Data-driven and Safe Networked Control with Applications to Microgrids

Today, automatic control is integrated into a wide spectrum of real-world systems such as electrical grids and transportation networks. Many of these systems comprise numerous interconnected agents, perform safety-critical operations, or generate large amounts of data. Their automation, therefore, must address the challenges pertinent to these three characteristics. Specifically, in multi-agent systems (MASs), one common objective is that agents reach consensus on state or output variables for achieving a desired collective behavior. For consensus in large MASs, distributed controllers employing a communication network are required. When distributed architectures are used, the security of the closed-loop system can be compromised due to the vulnerability of the communication network to adversarial interferences called cyber attacks. Timely detection of attacks is vital for ensuring security of safety-critical systems. Finally, it is relevant in data-rich systems to develop control and state-estimation methods that provide guarantees by using a finite amount of data and bypassing the need of knowing system models.Composed of three parts, this thesis contributes to address the abovementioned challenges of modern control systems. As an example, DC microgrids (DCmGs) are considered throughout the thesis for the development and validation of the proposed methods. The first part focuses on distributed consensus protocols. We first investigate the problem of state consensus in general linear interconnected MASs (LIMASs), for which we provide conditions based on the physical and communication graph properties as well as system matrices. Our results show that weak physical coupling and well-connected graphs are favorable features for consensus. We then focus on nonlinear DCmGs and propose a novel distributed controller to achieve voltage balancing and current sharing, which is a specific case of output consensus. By exploiting the structure of DCmG dynamics, we provide conditions on the attainment of these objectives and the stability of the closed-loop system. In the second part, a distributed cyber-attack detection scheme is developed for LIMASs controlled in a distributed fashion. The detection architecture comprises local monitoring units collocated with each agent and checking the presence of cyber attacks in variables communicated by neighboring agents. Each unit estimates the states of local and nearby agents, and detects an attack if a suitably defined error is sufficiently large. A thorough detectability analysis considering different types of attacks is also performed. The final part of the thesis provides direct data-driven methods for control and state estimation based on finite data. First, we look at the worst-case optimal tracking problem in presence of measurement noises satisfying a quadratic bound. Control design is formulated as a semidefinite program (SDP), whose computational complexity is independent of the amount of data. Then, we turn our attention to state estimation in presence of unknown inputs, for which we present data-driven necessary and sufficient conditions. Under these conditions, a novel data-driven state estimation method with stability guarantees is provided.

A comparison of model-based and data-driven controller tuning

Alireza Karimi, Klaske van Heusden

In many industrial applications, finding a model from physical laws that is both simple and reliable for control design is a hard and time-consuming undertaking. When a set of input/output measurements is available, one can derive the controller directly from data, without relying on the knowledge of the physics. In the scientific literature, two main approaches have been proposed for control system design from data. In the 'model-based' approach, a model of the system is first derived from data and then a controller is computed-based on the model. In the 'data-driven' approach, the controller is directly computed from data. In this work, the previous approaches are compared from a novel perspective. The main finding of the paper is that, although from the standard perspective of parameter variance analysis the model-based approach is always statistically more efficient, the data-driven controller might outperform the model-based solution for what concerns the final control cost. Copyright (C) 2013 John Wiley & Sons, Ltd.

Wiley-Blackwell2014

Real-Time Optimal Controls for Active Distribution Networks

Konstantina Christakou

Decentralized generation, distributed energy storage systems and active participation of end-users in the lower level of the electrical infrastructure, intelligently managed to provide grid support, define the notion of Active Distribution Networks (ADNs). The presence of distributed generation in ADNs incurs severe impacts on planning and operational procedures and calls for intelligent control techniques. This thesis focuses on the compelling problem of optimal operation and control of ADNs, with particular reference to the design of real-time voltage control and lines congestion management algorithms. In the first part of the thesis, we adopt a centralized architecture for voltage control and lines congestion management in ADNs. The goal of the proposed controller is to schedule the active and reactive power injections of a set of controllable resources, in coordination with traditional resources, in order to achieve an optimal grid operation. The controller relies on a linearized approach that links control variables and controlled quantities using sensitivity coefficients. Once the proposed algorithm is validated, as a further step, we relax the assumption that the DNO has an accurate knowledge of the system model, i.e., a correct admittance matrix and we adapt the proposed control architecture to such a scenario. When the controllable resources are heterogeneous and numerous, control schemes that rely on two-way communication between the controllable entity and the DNO cannot scale in the number of network buses and controllable resources. In this direction, in the second part of this thesis, we propose the use of broadcast-based control schemes that rely on state estimation for the feedback channel. We propose a low-overhead broadcast-based control mechanism, called Grid Explicit Congestion Notification (GECN), intended for provision of grid ancillary services by a seamless control of large populations of distributed, heterogeneous energy resources. Two promising candidates in terms of controllable resources are energy storage systems and elastic loads. Therefore, we choose to validate GECN in the case of aggregations of thermostatically controlled loads, as well as of distributed electrochemical-based storage systems. In the last part of the thesis, we formulate the control problem of interest as a non-approximated AC optimal power flow problem (OPF). The AC-OPF problem is non-convex, thus difficult to solve efficiently. A recent approach that focuses on the branch-flow convexification of the problem is claimed to be exact for radial networks under specific assumptions. We show that this claim, does not hold, as it leads to an incorrect system model. Therefore, there is a need to develop algorithms for the solution of the non-approximated, inherently non-convex OPF problem. We propose an algorithm for the AC-OPF problem in radial networks that uses an augmented Lagrangian approach, relies on the method of multipliers and does not require convexity. We design a centralized algorithm that converges to a local minimum of the original problem. When controlling multiple dispersed energy resources, it is of interest to define also a distributed method. We investigate the alternating direction method of multipliers (ADMM) for the distributed solution of the OPF problem and we show cases for which it fails to converge. As a solution we present a distributed version of the proposed OPF algorithm that is based on a primal decomposition.

EPFL2015

Real-Time Optimal Controls for Active Distribution Networks

Konstantina Christakou

EPFL2015