Model Predictive Control for Energy Efficient Buildings

The building sector consumes about 40% of energy used in the United States and is responsible for nearly 40% of greenhouse gas emissions. Energy reduction in this sector by means of cost-effective and scalable approaches will have an enormous economic, social, and environmental impact. Achieving substantial energy reduction in buildings may require to rethink the entire processes of design, construction, and operation of buildings. This thesis focuses on advanced control system design for energy efficient commercial buildings.Commercial buildings are plants that process air in order to provide comfort for their occupants. The components used are similar to those employed in the process industry: chillers, boilers, heat exchangers, pumps, and fans. The control design complexity resides in adapting to time-varying user loads as well as occupant requirements, and quickly responding to weather changes. Today this is easily achievable by over sizing the building components and using simple control strategies.Building controls design becomes challenging when predictions of weather, occupancy, re- newable energy availability, and energy price are used for feedback control. Green buildings are expected to maintain occupants comfort while minimizing energy consumption, being ro- bust to intermittency in the renewable energy generation and responsive to signals from the smart grid. Achieving all these features in a systematic and cost-effective way is challenging. The challenge is even greater when conventional systems are replaced by innovative heat- ing and cooling systems that use active storage of thermal energy with critical operational constraints.Model predictive control (MPC) is the only control methodology that can systematically take into account future predictions during the control design stage while satisfying the system operating constraints. This thesis focuses on the design and implementation of MPC for building cooling and heating systems. The objective is to develop a control methodology that can 1) reduce building energy consumption while maintaining indoor thermal comfort by using predictive knowledge of occupancy loads and weather information, (2) easily and systematically take into account the presence of storage devices, demand response signals from the grid, and occupants feedback, (3) be implemented on existing inexpensive and distributed building control platform in real-time, and (4) handle model uncertainties and prediction errors both at the design and implementation stage.The thesis is organized into six chapters. Chapter 1 motivates our research and reviews existing control approaches for building cooling and heating systems.Chapter 2 presents our approach to developing low-complexity control oriented models learned from historical data. Details on models for building components and spaces thermal response are provided. The thesis focuses on the dynamics of both the energy conversion and storage as well as energy distribution by means of heating ventilation and air conditioning (HVAC) systems.In Chapter 3, deterministic model predictive control problems are formulated for the en- ergy conversion systems and energy distribution systems to minimize the energy consumption while maintaining comfort requirement and operational constraints. Experimental and simu- lative results demonstrate the effectiveness of the MPC scheme, and reveal significant energy reduction without compromising indoor comfort requirement.As the size and complexity of buildings grow, the MPC problem quickly becomes com- putationally intractable to be solved in a centralized fashion. This limitation is addressed in Chapter 4. We propose a distributed algorithm to decompose the MPC problem into a set of small problems using dual decomposition and fast gradient projection. Simulation results show good performance and computational tractability of the resulting scheme.The MPC formulation in Chapter 3 and 4 assumes prefect knowledge of system model, load disturbance, and weather. However, the predictions in practice are different from actual realizations. In order to take into account the prediction uncertainties at control design stage, stochastic MPC (SMPC) is introduced in Chapter 5 to minimize expected costs and satisfy constraints with a given probability. In particular, the proposed novel SMPC method applies feedback linearization to handle system nonlinearity, propagates the state statistics of linear systems subject to finite-support (non Gaussian) disturbances, and solves the resulting optimization problem by using large-scale nonlinear optimization solvers