Search

Commercial buildings account for $200 billion per year in energy expenditures, with heating, ventilation, and air conditioning (HVAC) systems accounting for most of these costs. In energy markets with time-varying prices and peak demand charges, a significant potential for cost savings is provided by using thermal energy storage to shift energy loads. Since most implementations of HVAC control systems do not optimize energy costs, they have become a primary focus for new strategies aimed at economic optimization. Model predictive control (MPC) has emerged as one popular method to achieve this load shifting, while respecting system constraints. MPC uses a model of the system to make predictions and to solve an optimization problem. Much research has shown the benefits of MPC over alternative strategies for HVAC control [1]. However, some industrial applications, such as large research centers or university campuses, are too large to be solved in a single MPC instance. Decompositions have been proposed in the literature, but it is difficult to evaluate and to compare decompositions against one another when using different systems. In this paper, we present a large-scale relevant case study where solving a single MPC optimization problem is neither desirable nor feasible for real-time implementations. The study is modeled after the Stanford University campus, consisting of both an airside and waterside system [2]. The airside system includes 500 zones spread throughout 25 campus buildings along with the air handler units and regulatory building automation system used for temperature regulation. The waterside system includes the central plant equipment, such as chillers, that is used to meet the load from the buildings. Active thermal energy storage is available to the campus in addition to the passive thermal energy storage present in the form of building mass. The airside models describe the temperature dynamics in each of the 500 zones, and the waterside models describe the power consumption of the central plant equipment. The aim of the control system is to minimize costs in the presence of time-varying electricity prices and a peak demand charge as well as environmental disturbances such as weather while meeting constraints on comfort and equipment. We perform an economic optimization of the entire campus using a hierarchical system with distributed airside controllers to demonstrate the potential savings. The models from this case study are made publicly available for other researchers interested in designing alternative control strategies for managing chilled water production to meet airside loads. The aim of the case study release is to provide a standardized problem for the research community. A benchmark is provided for evaluating performance. References [1] A. Afram and F. Janabi-Sharifi. Theory and applications of HVAC control systems—A review of model predictive control (MPC). Building and Environment, 72:343–355, February 2014. [2] J. B. Rawlings, N. R. Patel, M. J. Risbeck, C. T. Maravelias, M. J. Wenzel, and R. D. Turney. Economic MPC and real-time decision making with application to large-scale HVAC energy systems. Computers & Chemical Engineering, 2017. In Press.

In this paper, we examine closed-loop operation of an HVAC central plant to demonstrate that closed-loop receding-horizon scheduling provides robustness to inaccurate forecasts, and that economic performance is not seriously impaired by shortened prediction horizons or inaccurate forecasts when feedback is employed. Using a general mixed-integer linear programming formulation for the scheduling problem, we show that optimization can be performed in real time. Furthermore, we demonstrate that closed-loop operation with a moderate prediction horizon is not significantly worse than a long-horizon implementation in the nominal case, and that closed-loop operation can correct for inaccurate long-term forecasts without significant cost increase. In addition, we show that terminal constraints can be employed to ensure recursive feasibility. The end result is that forecasts of demand need not be extremely accurate over long times, indicating that closed-loop scheduling can be implemented in new or existing central plants.