PaperlayerView exact revision on arXiv

Source-linked AI summary

Data-Optimized Contingency Screening: A Machine Learning Approach to Power System Security

Joshua Salako, Folajimi Osikomaiya, Olakorede Olamiju

arXiv:2609.04300v1cs.AIcs.LG

TL;DR

Power-system contingency classification matters for timely security assessment, but manual methods face scaling and consistency challenges. This study evaluates machine-learning models with preprocessing on contingency data from IEEE bus systems. Random Forest is reported as the most robust model, while preprocessing effects differ across models.

  • Problem

    Manual contingency classification can be time-consuming in large, complex systems and risk inconsistency, while robust detection of safe versus emergency conditions is needed [20].

    Source: §I · p. 1

  • Method

    The study assesses machine-learning models and the effects of SMOTE class balancing and PCA dimensionality reduction on IEEE-14 and IEEE-30 bus system datasets.

    Source: B. Synthetic Minority Over-sampling… · p. 6, CONCLUSION · p. 9

  • Results

    Random Forest was the most robust model across configurations, PCA generally contributed more than SMOTE, and SMOTE improved severe-class recall but could reduce precision through false positives.

    Source: D. Discussion of Models Efficiency · pp. 8–9, CONCLUSION · p. 10

  • Takeaways & Limitations

    The study presents machine-learning models as effective for real-time classification of emergency situations in energy systems.

    Source: CONCLUSION · p. 9

Abstract

from arXiv · show

Ensuring the security of the power system is essential for stability and reliability, especially in the event of disruption. Effective classification of contingency in power systems enables proactive decision-making and mitigates large-scale breakdowns and failures. This study explores the use of machine learning algorithms to classify security levels of contingencies in power systems into safe, moderate or severe classes. For this approach, Newton-Raphson load flow method extracts system data from contingency scenarios, using Overall Performance Index (OPI) as safety measure. For data pre-processing, Synthetic Minority Over-Sampling Technique (SMOTE) and Principal Component Analysis (PCA) is used to address class imbalance and reduce dimensionality, respectively. K-Nearest Neighbours (KNN), Random Forest (RF) and Support Vector Machines (SVM) is trained and evaluated on datasets generated through N-k contingency scenarios for k equal 1, 2, and 3 on IEEE-14 and IEEE-30 bus systems using four hybrid pre-processing configurations: normalized, SMOTE-balanced, PCA-transformed, and a combined SMOTE PCA-transformed. Performance is assessed by precision, recall and F1 score, with priority given to the severe contingency classes. The RF achieved the highest F1 scores of 0.97 in IEEE-30 and 0.86 in IEEE-14, SVM benefits significantly from PCA and improves the accuracy of the classification, while KNN is best suited for SMOTE and PCA conversion. The findings show that PCA contributes more than SMOTE to the overall performance of the model. However, SMOTE improves recall but can introduce false positives and is therefore a compromise of accuracy. This study highlights machine learning as a scalable and powerful alternative to traditional contingency analysis, which improves the assessment of security in real time.

I. INTRODUCTION

Contingencies threaten power-system reliability, while manual classification is difficult to scale for timely, comprehensive assessment. The study investigates machine-learning classification, against a background of work integrating security indices with machine learning for contingency analysis.

I. INTRODUCTION

  • The introduction frames robust detection of safe versus emergency operating conditions as an urgent need [20].The discussion notes that power-system contingencies may result from equipment failures, faults, natural disasters, human error, or deliberate attacks.Source: §I · p. 1
  • Manual contingency assessment is time-consuming in large, complex systems and risks inconsistent or biased classifications, motivating automated and timely decision support [5][6].Contingencies can lead to blackouts, voltage instability, cascading failures, and wider social and economic consequences.Source: §I · p. 1
  • The study investigates Random Forest, Support Vector Machines, and K-Nearest Neighbours for power-system contingency classification.The paper presents these models as an effort to overcome drawbacks of manual methods.Source: §I · p. 2

II. RELATED WORKS

  • Prior work describes integrating security indices with machine learning as a practical approach to real-time contingency analysis.The passage characterizes traditional contingency selection and ranking based on overall performance indices as computationally complex for real-time applications.Source: §II · p. 2

A. Contingency Analysis of Power Systems

The section outlines contingency analysis and a load-flow-based screening process that evaluates outages using power-system performance indices. Newton-Raphson load flow supplies system quantities used to assess and rank contingency severity.

A. Contingency Analysis of Power Systems

  • Contingency analysis defines potential events, selects a subset for further review, and evaluates their security implications.Evaluation can use power-flow analysis or dynamic simulation and compute measures such as voltage fluctuations, power-flow losses, and stability margins.Source: A. Contingency Analysis of Power Systems · p. 2

B. Load Flow Analysis

  • Newton-Raphson is an iterative load-flow technique with quadratic convergence, suited to large systems but requiring function evaluation and computation at each iteration [38].The passage notes that most power-flow problems converge in fewer than ten iterations.Source: B. Load Flow Analysis · p. 3
  • The active-power performance index assesses line congestion by identifying active-power flows that exceed line capacity.Its stated purpose is to quantify potential line congestion and help identify contingencies with excessive active-power flow.Source: B. Load Flow Analysis · p. 3
  • The voltage performance index identifies bus-voltage-limit violations and provides insight into system voltage stability.The passage notes that contingency voltage analysis must account for generator reactive-power constraints when reactive power approaches its limits.Source: B. Load Flow Analysis · p. 4
  • OPI combines the active-power and voltage performance indices to assess security across contingency scenarios.The study uses Newton-Raphson load flow to extract system parameters for contingency scenarios and uses OPI to inform security assessment and prioritization.Source: B. Load Flow Analysis · p. 4
  • The screening process calculates line flows and bus voltages, evaluates active-power and voltage indices, combines them into OPI, and ranks contingencies by OPI.The described flow runs load flow for baseline conditions and again after simulating an outage.Source: B. Load Flow Analysis · p. 4

C. Random Forests

Random Forest combines decision trees into an ensemble classification. The section presents bagging, feature selection, and majority voting as elements of its approach.

C. Random Forests

  • Random Forest trains multiple decision trees and aggregates their outputs by majority vote for classification [31].The trees are trained on different bootstrap samples of the original dataset.Source: C. Random Forests · pp. 4–5
  • Bootstrap aggregation and random feature selection are described as ways Random Forest reduces overfitting and improves generalization [31].The method uses subsets of features across trees, supporting diversity in the ensemble.Source: C. Random Forests · pp. 4–5
  • The paper applies Random Forest to prioritize contingency scenarios, identify high-risk failures, and assess network stability under different operating conditions.The section also describes Random Forest as suited to high-dimensional data and complex feature interactions.Source: C. Random Forests · p. 5

D. Support Vector Machines

Support Vector Machines classify by finding a margin-maximizing hyperplane between classes. The section also describes kernel functions for nonlinear data and notes the relevance of PCA to classification performance.

D. Support Vector Machines

  • SVM seeks a hyperplane that maximizes the margin between classes and can address linear and nonlinear classification problems using kernels [26].The margin is maximized subject to constraints described in the section.Source: D. Support Vector Machines · p. 5
  • For nonlinear data, SVM uses kernel functions to transform data into a higher-dimensional space where linear separation is possible.The section identifies the Radial Basis Function kernel and describes its hyperparameter as controlling the influence of individual training samples.Source: D. Support Vector Machines · p. 5
  • Prior work reports SVM use for contingency classification and high accuracy particularly when combined with feature-selection techniques such as PCA.Source: D. Support Vector Machines · p. 5

E. K-Nearest Neighbors

KNN is presented as a non-parametric classifier that predicts from the data rather than assuming an underlying distribution. Its class determination uses nearest neighbors and prior probabilities.

E. K-Nearest Neighbors

  • KNN is non-parametric and makes predictions from the data without assuming an underlying data distribution.Its predictions are based on the data itself.Source: E. K-Nearest Neighbors · p. 5
  • KNN uses the set of nearest neighbors to determine a query instance’s class.Source: E. K-Nearest Neighbors · p. 5
  • Class determination relies solely on prior probabilities, without considering the class distribution near the query point.Source: E. K-Nearest Neighbors · p. 5

III. DATA PREPROCESSING AND MODELLING

The study frames contingency prediction as an imbalanced three-class problem and applies normalization to its input and output parameters. Normalization scales values to 0.1–0.9 and is described as especially important for KNN.

III. DATA PREPROCESSING AND MODELLING

  • Contingencies are framed as a safe, moderate, or severe classification problem, with preprocessing used to address class imbalance and support fair training.Source: §III · p. 5

A. Data Normalization

  • Normalization scales data to the range 0.1 to 0.9 while preserving relative relationships between elements.Source: A. Data Normalization · p. 5
  • Normalization is particularly important for KNN because it makes parameters contribute proportionally to similarity measurement.The passage also says normalization benefits other algorithms by improving numerical stability and convergence.Source: A. Data Normalization · p. 5

B. Synthetic Minority Over-sampling Technique

The preparation workflow uses SMOTE to address underrepresentation of severe contingencies and PCA to reduce dimensionality. It generates and labels contingency data for IEEE-14 and IEEE-30 systems, then evaluates preprocessing configurations.

B. Synthetic Minority Over-sampling Technique

  • SMOTE addresses severe-class underrepresentation by generating synthetic minority-class samples along line segments between an instance and its nearest neighbors.The interpolation uses a random value in [0, 1] to place the sample along the connecting vector.Source: B. Synthetic Minority Over-sampling… · p. 6

C. Principal Component Analysis

  • PCA reduces dimensionality and redundant or correlated features by projecting data onto the top k principal components with the highest eigenvalues.The described procedure standardizes features, computes a covariance matrix, and derives eigenvectors and eigenvalues.Source: C. Principal Component Analysis · p. 6

D. Data Preparation for Contingency Classification

  • Contingency scenarios cover N-k events for k=1 to 3 on IEEE-14 and IEEE-30 systems, including line, generator, and load disturbances.Extracted features include bus voltages and angles, line loading, power flows, and generator outputs; OPI measures system security.Source: D. Data Preparation for Contingency… · p. 6
  • Normalized OPI assigns contingencies to Safe (≤0.2), Moderate (0.2–0.5), or Severe (>0.5), and is removed from model inputs to prevent data leakage.Source: D. Data Preparation for Contingency… · p. 6
  • Models are evaluated across normalized, SMOTE-balanced, PCA-transformed, and combined SMOTE-plus-PCA dataset configurations.The configurations test the effects of class balancing and dimensionality reduction.Source: D. Data Preparation for Contingency… · p. 7

E. Model Training and Evaluation

KNN, RF, and SVM are trained and evaluated across four preprocessing configurations, with hyperparameters selected by random search. Evaluation emphasizes precision, recall, and F1-score for severe contingencies because class imbalance can make accuracy misleading.

E. Model Training and Evaluation

IV. RESULTS AND FINDINGS

The section compares severe-contingency classification across models and preprocessing configurations on the IEEE-14 bus dataset. RF performed best with normalized or PCA-transformed data, while SVM and KNN showed different responses to preprocessing.

B. Model performance on IEEE-14 Bus Dataset

  • 0.86 F1-score was RF’s result on normalized and PCA data; SMOTE lowered it to 0.57 and SMOTE + PCA to 0.33.The reported combined-preprocessing result was the worst RF performance, attributed to excessive variance and poorer classification reliability.Source: B. Model performance on IEEE-14 Bus… · pp. 7–8, Table II · p. 8
  • 0.67 F1-score was achieved by SVM with PCA and SMOTE + PCA, after zero severe cases were correctly classified with normalized or SMOTE-balanced data.The passage attributes the improvement to feature reduction and reports no additional advantage from SMOTE.Source: B. Model performance on IEEE-14 Bus… · p. 8
  • F1-scores ranged from 0.57 to 0.75, with KNN best on normalized data; SMOTE and PCA did not significantly improve it.The passage describes KNN’s performance as relatively stable across configurations.Source: B. Model performance on IEEE-14 Bus… · p. 8

C. Model Performance on IEEE-30 Bus Dataset

On the IEEE-30 bus dataset, the section compares severe-contingency performance across preprocessing configurations. RF led overall, while PCA and combined preprocessing benefited SVM and KNN, respectively.

C. Model Performance on IEEE-30 Bus Dataset

  • 0.97 F1-score was achieved by RF on normalized and PCA-transformed data, compared with 0.95 on SMOTE and SMOTE + PCA data.The passage says SMOTE slightly reduced F1-score while improving recall, potentially introducing false positives and reducing precision.Source: C. Model Performance on IEEE-30 Bus… · p. 8
  • 0.89 F1-score was SVM’s normalized-data result; SMOTE improved performance, PCA increased it further, and SMOTE + PCA matched PCA.The passage suggests SVM benefits more from feature selection than from class balancing.Source: C. Model Performance on IEEE-30 Bus… · p. 8
  • 0.81 F1-score was KNN’s normalized-data result, rising to 0.83 with PCA; SMOTE improved F1-score, and SMOTE + PCA performed best.The passage reports that KNN had relatively lower performance than RF and SVM, particularly in severe-class recall.Source: C. Model Performance on IEEE-30 Bus… · p. 8

D. Discussion of Models Efficiency

Across the IEEE-14 and IEEE-30 systems, Random Forest was the most robust classifier, while the effects of preprocessing differed by model. The study presents Newton-Raphson load flow combined with machine learning as a framework for real-time contingency safety assessment.

D. Discussion of Models Efficiency

  • RF was the most robust model across IEEE-14 and IEEE-30, with confusion matrices showing strong correct classification and minimal misclassifications.The IEEE-14 matrix reports 761 correct Safe, 17 Moderate, and 3 Severe classifications; the IEEE-30 matrix has a dominant diagonal.Source: D. Discussion of Models Efficiency · pp. 8–9
  • 0.67 Severe-class F1 followed PCA reduction from 81 to 15 features for IEEE-14 SVM, which classified no Severe cases on normalized data.The passage reports only 14 Severe instances among 3,136 total instances in that system.Source: D. Discussion of Models Efficiency · p. 9
  • KNN performed best on normalized IEEE-14 data and on combined SMOTE-and-PCA IEEE-30 data, but generally underperformed RF and SVM.Its distance-based approach is sensitive to feature scaling and class distribution.Source: D. Discussion of Models Efficiency · p. 9
  • PCA contributed more than SMOTE to accuracy for most models; SMOTE improved Severe-class recall but could reduce precision through false positives.The combined SMOTE-and-PCA configuration helped KNN most but offered little or no benefit to RF and SVM.Source: D. Discussion of Models Efficiency · p. 9
  • IEEE-30 showed higher classification performance across models than IEEE-14, while RF remained the most reliable and SVM benefited from feature selection.The authors suggest the larger sample space and broader contingency scenarios may explain the system-level difference.Source: D. Discussion of Models Efficiency · p. 9
  • The study presents Newton-Raphson load flow integrated with machine learning as a framework for rapid, accurate real-time power-system safety assessment.Its analysis covers simulated IEEE-14 and IEEE-30 bus systems.Source: CONCLUSION · pp. 9–10
  • Reliance on simulated IEEE bus systems limits practical significance because these idealized models may not capture real-grid complexities such as dynamic load variation or equipment degradation.The authors call for validation using real-world grid data.Source: CONCLUSION · p. 10
Loading 2609.04300v1…