ABSTRACT
Predicting cooling loads in hot and humid climates using machine-learning approaches
Bingyan Jia
In residential, buildings cooling accounts for a significant amount of energy in hot, humid areas. In Qatar, cooling accounts for more than 60% of the country’s generated electricity. In previous years, overestimating cooling loads has led to wasted equipment costs and negatively impacted indoor thermal comfort. District cooling systems have gained popularity recently, both in Qatar and worldwide, due to advances in chillers, heat exchangers, and control systems. To determine the district cooling plant’s size and operation, it is vital to estimate building-level cooling loads. An accurate and fast prediction of building-level cooling loads is required to assist decision-making. There are several challenges in modeling building cooling loads, such as lack of detailed information regarding the buildings (e.g., building envelope), cooling energy data for validation, and computational effort. Furthermore, in multi-apartment residential buildings connected to a district or centralized cooling system, individual charging requires improvement. To save money and time on individual metering and charging, apartment-level cooling loads prediction is also required. Moreover, an apartment-level meta-model may be useful for optimizing building energy use and retrofit analysis at the apartment level.
This study aims to develop building- and apartment-level meta-models using machine learning approaches to predict the cooling load. Four machine-learning approaches are applied: multiple linear regression, support vector regression, artificial neural networks, and extreme gradient boosting. Critical parameters identified using sensitivity analysis are used as independent variables, which simultaneously consider the building envelope, climate, and internal heat gain parameters. New building energy models are created to test the meta-models’ performance.