Application of Neural Ordinary Differential Equations to Power System Frequency Dynamics

Electric power system control, supervision, and protection actions require an accurate system dynamics model. The swing equation-based modeling approach does not properly capture the dynamics of a converter-dominated future power grid. Therefore, this research investigates the applicability of a computationally efficient and accurate neural ordinary differential equations (NODEs) approach to model and infer critical state information of the power system frequency dynamics. The designed NODEs framework proposes a new neural network architecture that models input dynamics missing from existing NODEs methods. By actively perturbing the system using a known logarithmic chirp signal, the proposed Python-based NODEs framework is trained using only measured system states and inputs without requiring detailed system information. The trained model accurately predicts future states under square and step excitation signals with a goodness of fit 99.7% with an 18 times computational speedup than the detailed model in Simulink. The applicability of NODEs in modeling power system dynamics was validated and is a promising method to model future power grid dynamics integrated with large amounts of heterogeneous distributed energy resources.