Machine-Learning-Reinforced Massively Parallel Transient Simulation for Large-Scale Renewable-Energy-Integrated Power Systems

Renewable energy systems (RESs) are pivotal in the transition to eco-friendly smart grids. The complexity and uncertainty of RESs, driven by uncontrollable natural forces like sunlight and wind, bring challenges to integrating RESs into modern power systems. Electromagnetic transient (EMT) simulation is an effective method for studying the integration of RESs. Currently, the EMT simulation of RESs is limited to small-scale and lumped RES models due to the model complexity and nonlinearity, which cannot reflect the detailed characteristics of large-scale RESs in practice. This paper introduces a data-oriented, machine learning-enhanced approach to achieve massively parallel EMT simulation on CPU-GPU, designed to efficiently model and simulate large-scale, detailed RES. It incorporates data-driven machine learning modeling of RES via artificial neural networks and integrates these models using a data-oriented entity-component-system framework. The model training was based on reliable model data produced by traditional physical EMT models and the results were validated with MATLAB/Simulink. The RES components are grouped into a microgrid connected to a synthetic AC/DC system based on the IEEE 118-Bus system, achieving an acceleration performance of 400 times faster than traditional CPU nonlinear iterative computations with 2 million RES entities.