An Accessible PyTorch Implementation of Automatic Differentiation for Power System Model Parameter Identification and Optimization

Even though realistic dynamic models are a key prerequisite for managing modern power systems, accurately parameterized dynamic models are often not available. Therefore, this paper examines a gradient descent based optimization method for the optimization and identification of dynamic model parameters. The optimization method relies on PyTorch as an automatic differentiation tool for the computation of gradients. PyTorch has previously been used in the context of power systems only in combination with neural networks and here, it is used without them. It can both identify and optimize parameters with respect to a desired system behavior. The paper presents the theoretical background and shows exemplary use-cases. Additionally, the scalability of the method is investigated. The results imply a generic applicability for a range of problems.