Economic MPC and real-time decision making with application to large-scale HVAC energy systems

https://doi.org/10.1016/j.compchemeng.2017.10.038Get rights and content

Highlights

  • A hierarchical decomposition of economic MPC in large-scale HVAC systems is proposed.

  • Decomposition uses a two-layer approach to optimize both airside and waterside.

  • Sample optimization of a 500-zone campus is performed.

  • An application of these ideas at the Stanford University campus is presented.

Abstract

With the potential to decrease operating costs and improve energy efficiency, model predictive control (MPC) has been proposed as a replacement for traditional heuristic, PID, and other conventional control strategies for heating, ventilation, and air conditioning (HVAC) systems in commercial buildings. Due to the size of large commercial HVAC systems, implementing MPC as a single monolithic optimization problem is not practical nor desirable given real-time operating requirements. In this paper, we present a hierarchical decomposition for economic MPC in large-scale commercial HVAC systems using a two-layer approach. We show a sample optimization for a campus of 25 buildings with 500 total zones and a central plant consisting of eight chillers. Then, we discuss an application of the ideas presented here in the recently completed $485-million replacement of the Stanford campus heating and cooling systems and conclude with some of the control theory challenges presented by this new class of applications.

Introduction

Control of heating and cooling systems has become an active area of research due to the significant energy expenditures associated with commercial buildings. Most commercial applications use on/off and PID controllers for control of their heating, ventilation, and air conditioning (HVAC) systems (Afram and Janabi-Sharifi, 2014). Recent work has proposed model predictive control (MPC) as a promising solution for controlling HVAC systems (Oldewurtel et al., 2012, Mendoza-Serrano and Chmielewski, 2012).

Large commercial buildings are often subject to dynamic utility prices, most notably for electricity. These structures, aimed at flattening peaks in electricity consumption, come in the form of varying time-of-use charges, as well as peak demand charges based on the overall maximum rate of electricity usage over the course of a month (Ma et al., 2012a, Cole et al., 2012). MPC is able to take advantage of these price structures to shift the energy load from peak hours to off-peak hours using thermal energy storage (TES). Through load shifting, energy costs can be decreased by running equipment during periods of low resource prices. The mass of the building can be readily used as passive TES (Oldewurtel et al., 2010, Avci et al., 2013). Active TES, e.g., chilled water or ice tanks, can also be utilized for load shifting, generally at higher energy efficiency than passive storage (Henze, 2005, Hajiah and Krarti, 2012). It is important that such systems are effectively controlled, as some heuristic-based control strategies can actually lead to increased operating costs when a storage tank is added (Braun, 2007). Due to storage inefficiency, time-shifting production may actually increase secondary energy consumption (i.e., electricity usage). However, overall primary efficiency consumption is likely to decrease, as the additional electricity is produced during off-peak hours at higher efficiency. Thus, economic optimization can lead to both decreased costs and higher primary energy efficiency for building operators.

As shown in Fig. 1, there are two main parts of most large-scale HVAC systems: the airside and waterside systems. The airside system includes the zones in all of the buildings along with the associated air-handler units (AHUs) used in temperature regulation. The waterside system includes the central plant equipment used to meet heating and cooling loads. The aim of the control system is to decide zone temperature setpoints in the airside problem and equipment operation schedules in the waterside problem, which includes both continuous and discrete variables.

One barrier to the widescale deployment of MPC-based control systems in industry is the scale of commercial systems. Large-scale applications, such as the one depicted in Fig. 1, involve hundreds of buildings each with tens of zones. Solving a single optimization problem for these large systems is neither practical nor desirable.

