AI-driven power system transient simulation and angular stabilisation
Author(s)
Wei, Yunhe
Type
Thesis or dissertation
Abstract
Fast transient corrective control, within seconds after fault clearance, improves both the security and economic operation of modern power systems. In this narrow post fault window, the first rotor angle swing and coupled frequency and voltage dynamics evolve rapidly, so timely interventions can prevent loss of synchronism, suppress oscillations, and reduce unnecessary load shedding and equipment stress. Delivering these benefits in real time requires rapid and credible transient prediction, along with low latency decision making and recommendations that operators can understand and trust.
This paper develops an integrated capability that couples accelerated transient simulation with scenario aware corrective control. The prediction layer combines physical models with data driven surrogate models to approximate transient trajectories across diverse operating points, fault locations, and clearance times while preserving key physical behaviours. Physics informed training embeds system dynamics and network constraints during learning, improving reliability under unseen contingencies. Sequence modelling techniques capture both fast transients and longer dependencies with stable learning and modest memory demand, enabling rapid stability screening over many candidate scenarios. Reliability is assessed through validation on benchmark test systems against high fidelity simulation.
Building on these predictions, the control layer restores stability by reshaping active power flows through emergency demand response and optimal load shedding under voltage and frequency limits. A learning based controller evaluates candidate actions in a scenario dependent manner, balancing security and cost. Safety is enforced through action filtering that excludes infeasible or redundant actions and through penalties that discourage constraint violations. To support operator trust, explanatory tools attribute predicted outcomes to influential state variables and action components, producing transparent action rankings. Robustness to uncertainty is achieved by training and testing across scenario distributions and updating action values online as new measurements arrive.
This paper develops an integrated capability that couples accelerated transient simulation with scenario aware corrective control. The prediction layer combines physical models with data driven surrogate models to approximate transient trajectories across diverse operating points, fault locations, and clearance times while preserving key physical behaviours. Physics informed training embeds system dynamics and network constraints during learning, improving reliability under unseen contingencies. Sequence modelling techniques capture both fast transients and longer dependencies with stable learning and modest memory demand, enabling rapid stability screening over many candidate scenarios. Reliability is assessed through validation on benchmark test systems against high fidelity simulation.
Building on these predictions, the control layer restores stability by reshaping active power flows through emergency demand response and optimal load shedding under voltage and frequency limits. A learning based controller evaluates candidate actions in a scenario dependent manner, balancing security and cost. Safety is enforced through action filtering that excludes infeasible or redundant actions and through penalties that discourage constraint violations. To support operator trust, explanatory tools attribute predicted outcomes to influential state variables and action components, producing transparent action rankings. Robustness to uncertainty is achieved by training and testing across scenario distributions and updating action values online as new measurements arrive.
Date Issued
2025-10-01
Date Awarded
2026-02-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Strbac, Goran
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)