Research

This is the physics and criteria layer. Dynamic security is decided with models and margins, not with a black-box label of stable or unstable. I work on rotor-angle transient stability in two settings. In conventional multimachine networks I form a model-based transient stability index from synchronizing-torque contributions among generators, interpreted through electromechanical oscillation modes, so that a security margin can be read in torque terms and checked against CCT. In converter-connected systems I use Lyapunov-based transient energy functions for direct assessment of grid-forming (GFM) and grid-following (GFL) voltage source converters (VSCs), including CCT of the GFM unit when a GFL converter is in parallel. A review of analysis and enhancement methods for renewable-rich, low-inertia grids sits beside that work, because GFL and GFM pose different transient-stability problems. These criteria are what later AI models are asked to respect, and what a digital twin is asked to monitor.

Time-domain simulation remains the reference for TSA and CCT, but it is expensive when many post-fault trajectories or clearing times must be screened. This part asks whether scientific AI can speed that assessment without treating the grid as a black box. The main result is a physics-informed neural network (PINN) that predicts rotor-angle trajectories on a single-machine infinite-bus (SMIB) benchmark, conditioned on electrical power, initial conditions, and machine parameters. The same trained checkpoint is used for nominal-clearing TSA and for surrogate-assisted CCT search. Trajectory accuracy improves over a width-matched standard neural network, but better trajectories do not by themselves give protection-grade CCT; the surrogate is aimed at planning-oriented screening and comparative margin studies. A technical briefing on machine learning for TSA places that work among classification versus regression models, PINNs, and graph-based methods, and states the limits that still matter: scarce data, generalisation, and interpretability. The physics comes from pillar 01. Fast trajectories are what pillar 03 would run online.

Dynamic security is an operating-point question: given how the grid is running now, is the post-disturbance margin still adequate? This part is the operational setting I am building toward. A grid digital twin, here, is not a three-dimensional replica of substations. It is a living model of the dynamics that can be refreshed with measurements and used to screen contingencies and track security margins faster than repeating full time-domain simulation for every case. Pillar 01 defines the quantities to monitor: synchronism, energy-function margins, and CCT. Pillar 02 supplies a fast trajectory surrogate that can be queried inside that loop. The published energy-function and PINN studies already point at operational planning and contingency management; the twin is how those engines would sit in an online or near-online workflow.