Multivariate Time-series Diffusion Model-based Generation of Transient Trajectories for Power System Applications

Post-contingency transient trajectories are usually acquired via simulation and/or real-world measurements. However, time domain simulation is very time-consuming especially for electromagnetic transient (EMT). For real-world system, simulation models may be unavailable, and contingencies do not occur very frequently. Both situations may lead to insufficient data, which can be a major barrier to data-driven dynamic modeling applications. In the meanwhile, generative artificial intelligence has attained remarkable achievements in recent years, providing a powerful tool for data analytics. This paper develops a transient trajectory generation framework to tackle the data insufficiency challenge. The generation process is based on the diffusion model, which adds random noise to the original data in the forward process, and then denoises in the reverse process to construct desired data samples. A neural network based on structured state-space sequence model is utilized to capture the time dependency and reproduce multi-variable trajectories considering the simultaneous evolution and correlation of multiple states. A conditioner is thus supplemented in the neural network to guide the trajectory generation with desired dynamic behaviors. The algorithm is trained with simulated data from a hybrid EMT-phasor simulation platform, which consists of a 39-bus transmission system and a modified IEEE 33-node test feeder. The effectiveness of the proposed algorithm in generating trajectories with explicit and diverse dynamic features is verified.