Short-term forecasting for the electrical demand of Heating, Ventilation, and Air Conditioning systems

The heating, ventilation, and air conditioning systems (HVAC) of large scale commercial and institutional buildings can have significant contributions to the buildings overall electric demand. During periods of peak demand, utilities are faced with a challenge of balancing supply and demand while the system is under stress. As such, utility companies began to operate demand response programs for large scale consumers. Participation in such programs requires the participant to shift their electric demand to off-peak hours in exchange for monetary compensation. In such a context, it is beneficial for large scale commercial and institutional buildings to participate in such programs. In order to effectively plan demand response based strategies, building energy managers and operators require accurate tools for the short-term forecasting of large scale components and systems within the building. This thesis contributes to the field of demand response research by proposing a method for the short-term forecasting for the electric demand of an HVAC system in an institutional building. Two machine learning based approaches are proposed in this work: a component method and a system based method. The component-level approach forecasts the electric demand of a component within the HVAC system (e.g. air supply fans) using an autoregressive neural network coupled with a physics based equation. The system-level approach uses deep learning models to forecast the overall electric demand of the HVAC system through forecasting the electric demand of the primary and secondary system. Both approaches leverage available data from the building automation system (BAS) without the need for additional sensors. The system based forecasting method is validated through a case study for a single building with two data sources: measurement data obtained from the BAS and from an eQuest simulation of the building. The building used as the case study for the work herein consists of the Genomic building of Concordia University Loyola campus