Iterative partition-based moving-horizon state estimation

  • Iterative partitionierungsbasierte Zustandsschätzung auf bewegtem Horizont

Schneider, René; Marquardt, Wolfgang (Thesis advisor); Scattolini, Riccardo (Thesis advisor)

Aachen : Shaker Verlag (2017)
Book, Dissertation / PhD Thesis

In: Berichte aus der Verfahrenstechnik
Page(s)/Article-Nr.: xviii, 146 Seiten : Illustrationen, Diagramme

Dissertation, RWTH Aachen University, 2016


This thesis proposes a set of novel partition-based moving-horizon state estimation schemes for systems that are composed of interconnected and potentially geographically distributed subsystems. The state variables of every subsystem are viewed as partitions of the full state vector, and they are estimated in parallel by dedicated subsystem state estimators that can exchange information among each other. Not only can this approach be faster than a centralized moving-horizon estimator, but it also avoids the unfavorable dependence on a single central computer. The novelty of the proposed methods is their iterative nature. As a result, their estimation accuracy approximates the optimal estimation accuracy of centralized moving-horizon estimators arbitrarily well. Depending on which of the proposed methods is employed, the state of linear or nonlinear systems can be estimated. Moreover, different types of process and measurement uncertainties as well as additional inequality constraints of varying complexity can be taken into account. Theoretical results are developed and proven, which provide conditions for the convergence of the iterations at every sampling instant and for the stability of the estimation error as time proceeds. In particular, one of the proposed methods has the unique feature of simultaneous convergence and stability for a certain class of linear systems, independent of their subsystem topology. These theoretical results are validated by extensive numerical simulations, which also provide additional insights into the role of the different parameters and problem formulations on the dynamical behaviour of the proposed estimators. Further simulations illustrate the potential of the novel iterative partition-based moving-horizon estimators for industrial applications. First, the dynamic state estimation problem of large-scale power system networks is addressed. Assuming ideal parallelization, the proposed partition-based estimator is found to be nearly as accurate but faster than a centralized moving-horizon estimator. Secondly, by combining one of the proposed methods with a distributed model predictive controller from the literature, a completely distributed and optimization-based output feedback solution is developed and successfully applied to a chemical plant. Finally, a summary of the results and suggestions for promising future research directions conclude the thesis.