied energy. Volume 291(2021)
• Publication Date: 2021-06-01
• Subjects: Cooling energy -- Energy prediction -- Utility data -- Machine learning method -- XGBoost model -- Parameter tuning -- Building characteristics -- Weather features; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2021.116814
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: A user-friendly, infrastructure-free, and accurate model based on XGBoost. Hyperparameters tuning is applied to exploit optimal modeling performance. Interpreting cooling energy predictions and feature interactions by using SHAP. Surface to volume ratio estimated is beneficial for cooling energy use estimation. Degree hour data outperforms temperature for cooling energy estimation in XGBoost. Abstract: The energy used for space cooling in residential buildings has a significant influence on household energy performance. This study aims to develop a user-friendly, infrastructure-free, and accurate prediction model based on large-scale utility datasets from anonymized volunteer homes located in three different climate zones in the US, along with the corresponding weather data and building information. Notably, several new weather- and building characteristics-related parameters were designed in the modeling procedure and tested to be useful for enhancing the model's prediction performance. A few regression techniques were examined and compared through hyperparameter optimization and k-fold cross-validation. Subsequently, a workflow was also described for how to implement the developed model. The research results showed that the eXtreme Gradient Boosting (XGBoost) model offered optimal performance, and the feature importance analysis also identified as well as ranked the key predictors to enhance the interpretability of this model. An R 2 value of around 97% was obtained with that model on the whole dataset, while an R 2 value of 92% was achieved with various subsets of the dataset through the cross-validation approach. The RMSE and RAE for this model were 0.294 and 0.153, respectively. The resultant model for predicting cooling energy consumption will facilitate homeowners better understanding their buildings' performance levels with minimum input information and without additional hardware installations, ultimately aiding their decision making related to energy-saving strategies.