Combining finite element and reinforcement learning methods to design superconducting coils of saturated iron-core superconducting fault current limiter in the DC power system

Abstract

A saturated iron-core type superconducting fault current limiter (SI-SFCL) can effectively restrict the magnitude of the fault current and alleviate the strain on circuit breakers in DC power systems. Design of a superconducting coil (SC), which is one of the key tasks in the SI-SFCL design, requires guaranteeing a sufficient magnetic field, ensuring optimization of the shape and size, minimizing the wire cost, and satisfying the safety and stability of operation. Generally, finite element method (FEM) is used to calculate and evaluate the operating characteristics of SCs, from which it is possible to determine their optimal design parameters. When the coil is complex and large, the simulation time may range from hours to days, and if input parameters change even slightly, the simulations have to be redone from scratch. Recent advances in deep learning represent the ability to be effective for modeling and optimizing complex problems from training data or in real-time. In this paper, we presented a combination of the FEM simulation and deep Q-network (DQN) algorithm to optimize the SC design of a lab-scale SI-SFCL for a DC power system. The detailed design process and options for the SC of SI-SFCL were proposed. In order to analyze the characteristics related to the electromagnetic properties and operational features of the SC, a 3D FEM model was developed. Then, a DQN model was constructed and integrated with the FEM simulation for training and optimizing the design parameters of the SC in real-time. The obtained results of this study have the potential to effectively optimize the design parameters of large-scale SI-SFCL development for high-voltage DC power systems.