Fundamental Limitations of Distributed Feedback Control in Large-Scale Networks
Sammanfattning: Networked systems accomplish global behaviors through local feedback interactions. The purpose of a distributed control design is to select interaction rules and control protocols that achieve desired global control objectives. In this thesis, we address the question of fundamental limitations to such control designs, in terms of the global performance that is achievable in large-scale networks. We consider networked dynamical systems with single- and double- integrator dynamics controlled with linear consensus-like protocols. Such systems can be used to model, for example, vehicular formation dynamics and synchronization in electric power networks. We assume that the systems are subject to distributed disturbances and study performance in terms of H2 norm metrics that capture the notion of network coherence. In the context of power networks, we also show how such metrics can be used to quantify resistive losses caused by non-equilibrium, or transient, power flows due to a lack of synchrony. Distributed static feedback control based on localized, relative state measurements is subject to known limitations that, for example, cause coherence metrics to scale unfavorably with network size in lattices of low spatial dimensions. This causes an inevitable lack of rigidity in one-dimensional formations, such as strings of vehicles. We show here that the same limitations in general apply also to dynamic feedback controllers that are locally of first order. The proof relies partly on a fundamental limitation of localized relative feedback in networks of integrators of order three or higher, which we show to cause instability if the network grows beyond a certain finite size. This result holds unless the controller can access measurements of its local state with respect to an absolute reference frame, in which case dynamic feedback in the form of distributed derivative or integral control can fundamentally improve performance. This case applies, for example, to frequency control in power networks. However, if the absolute state measurements are subject to noise, the advantage of the distributed integral controller in terms of its performance scaling is lost. We show that scalable integral control of networks in principle requires centralization or all-to-all communication. For electric power networks, we show that performance in terms of transient power losses scales with the number of generator nodes in a network. However, in sharp contrast to the previous results, an increased connectivity does not in general improve performance. We discuss possible implications of these results in terms of the design of future power grids with increasingly distributed electricity generation.
KLICKA HÄR FÖR ATT SE AVHANDLINGEN I FULLTEXT. (PDF-format)