Deep Generative Model-Aided Power System Dynamic State Estimation and Reconstruction with Unknown Control Inputs or Data Distributions

TL;DR

This work tackles dynamic state estimation in power systems under unknown control inputs and data distributions by proposing a deep generative model–aided DSE framework. It combines a robust encoder with a GAN–VAE encoder–decoder to jointly estimate generator states and unknown controls, and employs a latent diffusion model (LDM) for latent-space reconstruction and data imputation, aided by a lightweight adaptor for fast adaptation to unforeseen events. The approach enables two-phase latent-space recovery, anomaly detection, and centralized DSE with reduced communication bandwidth, demonstrated on IEEE 39-bus and NPCC 140-bus systems where it outperforms UKF-VAR, VAE, and TSDM in accuracy and robustness. While offering significant improvements in resilience and efficiency, the method introduces higher computational complexity, highlighting the need for further optimization for real-time deployment.

Abstract

Fast and robust dynamic state estimation (DSE) is essential for accurately capturing the internal dynamic processes of power systems, and it serves as the foundation for reliably implementing real-time dynamic modeling, monitoring, and control applications. Nonetheless, on one hand, traditional DSE methods based on Kalman filtering or particle filtering have high accuracy requirements for system parameters, control inputs, phasor measurement unit (PMU) data, and centralized DSE communication. Consequently, these methods often face accuracy bottlenecks when dealing with structural or system process errors, unknown control vectors, PMU anomalies, and communication contingencies. On the other hand, deep learning-aided DSE, while parameter-free, often suffers from generalization issues under unforeseen operating conditions. To address these challenges, this paper proposes an effective approach that leverages deep generative models from AI-generated content (AIGC) to assist DSE. The proposed approach employs an encoder-decoder architecture to estimate unknown control input variables, a robust encoder to mitigate the impact of bad PMU data, and latent diffusion model to address communication issues in centralized DSE. Additionally, a lightweight adaptor is designed to quickly adjust the latent vector distribution. Extensive experimental results on the IEEE 39-bus system and the NPCC 140-bus system demonstrate the effectiveness and superiority of the proposed method in addressing DSE modeling imperfection, measurement uncertainties, communication contingencies, and unknown distribution challenges, while also proving its ability to reduce data storage and communication resource requirements.

Figures (8)

• Figure 1: The position and role of decentralized DSE in power systems, and the composition and control diagram of a 6th-order generator.
• Figure 2: The control block diagram of the utilized IEEEX1 exciter, TGOV1/TGOV1D governor, and IEEEST stabilizer in this paper, with parameters' definitions described in cui2020hybrid.
• Figure 3: The enhanced two-phase latent diffusion model for dynamic state estimation, composed of a robust encoder $E_{\bm{\phi}'}(\cdot)$, a TSDM model trained in the latent space $D_{\bm{\theta}}(\cdot, \cdot)$, and the generator $G_{\bm{\psi}}(\cdot)$ of GAN, which aims to mitigate the impact of communication uncertainty in centralized DSE.
• Figure 4: The deployment diagram of the deep generative model-aided data-driven DSE, which integrates both centralized and decentralized approaches in practical power systems. This can be used for local, accurate, and robust DSE on the power plant side, as well as monitoring in estimation centers, ensuring the stable and secure operation of both local and global control systems.
• Figure 5: Joint state and control variable estimation results for Generator 1 during a short-circuit fault event in the IEEE 39-bus system.