Power System Simulator based on Physics-Informed Neural Networks

This proposal puts forward a first-of-its-kind power system dynamics simulator based on neural networks: PINNSim. PINNSim aims to deliver a highly modular and scalable tool that can be up to 100x faster than any commercial power system simulator. This means that PINNSim can perform a power system stability assessment in 30 minutes instead of 2 days. We witness a dramatic shift from few large scale generators to millions of converter-based resources, such as solar, wind, batteries, and electric vehicles. Capturing the dynamic behavior of such a system is no longer possible with the conventional approximations. In case of fault, phenomena propagate much faster and an ElectroMagnetic Transient (EMT) simulation is necessary to accurately capture them. However, EMT is slow and computationally intensive. Uncertainty from solar and wind and the increased number of components requires the assessment of millions of critical scenarios so that power system operators eliminate blackout risk, and manufacturers design appropriate controllers for the safe operation of their equipment.PINNSim leverages physics-informed machine learning and GPU acceleration to model power system components. Its key novelty lies in the method it interconnects individual neural networks to form a “system of neural networks” that performs simulations at much higher speeds. Through this funding, we aim to scale PINNSim to handle larger, more complex power systems and incorporate a user interface, creating a user-friendly simulator capable of addressing a wide range of power system operation challenges.Our goal is to create a tool that will disrupt the power system simulation software landscape. We envision a library of component models based on neural networks, which can then be used in a plug’n’play fashion in PINNSim, similar to conventional simulators. If successful, PINNSim can naturally extend to almost any non-linear dynamic system such as robotics and biological systems.