The promise of harnessing the predictive and diagnostic powers of a well-calibrated building energy model for improved building operations is attracting a growing number of building sciences researchers around the globe. In fact, model predictive control (MPC) as applied to buildings, both commercial and residential, is currently experiencing a bona fide explosion of research activity.
What does MPC applied to buildings entail? Central to MPC is the notion that there is value to be found by considering future building and energy system behaviour as estimated by a suitable model – value that remains untapped by the building control architecture we are accustomed to today. While traditional control strategies consider only concurrent building information, in MPC the current control action is determined by solving a receding horizon optimal control problem in real time at each sampling instance, using the current state of the system under control as the initial state. While the optimization yields an optimal sequence spanning the planning horizon, only the first control action is applied to the controlled system, the initial state of the model is then reset at the next sampling instance, predictions are updated, and a new optimization is performed, repeating the cycle.
In the commonly applied receding horizon implementation, a suitable performance criterion of choice – with likely candidates of energy consumption, operating cost, indoor air quality, and peak demand –is minimized at every instance subject to constraints. It is this recognition of constraints on both action and state variables, from simple bounds on space temperatures to nonlinear constraints on equipment operation, combined with the advantages of using expected diurnal variations in weather patterns, utility pricing, and occupancy patterns, that allows for the design of effective control strategies spanning time scales of hours to days for a wide range of building control applications.
MPC is most effective when the presence of some form of thermal energy storage needs to be considered, such as passive building thermal mass or active systems such as ice and chilled water storage. Recent applications involve planning the charging and discharging control strategies for large university campus chilled water plants with thermal energy storage systems, radiant heating, and cooling systems, along with natural ventilation and shading control that manipulate free cooling and solar gains in concert with the available building thermal mass. Moreover, building systems with restricted state transitions, such as chillers engaging and disengaging in multiple chiller plants subject to lockout periods, may benefit from a planned operational strategy using a well-calibrated system model. Although MPC may be used for low-level implementations, such as controlling heating plant output to optimally anticipate building comfort requirements after periods of temperature setback, it is more often used for high-level implementations, such as determining the cooling temperature set point profile for the next day, with the underlying local control loops previously confirmed or commissioned to operate satisfactorily – MPC in most cases thus working to improve building operation in tandem with an existing building automation system by serving the role of an operations advisor.
The required system model can be an explicit linear model such as a transfer function, a state space model, or a set of ordinary differential equations and differential algebraic equations, either defined from known information in a forward fashion or inversely identified using parameter estimation. Alternatively, to account for nonlinearities, for example due to saturation and hysteresis effects, building models may be expressed in an implicit fashion using detailed building simulation programs, with the optimization seeing the function evaluator as a black box.
What is behind the rapid growth in attention to MPC in buildings? Without claim of completeness, a few likely drivers are
(1) The building science community is looking to identify cost-effective energy efficiency opportunities, and operational improvements of existing buildings are a dominant part of the opportunity. | |||||
(2) Advances in computing power enable higher dimensional investigations with quicker convergence. | |||||
(3) The growth in building automation and availability of near-real-time information about building operation have increased the stock of existing buildings that can potentially benefit from MPC applications. | |||||
As evidence for this trend, during the first MPC in Buildings Workshop held in Montreal, 24–25 June 2011, approximately 80 building science researchers flocked from around the globe to learn new approaches and share their findings, all of which can be found at http://www.ibpsa.us/mpc2011/. The exciting two-day programme included 23 presentations, many of which dealt with MPC simulation studies, which to date dominate the literature, but the sessions also considered field and experimental studies, MPC environments and methodology, as well as contributions on model complexity and model mismatch. The meeting ended with a roundtable discussion that synthesized the state- of-the-art, discussed visions and collective research goals, and identified a set of obstacles to be overcome for MPC to fulfil its potential as a disruptive building control paradigm.
