Model Predictive Control
Deploying model predictive controllers on large-scale distributed systems is challenging since it requires solving large-scale optimization problems in real time on embedded hardware. One popular solution to this problem is to distribute the computations performed by the optimization algorithm amongst distributed platforms. Methods include matrix splitting, dual decomposition, the Alternating Direction Method of Multipliers (ADMM), an alternating minimization algorithm, a distributed inexact Newton method, and more. A challenge of creating a distributed controller for building HVAC systems is that the communication delay is quite large. The effect of this large communication delay is detrimental to the overall performance of the aforementioned distributed control solutions.
We have developed a novel primal-dual active-set algorithm that is particularly well suited for distributed implementation in systems with high communication costs. The algorithm is naturally decomposable for optimization on a star communication network for large-scale systems having dynamics coupled through the control inputs. Additionally, the decomposition fits a typical hardware architecture where the leaf nodes of the network have limited computational resources and the central hub of the star network has substantially more computing power. [link].
A Primal-Dual Active-Set Method for Distributed Model Predictive Control
Sarah Koehler, Claus Danielson, and Francesco Borrelli [pdf]
This material is based upon work supported by the National Science Foundation under grants CPS-1239552, and the NSF Graduate Research Fellowship Program. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.