Date of Award

Spring 1-1-2012

Document Type


Degree Name

Doctor of Philosophy (PhD)

First Advisor

Gregor P. Henze

Second Advisor

Balaji Rajagopalan

Third Advisor

Michael J. Brandemuehl

Fourth Advisor

Moncef Krarti

Fifth Advisor

Gail S. Brager


Model predictive control (MPC) is a powerful technique that can be used to reduce the operational cost, energy consumption, and environmental footprint of buildings. MPC optimizes control decisions to minimize the objective function produced by a building energy model and has been successfully applied to a range of control problems in buildings, usually thermal mass storage. Parametric simulation studies are typically conducted, and the resulting solution patterns are used to inform control strategies. A model predictive controller can also directly control building equipment, but in order to achieve faster solution convergence needed for real-time implementation, reduced-order gray- and black-box models are often employed that can be optimized through linear or quadratic programming.

Despite the widespread potential for thermal mass control in buildings, MPC of this kind is challenging to implement due to the necessity of reduced-order models and the need to integrate with building automation systems (BAS). This dissertation examines the possibility of using MPC conducted on white-box building energy models—the same types used to evaluate building designs—to develop datasets from which near-optimal control rules can be extracted using supervised learning techniques. This allows for the development of custom supervisory controllers that more closely approximate optimal energy and thermal comfort results compared to conventional control heuristics. Rules are developed in such a form that they can be implemented in a conventional BAS. The dissertation uses the case of mixed mode (MM) buildings to test these techniques. A proof-of-concept rule extraction case is first presented for a simple binary natural ventilation control problem to test the utility of several data mining and statistical techniques to the problem, including generalized linear models (GLM), classification and regression trees (CART) and adaptive boosting. Next, a simulation study is conducted to explore a variety of more complex MM optimal control problems on four different MM building types and in five different climates. Two of these cases form the training set for further rule extraction, testing the applicability of this technique beyond simple binary decisions. CARTs were found to be successful in reproducing optimal supervisory control sequences, often yielding greater than 90% of optimizer energy savings with minimal thermal comfort consequences. Robustness of extracted rules and generalizability to broader cases (e.g. other building types and climates) is examined. Finally, an experiment is presented in which the energy and comfort performance of extracted rules are tested on a radiantly cooled test cell. The impacts of model calibration mismatch and weather forecast uncertainty are examined and are found to contribute significantly to the reduced experimental performance of the rules.

The research provides two key outcomes for the larger building community. For designers of MM buildings, the simulation study provides for the first survey of MM performance under optimal control and identifies preferred strategies by climate and building type. For building control engineers, the rule extraction framework provides a new and innovative means for analyzing MPC solutions and implementing near-optimal rules based on those solutions. The research presents the first step in what will hopefully be a new vein of building controls research and eventually, controls practice. Future research must further examine the robustness of the approach and its operational performance in “live” buildings.