The ideal implementation scenario is one in which the operator embraces the new technology, helps improve it, and ends up emerging as a more informed and proactive building operator, achieving a higher performance building, all while playing at the high end of the skill range. To an extent, the MPC system ends up educating the building operator. First-hand field experience confirms this notion. A large, 100,000 m2 building in downtown Chicago was previously never operated during a hot August weekday without all three chillers running during the late afternoon. Using a cloud-based real-time MPC system that relied on a detailed and calibrated model of the building, an alternative demand reducing and real-time price-based strategy was found and employed that required only two instead of three chillers, to the notable astonishment of the operating team. Having gained trust in the model, the same building operators now inquired to conduct what-if scenarios using the same calibrated building energy model, from operational changes to investment decisions such as chiller upgrades.
Yet, there are obstacles. Prime among them is articulating convincing arguments for the adoption of MPC to relevant stakeholders such as building owners or operators and to create buy-in at the management level. These arguments include operating cost savings, improved thermal comfort and indoor air quality, higher occupant satisfaction, and a greener and more sustainable image. Of course, measurement and verification strategies of said performance improvements along with performance metrics must be agreed upon, similar to retrofit projects within energy savings performance contracts.
Through highly visible showcase projects, the evidence must be delivered that MPC can be cost-effective in spite of the high engineering effort associated with creating a high-fidelity model, e.g. by avoiding the installation of the last chiller through a demand limiting strategy. A public database of successful projects employing MPC will be essential in scaling from individual projects to a repository of best practices, serving both to identify appropriate building types and to help educate personnel in an effort to build the skilled human capacity needed for such a quantum leap in control system complexity.
Given MPC's reliance on continuous optimization of a potentially opaque building model, the risk of disenfranchising the building operator must be addressed. Hence, the operator who will need to work alongside a real-time MPC system must be motivated to become and stay a friend, not foe, of such a sophisticated operations advisor. Obviously, achieving the requisite rapport between man and machine is a process that must involve mechanisms for collecting feedback on questionable MPC decisions and helping the building operator to develop confidence in the MPC system rather than suspicion.
While historically separate disciplines, building simulation and building controls may thus converge in a joint role: ambitious performance targets enabled by predictive controls are scrutinized and verified by a motivated building personnel that has built trust in the model-based operations approach and can periodically explore additional savings through retrofit measures that have likely suggested themselves through the review of the MPC strategies.
On the technical side, a recurring topic is the requisite model complexity, with workshop attendees offering estimates of 70% of project costs attributable to model creation and calibration. While simpler yet physically motivated inverse grey box models dominated the applications presented, advocates of detailed building energy simulation models offered valid motivation for their use such as the recognition of spatial diversity in utilization patterns and comfort requirements as well as the ability to model complex heating, ventilating, and air-conditioning systems and their operational strategies. Moreover, calibrated detailed models can be more easily used for other related applications, such as retrofit analysis, or fault detection and diagnosis.
With this much research activity currently ongoing and likely to increase, the question begs whether the time has come for a unified and open-source software environment for building MPC; one that avoids duplicating efforts and can be applied to both offline simulation and online control cases, offering benefits to the research community that offset the loss of advantages of a potentially superior individual effort. Such an environment would make extensive use of distributed computing to leverage high-performance computing resources in the cloud, would allow for exploration of model complexity in response to MPC objectives, and would consider a variety of hardware implementation options.
This special issue on MPC in buildings boasts a cross-section of contributions that were developed based on works presented at the 2011 workshop: from a discussion of two simulation environments using both detailed building simulation and simplified inverse grey box models, offline rule extraction by means of machine learning as well as lookup tables as an approach to tackle real-time control challenges, and finally an investigation into the existence of local minima and methods to guarantee global optimality in a building temperature control application. At the time of writing, a second workshop on intelligent building operation was being organized for the summer of 2013 in Boulder, Colorado; testament to the vitality of this exciting research domain.