Resilient and Stability-Aware Operation of Modern Power Systems

QC 20260925

Abstract

Modern power systems are undergoing a profound transition, with the growing integration of distributed energy resources (DERs) and converter-interfaced devices shifting grid operation from traditional centralized synchronous generation toward a mode dominated by renewable energy and power electronics. This transition enhances system flexibility but also introduces new challenges, requiring higher levels of real-time system awareness, resilience against extreme events, and dynamic stability in converter-dominated power systems.   

In terms of system awareness, existing identification methods largely rely on synchronized phasor measurements or assume known network parameters, and are mostly validated on balanced systems, making them difficult to adapt to the dynamic variation of three-phase unbalanced distribution networks in actual operation. For resilience enhancement, existing methods typically treat safety constraints as soft penalty terms, offering limited guarantees for restoration decisions following extreme events. For stability analysis for converter-dominated systems, existing methods mostly remain at the offline evaluation stage and are not integrated with actual operational decision-making.   

To address these challenges, building on digital twin (DT) driven accurate system awareness, this thesis develops methods that embed safety constraints as hard constraints within the learning and optimization based decision-making process, to enhance the resilience of unbalanced distribution grids against extreme events and the dynamic stability of converter-dominated power systems, and extends this research as an application to railway traction power supply systems with coordinated wayside energy storage system (ESS) energy management.   

First, for three-phase unbalanced distribution networks, this thesis proposes a DT-integrated identification method that combines a data-driven regression step with a model-based Newton-Raphson iteration, enabling the system to continuously identify the real-time network topology and parameters during operation. Based on the accurate and continuously updated DT model, a DT-integrated safe DRL-based critical load restoration method is further proposed, in which a projection layer is introduced into the decision-making process to enforce safety constraints as hard constraints throughout the learning process.   

To further address stability challenges in converter-dominated power systems, this thesis proposes a stability-aware safe DRL method. A double deep neural network (DNN)-based converter-grid interaction stability (CIS) region is constructed and embedded into the decision constraints of reinforcement learning, so that interaction stability is maintained while pursuing control objectives such as frequency support.   As an extended application of this research, this thesis extends this stability-aware perspective to railway traction power supply systems (RPSSs), where train operation, wayside ESS, and NN-accelerated power flow are jointly optimized. A stability-constrained joint optimization method is proposed, which embeds a Railway-ESS Coordinated (REC) stability projection to ensure the resulting dispatch remains dynamically stable.   

The proposed methods are validated respectively on modified IEEE test systems for three-phase unbalanced distribution networks, hardware validation for converter-dominated power systems, and real-world Swedish railway traction power supply systems, demonstrating the effectiveness and practical value.

Link to DiVA