To divide the single optimization problem into smaller ones, distributed MPC can be used as discussed by Rawlings and Mayne (2009, Chapter 6) and Christofides et al. (2013). Typically, iterative methods have been used for this purpose (Cai et al., 2016, Elliott and Rasmussen, 2012, Scherer et al., 2014, Lamoudi et al., 2011). However, these methods may involve many information exchanges and iterations, which are undesirable in practice due to the limitations to information exchanges based on existing HVAC communication protocols (Moroşan et al., 2010). Additionally, the computation time to convergence may be longer than solving the single large optimization. Finally, demand charges require coordination amongst the zones in the airside system in order to avoid a high peak cost. Existing distributed methods do not address the full complexity of commercial applications, including peak demand charges, active TES, or detailed waterside equipment models.

In the central plant, the key challenge is the combinatorial problem that must be solved to sequence the multiple chillers, pumps, etc., while accounting for variable equipment efficiency. Although near-optimal heuristics are available (Braun, 2007), additional benefit can be achieved via optimization. Global search heuristics like particle swarm optimization (Ali et al., 2013) or optimization methods like mixed-integer linear programming (MILP) (Deng et al., 2015) have been applied to make the necessary discrete choices. Due to its small contribution to total cost, auxiliary equipment like pumps and cooling towers can be modeled to relatively low accuracy, e.g., in a single lumped model (Kapoor et al., 2013). This simplification reduces the number of discrete choices and thus also the computation time. For models where system dynamics depend on whether the TES is charging or discharging, an additional discrete decision variable can be added, often leading to a linear switched-system model (Mayer et al., 2015, Feng et al., 2015). Many strategies for the waterside subsystem assume that the total cooling load is a known parameter (Ma et al., 2012b, Powell et al., 2013, Deng et al., 2015), which effectively decouples the two subsystems, assuming the load estimate is near-optimal.

A final challenge is the presence of nonlinearities throughout the system. For the airside system, bilinear terms arise from the product of temperatures and flow rates (Ma et al., 2015), while for the waterside, equipment models are often polynomial in input variables (Lee et al., 2012). These difficulties can be eliminated by linearizing along a reference trajectory (Deng et al., 2015), but the resulting solution can be inaccurate if the reference trajectory is suboptimal. Successive linearizations can be performed until convergence (Oldewurtel et al., 2012), although convergence to a global optimum is not guaranteed. For both the airside and waterside systems, care is needed to balance model accuracy with computational complexity.

The remainder of this paper is organized as follows. In the next section, we present the hierarchical structure for the proposed controller and compare to other approaches in the literature. We then discuss some of the practical considerations for this controller. Next we present a brief case study simulating the proposed controller on a large cooling plant. The paper concludes with a brief description of the first, recently completed large-scale application of the ideas in the Stanford Energy Systems Innovations Project, and a discussion of what new control theoretic results might prove useful in this new class of applications.

Section snippets

Control structure

Due to the large number of zones on the airside and the discrete decisions on the waterside, a monolithic optimization problem considering all aspects of the problem in detail would be difficult to maintain and computationally intractable for online use. Therefore, we decompose the problem into a high-level subproblem that considers both the airside and waterside subsystems in approximate form and two low-level subproblems that consider each of the subsystems in more detail. Setpoints

Assumptions and simplifications

In the high-level problem, both the airside and waterside systems are modeled in only approximate form. For the airside system, two main simplifications are made. First, it is assumed that the decision variable is the cooling load delivered to the zone, rather than temperature setpoint for a local controller. This reduction allows the local controller model to be omitted. Second, the detailed zone models are reduced by assuming certain collections of zones are completely coupled and can thus be

System description

As an example system, we consider a large central cooling plant serving a campus of buildings. The cooling plant consists of eight conventional chillers with supporting pumps and cooling towers. Each chiller has minimum and maximum cooling capacities of 2.5 MW and 12.5 MW for a total plant capacity of 100 MW cooling. The nonlinear model and piecewise-linear approximation for each chiller is shown in Fig. 3. Chilled water supply temperature is fixed at 5.5 °C. The central plant also contains a

