Abstract
As power grids become increasingly digitized, they face growing threats from cyberattacks, such as false data injection, which can trigger cascading failures. Traditional machine learning models often treat sensor data as isolated time series, neglecting the physical topology of the grid. This research introduces a hybrid deep learning architecture that combines Long Short-Term Memory (LSTM) networks with Graph Isomorphism Networks (GIN). By processing both temporal sensor sequences and structural grid embeddings within an autoencoder framework, the model achieves a near-perfect 99.92% F1 score. This approach demonstrates that incorporating spatial awareness is critical for robust anomaly detection in critical infrastructure.
Introduction
The modern power grid is a complex, interconnected web of generators, substations, and consumers. While digitization improves efficiency, it also expands the attack surface for malicious actors. A single localized anomaly—whether caused by physical failure or a cyberattack—can propagate through the system, leading to widespread blackouts.
Existing detection methods face two primary hurdles. First, they often ignore the grid's topology, treating data from different nodes as independent variables. Second, they are prone to overfitting, making them less effective against "stealthy" attacks designed to mimic normal fluctuations. This study addresses these gaps by proposing a model that understands both the timing and the location of suspicious readings.
Methodology: The Hybrid LSTM-GIN Architecture
The core of the proposed solution is a hybrid autoencoder designed to learn the normal operating state of a power grid. The architecture consists of three main components:
- Temporal Encoding: An LSTM encoder processes a sequence of ten timestamps of grid data, including power output and voltage. LSTMs are utilized for their ability to capture long-term dependencies in time-series data.
- Structural Encoding: Simultaneously, a Graph Isomorphism Network (GIN) analyzes the grid's physical graph. Nodes represent components like generators, while edges represent power lines. The GIN generates a unique structural "signature" or embedding for each node based on its neighborhood.
- Integration and Decoding: The temporal features from the LSTM are concatenated with the structural embeddings from the GIN. An LSTM decoder then attempts to reconstruct the original input sequence from this combined representation.
The model is trained to minimize Mean Squared Error (MSE) on normal data. During inference, any reconstruction error exceeding the 0.99 percentile of training errors is flagged as an anomaly.
Experimental Setup and Datasets
The researchers evaluated the model using a simulated 123-node power grid generated via the NATI[P]G framework. The dataset included baseline data representing normal operating conditions and anomaly data containing simulated cyberattacks. Features tracked included real energy levels for generators and inverters, as well as voltage consumption levels for load nodes.
Results and Discussion
The inclusion of structural awareness via the GIN proved to be the decisive factor in model performance. Nodes with higher connectivity are naturally more susceptible to load fluctuations; by incorporating the GIN, the model "learns" to expect these patterns based on a node's position within the network topology.
As shown in the results, the hybrid LSTM-GIN model achieved an F1 score of 99.92%, compared to 97.30% for a standard LSTM-only model. While a 97% accuracy might seem sufficient, the 2.6% improvement in the hybrid model represents a significant reduction in false negatives, which is vital for preventing catastrophic grid failures.
Conclusion
This research highlights the importance of "topology-aware" machine learning for infrastructure security. By combining LSTM's temporal processing with GIN's structural analysis, the hybrid model provides a highly accurate and generalizable solution for detecting grid anomalies. Future work will focus on scaling this architecture to regional grids with thousands of nodes and testing its resilience against a broader range of sophisticated attack vectors.
Summary updated