A Dynamic Recurrent Adjacency Memory Network for Mixed-Generation Power System Stability Forecasting

TL;DR

The paper tackles real-time stability forecasting in power systems with high inverter-based resource penetration. It introduces DRAMN, a dynamic recurrent adjacency memory network that fuses sliding-window dynamic mode decomposition with graph-convolutional gating inside an LSTM to jointly model evolving spatial interactions and temporal dynamics. The approach yields a multi-layer dynamic adjacency tensor, achieves state-of-the-art accuracy (up to ~0.999 AUROC) with significant feature reduction (≈82%), and demonstrates strong generalization across 9-bus, HVDC, and 39-bus systems. Its real-time performance and interpretability hold practical promise for deployment in modern control centers, though scalability and topology-change robustness remain areas for future work.

Abstract

Modern power systems with high penetration of inverter-based resources exhibit complex dynamic behaviors that challenge the scalability and generalizability of traditional stability assessment methods. This paper presents a dynamic recurrent adjacency memory network (DRAMN) that combines physics-informed analysis with deep learning for real-time power system stability forecasting. The framework employs sliding-window dynamic mode decomposition to construct time-varying, multi-layer adjacency matrices from phasor measurement unit and sensor data to capture system dynamics such as modal participation factors, coupling strengths, phase relationships, and spectral energy distributions. As opposed to processing spatial and temporal dependencies separately, DRAMN integrates graph convolution operations directly within recurrent gating mechanisms, enabling simultaneous modeling of evolving dynamics and temporal dependencies. Extensive validations on modified IEEE 9-bus, 39-bus, and a multi-terminal HVDC network demonstrate high performance, achieving 99.85%, 99.90%, and 99.69% average accuracies, respectively, surpassing all tested benchmarks, including classical machine learning algorithms and recent graph-based models. The framework identifies optimal combinations of measurements that reduce feature dimensionality by 82% without performance degradation. Correlation analysis between dominant measurements for small-signal and transient stability events validates generalizability across different stability phenomena. DRAMN achieves state-of-the-art accuracy while providing enhanced interpretability for power system operators, making it suitable for real-time deployment in modern control centers.

Figures (6)

• Figure 1: Overview of the proposed DRAMN framework. Time-series measurements are processed in a sliding window to extract modal features using DMD. These features are mapped to multi-layer dynamic adjacency matrices representing evolving modal interactions. The recurrent network utilizes both node features and dynamic graph structures to forecast small-signal and transient instability probabilities.
• Figure 2: U-shaped HVDC network interconnecting a diverse generation mix composition. The bipolar links 1 and 2 are interconnected by a tie-line. Terminals B of both links are fed by PV generation and BESS. Terminal A of link 1 is connected to a PH station, while terminal A of link 2 is connected to a PH station and to a grid equivalent through a transmission line.
• Figure 3: Voltage trajectories measured at Buses 4, 7, and 9 during a load increase event under the generation mix SG/GFM/GFL = 58/38/4.
• Figure 4: Top 13 dominant nodes of the HVDC system.
• Figure 5: AUROC of DRAMN trained on clean versus noise-augmented data, evaluated on HVDC measurements corrupted with AWGN at varying SNR levels.