State Monitoring and Fault Diagnosis of HVDC System via KNN Algorithm with Knowledge Graph: A Practical China Power Grid Case

Qian Chen, Qiang Li, Jiyang Wu, Jingsong He, Chizu Mao, Ziyou Li, Bo YangSustainability 2023 · DOI

TL;DR

This paper introduces a novel fault diagnosis method for High-Voltage Direct Current (HVDC) transmission systems using the K-Nearest Neighbors (KNN) algorithm combined with a knowledge graph (KG). The approach demonstrates high accuracy in identifying four common fault types, offering a practical solution for improving the reliability and intelligence of China's power grid.

What Problem Does This Paper Address?

Modern power systems are increasingly relying on HVDC technology for efficient long-distance power transmission, especially in regions like China with uneven energy distribution. However, HVDC systems are complex and prone to various faults, which can lead to significant economic losses and disruptions if not diagnosed and resolved quickly. Existing fault diagnosis methods have limitations in terms of complexity, cost, and speed, particularly for the small-sample data characteristic of HVDC systems due to their high reliability. This paper aims to address this by developing an accurate and efficient fault diagnosis method that can contribute to the construction of an HVDC system knowledge graph for enhanced operational intelligence.

Methods

The core of this research is the application of the K-Nearest Neighbors (KNN) algorithm for fault diagnosis. The authors propose a method that involves:

  • Data Preprocessing: Normalizing fault data from 15 different signal channels.
  • Feature Engineering: Connecting the normalized data from these channels into a single feature vector for each fault instance.
  • Data Labeling and Splitting: Labeling the fault data into four distinct types and then splitting it into training (80%) and testing (20%) sets.
  • Model Establishment: Building a KNN fault diagnosis model, selecting Euclidean distance as the metric function, and setting the parameter K (number of neighbors) to 7.
  • Training and Testing: Training the KNN model with the training data and then evaluating its performance on the test data.
  • Knowledge Graph Integration: The paper emphasizes that this fault diagnosis capability is a crucial component for building a comprehensive HVDC system knowledge graph, which can store and analyze fault information for future reference and improved decision-making.

The performance of the proposed KNN method is compared against two other common machine learning algorithms: Support Vector Machine (SVM) and Bayesian Classifier (BC).

Datasets

The study utilizes fault data collected from a practical LCC-HVDC transmission project: the Tianshengqiao (Guangxi Province, China)–Guangzhou (Guangdong Province, China) Transmission Project. This project operates at ±500 kV, spans 960 km, and has a rated power of 1800 MW. The fault data used in the paper were monitored over the past three years of the project's operation. The dataset comprises 15 representative signal channels, and the fault data are categorized into four types: AC fault, DC fault, converter valve fault, and commutation failure fault. The paper describes three different test sets (Y1, Y2, Y3) for evaluation, with varying proportions of training and testing data, including using the training data itself as a test set for cross-validation.

Key Insights and Ideas

  • KNN's Suitability for Small-Sample HVDC Data: The paper highlights that HVDC systems, due to their high reliability, present a "small-sample data problem" for fault diagnosis. KNN's effectiveness in handling such datasets, coupled with its speed and simplicity, makes it a strong candidate.
  • Fault Diagnosis as a Core KG Component: The research underscores that accurate fault diagnosis is not just an isolated task but a fundamental building block for constructing a robust HVDC system knowledge graph, enabling more intelligent operation and maintenance.
  • Practical Application and Engineering Value: By using data from a real-world HVDC project, the study provides practical engineering value and a benchmark for future applications in similar scenarios.
  • Comparative Performance: The study provides a direct comparison of KNN against SVM and BC, offering insights into the relative strengths of these algorithms for HVDC fault diagnosis.

Key Results

The KNN algorithm demonstrated superior performance compared to SVM and BC across multiple test sets:

  • Highest Accuracy: The KNN algorithm achieved the highest diagnostic accuracy, reaching up to 100% in some test sets (Y2 and Y3).
  • Overall Accuracy: Across all test sets, KNN consistently outperformed SVM and BC, with an overall accuracy exceeding 80%.
  • Specific Accuracy: For the Y1 test set, KNN achieved 83.3% accuracy, while SVM reached 75% and BC achieved 66.7%.
  • Error Analysis: The study noted that DC faults were more prone to classification errors across all tested algorithms. The confusion matrices provided detailed breakdowns of correct and incorrect classifications for each fault type and algorithm.

Limitations and Future Work

  • Limited Fault Types: The study acknowledges that the dataset only included four types of faults, and future work should aim to incorporate a wider range of faults (e.g., transformer faults, generator faults) to provide a more comprehensive diagnosis.
  • Dataset Size: While KNN is well-suited for small-sample problems, the paper notes that further improvement and enrichment of datasets are needed for even more robust performance.
  • Future Applications: The authors suggest extending this fault diagnosis technology to flexible DC transmission and AC/DC hybrid systems. They also propose exploring more advanced AI technologies and developing better filtering techniques for HVDC systems to mitigate noise and packet loss in long-distance transmission.

Summary updated

Abstract

Based on the four sets of faults data measured in the practical LCC-HVDC transmission project of China Southern Power Grid Tianshengqiao (Guangxi Province, China)–Guangzhou (Guangdong Province, China) HVDC transmission project, a fault diagnosis method based on the K-nearest neighbor (KNN) algorithm is proposed for an HVDC system. This method can effectively and accurately identify four different fault types, aiming to contribute to construction of a future HVDC system knowledge graph (KG). First, function and significance of fault diagnosis for KG are introduced, along with four specific fault scenarios. Then, the fault data are normalized, classified into a training set and a test set, and labeled. Based on this, the KNN fault diagnosis model is established and Euclidean distance (ED) is selected as the metric function of the KNN algorithm. Finally, the training data are conveyed to the model for training and testing, upon which the diagnosis result obtained by the KNN algorithm with a knowledge graph is compared with that of the support vector machine (SVM) algorithm and Bayesian classifier (BC). The simulation results show that the KNN algorithm can achieve the highest diagnosis accuracy, with more than 83.3% diagnostic accuracy under multiple test sets among all three diagnosis methods.
BibTeX
@inproceedings{chen2023state,
  author = {Qian Chen and Qiang Li and Jiyang Wu and Jingsong He and Chizu Mao and Ziyou Li and Bo Yang},
  title = {{State Monitoring and Fault Diagnosis of HVDC System via KNN Algorithm with Knowledge Graph: A Practical China Power Grid Case}},
  booktitle = {Sustainability},
  year = {2023},
  doi = {10.3390/su15043717},
}
Figure