Browsing by Author "Khatana, Vivek"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Balancing information mixing and optimality: a framework for robust and efficient distributed decision-making
    (2025-01) Khatana, Vivek
    Driven by the need for resiliency, scalability, and plug-and-play operation, distributed decision-making is becoming increasingly vital. This thesis develops a comprehensive framework to address the challenges in distributed decision-making, focusing on distributed optimization and control of multi-agent networks. It combines theoretical insights, algorithmic tools, and experimental validation to enhance decision-making in networked dynamical systems (NDS) with applications aimed at advancing clean and renewable energy systems. The proposed solutions emphasize distributed consensus, distributed optimization under practical constraints, and resilience against malicious agents and natural disasters in modern power systems. Achieving consensus in multi-agent systems is a cornerstone of distributed decision-making. The first part of this thesis makes a significant contribution in the development of distributed average consensus algorithms in multi-agent networks. By analyzing the geometry of the ratio consensus algorithm, this thesis introduces a finite-time distributed stopping criterion that guarantees convergence within any specified tolerance, regardless of the dimensionality of the state variables. The approach leverages the monotonicity of network state polytopes indexed by time. Additionally, the thesis presents a finite-time stopping criterion for networks with dynamic interconnection topologies, demonstrating that global maximum and minimum values remain strictly monotonic, even with dynamic links. The practicality of these algorithms is showcased through MATLAB simulations and experiments with Node.js-based agents. To address communication noise in the cyber-physical components of NDS, a resilient average consensus algorithm is proposed. Each agent updates its estimates using local information while incorporating weighted, noise-free initial values during iterations. The resilient average consensus algorithm has a geometric rate of convergence under noiseless conditions and almost-sure consensus under noisy communication. Numerical experiments confirm its effectiveness under varying noise scenarios and parameters. Part two of this thesis introduces a framework for distributed multi-agent optimization problems involving a common decision variable. A novel optimize then agree approach decouples optimization and consensus steps, ensuring disagreement between agents’ estimates remains below a predefined tolerance; existing algorithms do not provide such a guarantee which is required in many engineering scenarios. For the class of distributed optimization problems with local linear equality, inequality, and set constraints, we develop an algorithm that works over directed communication topologies and accrues all the benefits of the Alternating Direction Method of Multipliers approach. The algorithms synthesize distributively, communication overhead remains within a log factor of the lower bound, and guarantees strong convergence properties, achieving global geometric convergence rates for strongly convex and Lipschitz differentiable functions and global sublinear convergence rates for general convex functions. The efficacy of this framework is demonstrated through comparisons with state-of-the-art algorithms on distributed optimization and learning problems. The last part of the thesis focuses on developing methods for clean and renewable energy adoption in power systems. We develop a distributed controller for secondary control in microgrids with grid-forming inverter-based resources (GFM IBRs). The controller uses distributed optimization, enabling decentralized measurements and neighborhood information exchange to achieve voltage regulation and reactive power sharing. Additionally, a framework for distributed detection and isolation of maliciously behaving agents is proposed for resilient power apportioning between distributed energy resources (DERs). To address challenges posed by the absence of power grids in catastrophic events, this thesis introduces a net-load management engine (horizon of viability (HoV) engine) that ensures reliable power supply to critical infrastructure over a given time-horizon by generating cost-optimal portfolios of local generation sources and loads using mixed-integer convex programming. Controller-hardware-in-the-loop (CHIL) platforms validate the proposed secondary controller, the resilient power apportioning protocol, and the HoV engine across diverse DERs and loads, demonstrating the robustness of the developed methods.