Abstract

For system-level studies, manufacturers commonly provide black-box models. However, their limited transparency regarding control structures and parameterization can impede model validation, interoperability, and stability assessment. They may also be difficult to integrate into simulations, computationally demanding, and poorly suited to analytical methods. Conversely, generic models calibrated solely through parameter optimization often provide only limited fidelity outside the operating conditions represented in the available data. The Dynamic Power System Open Solutions for Hybrid Dynamic Equivalents (DyPSOS) project aims to replace black-box models with accurate, analyzable and easy-to-integrate dynamic equivalents. We pursue a hybrid, data-driven approach, combining automatic parameter optimization of known model components with Neural Ordinary Differential Equation (NODE) and Universal Differential Equation (UDE) methods. We use the open-source Julia SciML ecosystem, together with symbolic modeling (ModelingToolkit.jl) and the power grid simulation framework PowerDynamics.jl, to train and implement the models. The goal is to develop an optimization toolbox for dynamic equivalents with high accuracy that base upon known physics and can be integrated into existing differential-algebraic equation (DAE) systems.

About the speaker

Lars Bittcher is a researcher in the Power System Control and Dynamics Division of the System Stability and Grid Integration Department at the Fraunhofer Institute for Energy Economics and Energy System Technology (IEE). His research focuses on hybrid data-driven modeling of power system dynamics, combining physics-based models with machine learning approaches such as Neural Ordinary Differential Equations (NODEs) and Universal Differential Equations (UDEs).