Event-based optimization with non-stationary uncertainties to save energy costs of HVAC systems in buildings

Abstract

Building accounts for nearly 40% of global energy consumption, and about 40% of that is consumed by HVAC systems. A typical way of saving HVAC energy cost is to formulate and solve the HVAC operation problem which minimizes the HVAC energy cost in 24 hours ahead. Traditionally, the problem is solved by using time-based approaches where decisions are calculated and executed at each discrete time instant. In this paper, an innovative event-based approach is developed in the Lagrangian relaxation framework so that the decisions are only calculated and executed on an “as needed” basis to reduce computational requirements and extend device lifetimes. Developing such an event-based approach is challenging since with a finite time horizon of 24 hours and non-stationary uncertainties in weather, cooling load, etc., there is no steady-state solution. Events and actions are therefore time-dependent, causing the policy space to be extremely large. Our key idea to overcome this difficulty is to 1) include time-dependent variables that affect decisions in the definition of events so that events and actions will become time-independent and the size of event-based policy will be reduced significantly; and 2) develop a Q-learning method based on events within the Lagrangian relaxation framework to obtain the optimal actions. Numerical results demonstrate significant reductions of computational efforts as compared with time-based approaches with similar levels of energy savings and human comfort.