A low cost multi-agent control approach for building energy system management

Jie Cai, Purdue University

Abstract

More than 40% of the primary energy is related to energy consumption in buildings in the United States and if buildings are not operated properly, a significant amount of energy is wasted. This matter is becoming widely recognized and the subject of building optimal controls has attracted growing research efforts in the past few years. However, the deployment of advanced controls in buildings has been progressing very slowly because of the high implementation (sensors, software programming, etc.) cost. In particular, modeling of building energy systems is a challenging task due to the internal complexities, which poses an important barrier for model-based controller designs in buildings. This thesis presents a low-cost multi-agent control approach for managing building energy systems. Firstly, a general multi-agent framework is reported which defines a general agent structure along with the physical connections between agents. With the help of this framework, a multi-agent system can be easily setup and configured for any given building energy system. With some symbolic manipulations, the framework is able to automatically compose an optimization problem for the target system that can be solved either in a centralized manner or with some distributed optimization techniques. The controller design procedure is automated within the framework, which would reduce the engineering effort and, thus, the implementation cost significantly. The agent behavior can be either integrated in devices by manufacturers or identified on the fly using collected operating data. A specific effort has been spent on data-driven modeling of building energy systems to enable multi-agent controls. Building envelope and heating, ventilation and air conditioning (HVAC) equipment models are dealt with separately due to the significant difference in their time constants. A building envelope has relatively slow response, especially for buildings with high thermal mass elements made from concrete, brick, and/or stone. Thus, a dynamic thermal network model is used and the key parameters are estimated with training data. Different problems associated with developing data-driven models of envelopes are investigated including modeling of single-zone and multi-zone buildings as well as optimal excitation design to obtain an informative training data set. In contrast to building envelopes, HVAC equipment is usually fast in response and static models are typically used. This study considers two types of air-conditioning (AC) systems that are dominantly utilized in commercial buildings: direct-expansion (DX) and chilled water systems. A correlation-based model and a physical-component-based model are presented for DX units with capacity modulation and variable airflow, which were trained with field data and their performances were compared. For chilled water systems, a gray-box model is developed for the cooling coil, which is then integrated with an empirical chiller model to represent the overall system characteristics. Control-oriented models for other HVAC devices are also reported including models for supply fan, chilled water pump, etc. A multi-agent controller should be able to combine control heuristics with optimization-based approaches to provide a scalable and computationally tractable solution. Heuristic control approaches are based on some well-developed and general rules for a specific type of system. They should be easily integrated in devices and, most of the time, should be able to provide near-optimal performance. In this study, a simple control heuristic is proposed for a specific DX unit based on optimization results. With a small modification, the proposed heuristic is generalized to be applicable to any DX unit and the impact of different system configurations, pressure control schemes and climate conditions on the control performance is studied with the help of simulation models. For chilled water AC systems, a simple heuristic rule is also identified from optimization results of a representative system. By virtue of the developed control heuristic, the optimization problem of chilled water cooling systems can be formulated under a convex form, which is critical in ensuring convergence of the adopted optimization algorithms. For cases where near-optimal heuristics are not available, an optimization-based controller is needed. The developed multi-agent framework is able to synthesize a controller either in a centralized or a distributed (multi-agent) scheme depending on the complexity of the target building. As a test case, a multi-agent controller was synthesized and applied to a centralized AC system serving multi-zone buildings where the proposed control method was able to recover most of the energy savings. A heuristic-optimization combined control appears to be a promising approach since a general HVAC system usually consists of both heuristic-mature devices and heuristic-lacking devices. To test this idea and to demonstrate a scalable heuristic-integrated model predictive control (MPC) method for buildings, the proposed DX unit heuristics were integrated with a simplified MPC and performances of different control strategies were evaluated and compared. In addition, the developed heuristics for chilled water systems were utilized to develop a scalable and robust distributed MPC approach for a multi-zone building or a building cluster under a demand response (DR) scenario. With some moderate modification, a distributed optimization approach for a long-term energy management problem is proposed whose solution could be used as a benchmarking tool to study different DR strategies. Both the DMPC and long-term optimization approaches can be easily implemented within the multi-agent control framework where the heuristics are embedded in the HVAC agents.

Degree

Ph.D.

Advisors

Braun, Purdue University.

Subject Area

Architectural|Mechanical engineering

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS