Scholarlink

<p>Minimizing gas losses has become a crucial factor in optimizing the utilization of energy resources. Multiple factors contribute to the complexity of distinguishing between normal gas loss and anomalous behavior within the gas network. This study aims to identify abnormal behavior in looped, multi-pressure regime, natural gas networks. Anomaly detection is used to identify the type of discrepancy: gas leak, malicious activity, or excessive unaccounted for gas (UFG) quantity, which can trigger immediate action to reduce gas loss. A machine learning anomaly detection algorithm using multivariate Gaussian density estimation combined with principal component analysis (PCA) is developed. The algorithm is applied on datasets of 21 natural gas networks collected over a period of 5&#xa0;years in two streams. In the first stream, each algorithm learns only from its own dataset. In the second stream, a novel approach to construct a global network dataset is adopted. In each stream, two scenarios are applied (with and without the application of PCA). The outcome from the algorithm’s application in various scenarios was evaluated using four performance metrics, namely, Precision, Recall, F1score, and Accuracy. </p><p>The results showed that the anomaly detection algorithm is able to identify anomalies, in the best scenario, with an average Recall (0.99). Therefore, the Recall metric can be used as a strong measure to detect true anomalies. Although the metrics F1score, Precision, and Accuracy have substantially lower values than that of the Recall, they were useful in detecting data inhomogeneity. The PCA was used for better visualization and understanding of the networks under consideration. The globalization technique was useful for generalizing over different networks having various numbers of industrial, commercial, and residential customers with their corresponding diverse loading profiles. Despite the reduced value of Recall for the global dataset with a recorded Recall (0.67), it can help with anomaly detection in situations where no historical data is available. </p><p>The application of the developed algorithm proved capable of identifying the abnormality in network balance and the consumer identification number. When linked to a digital map, emergency and/or investigation crews can be easily deployed for immediate action according to specific approved procedures. Furthermore, with the use of the Internet of Things (IoT), online devices such as meters, pressure and temperature online transmitters, and anomaly detection in real-time can trigger immediate action, saving gas loss.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning anomaly detection of lost and unaccounted for gas in natural gas networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月5日

查看全文
  • Omar Fayez Mohamed ElMahdy,
  • Mohamed Ezz Hassan,
  • Sayed M. Metwalli

摘要

Minimizing gas losses has become a crucial factor in optimizing the utilization of energy resources. Multiple factors contribute to the complexity of distinguishing between normal gas loss and anomalous behavior within the gas network. This study aims to identify abnormal behavior in looped, multi-pressure regime, natural gas networks. Anomaly detection is used to identify the type of discrepancy: gas leak, malicious activity, or excessive unaccounted for gas (UFG) quantity, which can trigger immediate action to reduce gas loss. A machine learning anomaly detection algorithm using multivariate Gaussian density estimation combined with principal component analysis (PCA) is developed. The algorithm is applied on datasets of 21 natural gas networks collected over a period of 5 years in two streams. In the first stream, each algorithm learns only from its own dataset. In the second stream, a novel approach to construct a global network dataset is adopted. In each stream, two scenarios are applied (with and without the application of PCA). The outcome from the algorithm’s application in various scenarios was evaluated using four performance metrics, namely, Precision, Recall, F1score, and Accuracy.

The results showed that the anomaly detection algorithm is able to identify anomalies, in the best scenario, with an average Recall (0.99). Therefore, the Recall metric can be used as a strong measure to detect true anomalies. Although the metrics F1score, Precision, and Accuracy have substantially lower values than that of the Recall, they were useful in detecting data inhomogeneity. The PCA was used for better visualization and understanding of the networks under consideration. The globalization technique was useful for generalizing over different networks having various numbers of industrial, commercial, and residential customers with their corresponding diverse loading profiles. Despite the reduced value of Recall for the global dataset with a recorded Recall (0.67), it can help with anomaly detection in situations where no historical data is available.

The application of the developed algorithm proved capable of identifying the abnormality in network balance and the consumer identification number. When linked to a digital map, emergency and/or investigation crews can be easily deployed for immediate action according to specific approved procedures. Furthermore, with the use of the Internet of Things (IoT), online devices such as meters, pressure and temperature online transmitters, and anomaly detection in real-time can trigger immediate action, saving gas loss.