Application – Stanford Energy Systems Innovations project

Stanford University recently completed a $485-million Stanford Energy System Innovations (SESI) project to replace completely their 155-building campus heating and cooling system (Fig. 9). Johnson Controls, Inc. designed the control system for the new system. The deployed control system is approximately one-half of the top and bottom levels shown in Fig. 2. The waterside system is modeled similarly to the approach described in this paper, while the airside load is determined from historical

Economic versus tracking problems

After even a cursory examination of Figs. 5–, we can conclude that there is no tracking problem of the type considered in the vast majority of the MPC theory literature. Traditional MPC theory (stabilize the origin) was fashioned after the dominant application in the chemical process industries, which was to maintain a plant for reasonably long times at a fixed steady-state setpoint determined by another layer, the so-called real-time optimization (RTO) layer. One can debate whether this early

Conclusions and future outlook

Large-scale commercial applications provide a great opportunity for cost savings through optimal control of their HVAC systems. The existing industry-standard technology is not capable of achieving these savings. In this paper, a scalable decomposition has been presented for control of the coupled airside and waterside systems. The simulation results show that this control architecture can be used for large-scale applications with a large central plant and hundreds of zones.

Future research

Acknowledgments

The authors are grateful to Johnson Controls, Inc. for sample data, equipment models, and research funding. The authors gratefully acknowledge the financial support of the NSF through grant #CTS-1603768.

References (57)

  • G.P. Henze et al.

    Optimal design and operation of a thermal storage system for a chilled water plant serving pharmaceutical buildings?

    Energy Build.

    (2008)
  • M. Killian et al.

    Ten questions concerning model predictive control for energy efficient buildings

    Build. Environ.

    (2016)
  • M. Killian et al.

    Hierachical fuzzy MPC concept for building heating control?

    IFAC Proc. Vol.

    (2014)
  • T.-S. Lee et al.

    Evaluation of the suitability of empirically-based models for predicting energy performance of centrifugal water chillers with variable chilled water flow?

    Appl. Energy

    (2012)
  • J. Ma et al.

    Demand reduction in building energy systems based on economic model predictive control

    Chem. Eng. Sci.

    (2012)
  • B. Mayer et al.

    Management of hybrid energy supply systems in buildings using mixed-integer model predictive control

    Energy Convers. Manage.

    (2015)
  • P.-D. Moroşan et al.

    Building temperature regulation using a distributed model predictive control

    Energy Build.

    (2010)
  • F. Oldewurtel et al.

    Use of model predictive control and weather forecasts for energy efficient building climate control

    Energy Build.

    (2012)
  • K.M. Powell et al.

    Optimal chiller loading in a district cooling system with thermal energy storage?

    Energy

    (2013)
  • J.B. Rawlings et al.

    Model predictive control with discrete actuators: theory and application

    Automatica

    (2017)
  • M.J. Risbeck et al.

    A mixed-integer linear programming model for real-time cost optimization of building heating, ventilation, and air conditioning equipment

    Energy Build.

    (2017)
  • H. Scherer et al.

    Efficient building energy management using distributed model predictive control?

    J. Proc. Control

    (2014)
  • K. Subramanian et al.

    A state-space model for chemical production scheduling

    Comput. Chem. Eng.

    (2012)
  • C.R. Touretzky et al.

    A hierarchical scheduling and control strategy for thermal energy storage systems

    Energy Build.

    (2016)
  • R.P. Aguilera et al.

    Stability analysis of quadratic MPC with a discrete input alphabet?

    IEEE Trans. Automat. Control

    (2013)
  • D.A. Allan et al.

    Stability and robustness of model predictive control with discrete actuators

  • D. Angeli et al.

    On average performance and stability of economic model predictive control?

    IEEE Trans. Automat. Control

    (2012)
  • M. Baotic et al.

    Constrained optimal control of hybrid systems with a linear performance index?

    IEEE Trans. Automat. Control

    (2006)
  • Cited by (90)

    View all citing articles on Scopus
    View full text