Data-driven Whole Building Energy Forecasting Model for Data Predictive Control

Liang Zhang

doi:10.17918/3k51-2078

In the United States, the buildings sector accounted for about 41% of primary energy consumption. Building control and operation strategies have a great impact on building energy efficiency and the development of building-grid integration. Model predictive control (MPC) has received extensive attention from researchers in the field of whole building control and operation strategies. To develop MPC for whole building control and operation, high-fidelity building energy forecasting model is one of the most critical components. Data-driven energy forecasting model is typically developed using statistical methods to capture the relationship between building energy consumption and collected building data, such as operation data. MPC built with a data-driven model is also termed as data predictive control (DPC). Due to the surge of machine learning and the advances of building automation system (BAS), data-driven energy forecasting model and DPC for building control are increasingly studied in academia and applied in industry. However, three gaps impede the development of high-fidelity and cost-effective data-driven building energy forecasting models and predictive control strategies: Gap 1: Active learning, the key to defy data bias in building operation data, is hardly studied and applied to the area of data-driven building energy forecasting modeling; Gap 2: Feature selection to defy high data dimensionality is widely applied to building energy modeling process but there lacks a systematic and scalable methodology; Gap 3: Active learning and feature selection have not been systematically integrated for whole building DPC application. In this dissertation, to address the three gaps mentioned above, three research objectives are proposed: Objective 1: Develop active learning strategies in the application of data-driven building energy forecasting modeling to defy data bias; Objective 2: Develop a systematic feature selection procedure in the application of data-driven building energy forecasting modeling to defy high data dimensionality; Objective 3: Develop an integrated active learning and feature selection framework for data-driven building energy forecasting modeling used for whole building DPC application. In this thesis, the integrated framework of active learning and feature selection is developed to improve the performance of data-driven building energy forecasting modeling that can be used for future DPC applications. The framework provides a systematic methodology and automatic workflow that starts with collecting raw data from BAS to the establishment of data-driven energy models and DPC controllers. The developed strategies and framework are evaluated in a number of virtual and real building testbeds. Improved performance is observed from the building energy forecasting models built using the developed active learning strategy, systematic feature selection procedure, and integrated framework of active learning and feature selection, respectively. A DPC controller is also developed using an energy forecasting model built with the developed framework. Using virtual testbeds, the developed DPC controller is demonstrated to have better performance, in terms of total electricity cost, peak load shifting capability, and average CPU time, which further shows the effectiveness of the developed framework.

Data-driven Whole Building Energy Forecasting Model for Data Predictive Control

Files and links (1)

Abstract

Metrics

Details

Data-driven Whole Building Energy Forecasting Model for Data Predictive Control

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media