Abstract

Heating, Ventilation and Air Conditioning (HVAC) systems of large buildings have a high contribution to the electric grid peak demand. During those periods, electric utilities face important mismatch issues in power supply and demand. In the context of Demand Response (DR) programs, there is a need from building energy managers for tools to forecast the electric demand of HVAC systems to plan for fast-DR control strategies. This thesis contributes to the DR research field by proposing a method for multi-step forecasting of the electric demand of existing HVAC cooling systems on the short-term in large commercial and institutional buildings.
Two forecasting methods are proposed: a cascade-based (global) method and a component-based method. The cascade-based method includes a sequence of forecasts of target variables. First, the air flow rate supplied by the AHUs is forecasted, followed by the cooling coils load, the cooling load of the whole building, and finally the electric demand of the primary cooling system is forecasted. The component-based method forecasts the electric demand of one component of the HVAC system such as a fan. Data-driven models are developed based on Building Automation System (BAS) trend data for electric demand forecasting of HVAC cooling system over the next six hours with a time-step of 15 minutes.
The large amount of data collected through the BAS presents a gold mine of information which could be used for better understanding of the actual building operation and performance. Data mining techniques are used as pre-processing steps to help in the development of the forecasting models, for the selection of regressors, to identify typical daily profiles of the target variable and to better understand the operation of HVAC systems. Different sequences of preprocessing steps are tested and their impact on the forecasting performance is compared.
The proposed forecasting methods are validated on two case studies: the Genomic research center on Loyola Campus of Concordia University and an office building located in Shawinigan-Sud (Québec). The thesis compares the forecasts with measurements, and discusses the quality of forecasting results.