[Paper Review] Deep Learning for Power System Security Assessment

José-María Hidalgo-Arteaga, Fiodar Hancharou+2 more|arXiv (Cornell University)|Mar 31, 2019

Power System Optimization and Stability19 references4 citations

TL;DR

This paper proposes a novel deep learning framework that represents power system snapshots as 2D images to enable fast and accurate security assessment using convolutional neural networks (CNNs). It achieves over 255x speedup compared to traditional small-signal stability analysis while maintaining 99.14% recall and 98% accuracy in classifying N-1 and small-signal stability violations on the 162-bus NESTA system.

ABSTRACT

Security assessment is among the most fundamental functions of power system operator. The sheer complexity of power systems exceeding a few buses, however, makes it an extremely computationally demanding task. The emergence of deep learning methods that are able to handle immense amounts of data, and infer valuable information appears as a promising alternative. This paper has two main contributions. First, inspired by the remarkable performance of convolutional neural networks for image processing, we represent for the first time power system snapshots as 2-dimensional images, thus taking advantage of the wide range of deep learning methods available for image processing. Second, we train deep neural networks on a large database for the NESTA 162-bus system to assess both N-1 security and small-signal stability. We find that our approach is over 255 times faster than a standard small-signal stability assessment, and it can correctly determine unsafe points with over 99% accuracy.

Motivation & Objective

• To address the computational intractability of security assessment in large power systems with millions of operating points.
• To explore the potential of deep learning, particularly CNNs, for accelerating power system security assessment.
• To develop a novel 2D image representation of power system snapshots that preserves system topology and power flow characteristics.
• To train and evaluate deep neural networks on a large-scale database of 1M+ operating points for N-1 and small-signal stability assessment.
• To demonstrate the feasibility of integrating deep learning models into security-constrained optimal power flow frameworks.

Proposed method

• Power system snapshots are converted into 2D images using three channels: active power (P), reactive power (Q), and voltage magnitude (V), mapped to spatial coordinates based on bus and line topology.
• The image representation preserves the network structure by assigning each bus and line to a specific grid location, enabling spatial pattern learning via CNNs.
• A deep convolutional neural network (ConvNet) is trained on a large database of 1,000,000 operating points generated using an efficient algorithm for the NESTA 162-bus system.
• The model is trained to classify each operating point as secure or insecure based on N-1 contingency criteria and small-signal stability (damping ratio ≥ 3%).
• Model performance is evaluated using metrics including accuracy, F1-score, Matthews Correlation Coefficient (MCC), and recall, with k-means clustering used to analyze misclassified regions.
• The framework is benchmarked against standard small-signal stability analysis in terms of computation time on a single-core CPU.

Experimental results

Research questions

• RQ1 Can power system snapshots be effectively represented as 2D images to enable application of pre-trained deep learning models from computer vision?
• RQ2 How accurately can a deep neural network classify N-1 and small-signal stability violations using image-based representations of power system data?
• RQ3 What is the computational speedup of the proposed deep learning model compared to conventional small-signal stability assessment?
• RQ4 Which regions of the security boundary are most challenging for the model to classify correctly, and why?
• RQ5 Can the deep learning model be integrated into a security-constrained optimal power flow framework to improve computational efficiency?

Key findings

• The proposed image-based representation enables effective transfer of deep learning techniques from computer vision to power system security assessment.
• The deep neural network achieves a classification accuracy of over 98% and a recall of 99.14% for unsafe operating points, indicating strong performance in identifying critical stability violations.
• The model reduces computation time to 86 ms per operating point, representing a 255-fold speedup compared to standard small-signal stability analysis (21,950 ms).
• A significant portion of misclassified points (over 99%) are concentrated in a single, small region of the security boundary, suggesting localized modeling challenges.
• The model maintains high performance across multiple evaluation cases, with the best results observed in balanced datasets and moderate class imbalance.
• Preliminary integration into an optimal power flow framework shows an average computation time of 19.99 seconds over 50 tests, indicating feasibility for future optimization integration.