Abstract: Continuous monitoring of the spatio-temporal dynamic behavior of critical infrastructure networks, such as the power systems, is a challenging but important task. In particular, accurate and timely prediction of the (electro-mechanical) transient dynamic trajectories of the power grid is necessary for early detection of any instability and prevention of catastrophic failures. Existing approaches for the prediction of dynamic trajectories either rely on the availability of accurate physical models of the system, use computationally expensive time-domain simulations, or are applicable only at local prediction problems (e.g., a single generator). In this paper, we report the application of two broad classes of data-driven learning models -- along with their algorithmic implementation and performance evaluation -- in predicting transient trajectories in power networks using only streaming measurements and the network topology as input. One class of models is based on the Koopman operator theory which allows for capturing the nonlinear dynamic behavior via an infinite-dimensional linear operator. The other class of models is based on the graph convolutional neural networks which are adept at capturing the inherent spatio-temporal correlations within the power network. Transient dynamic datasets for training and testing the models are synthesized by simulating a wide variety of load change events in the IEEE 68-bus system, categorized by the load change magnitudes, as well as by the degree of connectivity and the distance to nearest generator nodes. The results confirm that the proposed predictive models can successfully predict the post-disturbance transient evolution of the system with a high level of accuracy.
Introduction
Modern power grids are increasingly operated close to their stability limits. When disturbances occur, such as sudden load changes or equipment failures, the grid experiences fast electro-mechanical transients. If operators cannot accurately predict how these transient trajectories will evolve over the next few seconds, they risk missing the window to prevent cascading failures.
Traditionally, predicting these dynamics requires time-domain simulations based on highly detailed physical models of the grid's generators and components. These simulations are computationally expensive and rely on parameters that are often unavailable or unreliable. This paper addresses the need for a system-wide, multi-timestep prediction model that operates purely on limited, noisy streaming data (like Phasor Measurement Units, or PMUs) without needing the underlying physics equations.
Data-Driven Modeling Approaches
The authors compare two broad classes of models to capture the grid's nonlinear dynamics: Koopman Operator Theory and Graph Neural Networks.
Koopman Operator Theory (KOT)
This mathematical framework transforms complex, nonlinear system dynamics into a globally linear system within a higher-dimensional "lifted" space. Because finding the exact infinite-dimensional operator is impossible, the study uses two finite-dimensional approximations:
- Robust DMD (Dynamic Mode Decomposition): An analytical, matrix-math approach that is incredibly fast to compute.
- deepDMD: Uses a neural network to learn the optimal observable functions that map the sensor data into the lifted linear space.
Neither KOT method requires knowledge of the grid's topology; they only need a single time-step snapshot of measurements to perform inference.
Spatio-Temporal Graph Convolutional Networks (STGCN)
STGCNs model the grid as a dynamic graph to explicitly capture spatial (topology) and temporal correlations. Standard STGCNs are typically built for single-step predictions. The authors enhanced this architecture with an incremental inference strategy, allowing the model to autoregressively feed its own predictions back into itself. This enables variable-length future predictions (e.g., 1 to 10 seconds) without needing to retrain the model for different time horizons.
Experimental Setup and Datasets
The authors generated a synthetic dataset using the GridSTAGE simulation framework applied to the standard IEEE 68-bus system. They simulated 330 different load change events, varying the magnitude of the disturbance, the degree of node connectivity, and the distance to the nearest generators. To mimic real-world PMU sensors, the data was sampled at 50 Hz and intentionally corrupted with 85 dB Signal-to-Noise Ratio (SNR) noise. The dataset and model code were made publicly available to serve as a reproducible benchmark for the community.
Performance Evaluation
Accuracy and Speed
All models successfully recovered the post-disturbance transient evolution with very low error: Root Mean Square Error (RMSE) < 21 × 10⁻³ Hz and Mean Absolute Percentage Error (MAPE) < 23 × 10⁻⁵ across all locations.
- STGCN achieved the absolute best accuracy (RMSE of 1.13 × 10⁻³ Hz) on low-magnitude load changes, benefiting from a 4-second historical observation window.
- Robust DMD and deepDMD are remarkably fast, requiring only ~0.1 seconds of inference time to predict a full 10-second future window. STGCN was slower, taking between 1.2s and 2.1s depending on the historical observation window used.
- Training Time: Robust DMD trains in about 5 seconds due to its analytical nature, while deepDMD and STGCN require standard neural network training times (10 to 60 minutes).
Impact of Disturbances
For all models, prediction errors roughly doubled when the magnitude of the load disturbance increased, or when the disturbance occurred very close to a generator bus (1-hop away vs. 3-hops away). Interestingly, spatial prediction errors were not randomly distributed; they clustered around specific load and generator buses, suggesting that the spatial error distribution reflects inherent physical bottlenecks of the power network itself.
Conclusion and Future Directions
This study demonstrates that physics-free prediction is highly viable for real-time grid monitoring. Purely data-driven models can capture nonlinear dynamics accurately enough to provide early warnings. However, some limitations remain: STGCN requires a history of observations and can exhibit abrupt erroneous jumps under high-magnitude disturbances, while deepDMD shows wider error confidence intervals immediately following a disturbance.
Future work will explore hybrid models that combine the speed and robustness of Koopman-based linear lifting with the spatial awareness of graph convolutional neural networks. Additionally, the authors plan to investigate distributed learning techniques to scale these models for massive, real-world dynamical networks.
Summary updated