Session: 20-04: Optimization Through Digital Tools

Paper Number: 153126

Data-Driven Based Digital Twin Leak Detection for Hydrogen Networks for the Gas Turbines’ Applications 

Introduction

The transition to a hydrogen economy necessitates the development of reliable monitoring systems for extensive gas distribution networks, including those used in gas turbines, to ensure safety and operational efficiency. Detecting leaks in such networks is a critical challenge due to the unique physical properties of hydrogen, such as its low molecular weight and high diffusivity, which complicate traditional detection methods. This study extends the application of deep learning techniques, previously explored in natural gas systems [1], to detect potential leaks within hydrogen or mixed-gas networks using a digital twin model. The proposed approach leverages steady-state data and simulated scenarios to train deep learning models capable of identifying subtle deviations indicative of gas leaks.

Hydrogen infrastructure is rapidly expanding, making it essential to maintain the integrity of these networks to prevent environmental damage, property destruction, and safety hazards. Traditional methods, such as the negative pressure wave technique and acoustic wave detection, along with methods like Pressure Point Analysis (PPA) and wavelet transform technology, have been widely applied in gas distribution systems [2,3]. Despite their effectiveness, these techniques often struggle with the complexities of interconnected networks, requiring advanced approaches to detect leaks accurately. Recent advancements in sensor technology have enabled real-time monitoring [4], but these solutions are susceptible to false positives caused by environmental interference and noise.

Deep learning offers a promising pathway to enhance leak detection in complex gas networks by training models on large datasets that include both normal and leakage conditions [5-6]. These models can achieve high accuracy in identifying leaks and localizing their source by recognizing subtle patterns that may be missed by conventional methods. However, the limited availability of operational data for hydrogen networks, combined with the scarcity of anomalous events, poses a significant challenge. To address this, the study employs a digital twin framework based on steady-state simulations, providing a comprehensive training dataset that facilitates accurate leak detection and monitoring.

Result

This study addresses these challenges by employing a two-stage neural network model capable of detecting and identifying multiple leaks simultaneously while reducing false positives. The approach effectively handles complex interferences such as noisy sensor readings and unpredictable consumption patterns, demonstrating robustness across a range of leakage scenarios.

To further enhance model performance, data augmentation techniques are integrated to mitigate the variability and noise present in sensor readings. The deep learning model is validated using simulation data from a hydrogen network, particularly accommodating rare leakage scenarios and potential network changes like aging and topology modifications. The method outperforms traditional models, especially in scenarios involving variable sinks and environmental disturbances. By recognizing the random nature of sinks and sources in real-world gas networks, this approach is positioned as a viable solution for practical applications.

Conclusion

In conclusion, this study contributes to hydrogen network leak detection by presenting a robust, adaptable deep learning framework. The results highlight the potential of integrating advanced digital twin technologies with deep learning, leading to improved safety measures and environmental sustainability in hydrogen energy transport and storage.

References

[1] Ebrahimi, E., Kazemzadeh, M., & Ficarella, A. (2024). Leak identification and quantification in gas network using operational data and deep learning framework. Sustainable Energy, Grids and Networks, 39, 101496.

[2] Karkulali, P., Mishra, H., Ukil, A., & Dauwels, J. Leak detection in gas distribution pipelines using acoustic impact monitoring, in: IECON 2016-42nd Annual Conference of the IEEE Industrial Electronics Society, IEEE, 2016, pp. 412–416.

[3] A. bin Md Akib, N. bin Saad, V. Asirvadam, Pressure point analysis for early detection system, in: 2011 IEEE 7th International Colloquium on Signal Processing and its Applications, IEEE, 2011, pp. 103–107.

[4] Ukil, A., Braendle, H., & Krippner, P. Distributed temperature sensing: Review of technology and applications, IEEE Sensors Journal, 12(5) (2011) 885–892.

[5] Zheng, J., Wang, C., Liang, Y., Liao, Q., Li, Z., Wang, B. Deeppipe: A deep-learning method for anomaly detection of multi-product pipelines, Energy, 259 (2022) 125025.

[6] Liang, J., Ma, L., Liang, S., Zhang, H., Zuo, Z., Dai, J. Data-driven digital twin method for leak detection in natural gas pipelines, Computers and Electrical Engineering, 110 (2023) 108833.

Presenting Author: Elham Ebrahimi University of Salento

Presenting Author Biography: I earned my master's degree in Environmental and Civil Engineering from Shiraz University, Iran. Currently, I am a PhD student at Unisalento, conducting research on sustainable mobility within energy networks.

Authors:

Elham Ebrahimi University of Salento
Antonio Ficarella University of Salento