Multi-modal machine learning prediction of fleetwide switchgear SF6 escape from historical maintenance records, online monitors and domain expertise

Tools

Liu, Ting and Irwin, Fiona and de la Barba, Luis and Holton, Allan and Jones, Dan and Terret-Hensman, Rob and Wilson, Gordon and Riccardi, Annalisa and Brown, Blair David and McArthur, Stephen and Stewart, Brian and Stephen, Bruce (2025) Multi-modal machine learning prediction of fleetwide switchgear SF6 escape from historical maintenance records, online monitors and domain expertise. In: CIGRE 2025 International Symposium, 2025-09-29 - 2025-10-03, Palais des Congrès de Montréal.

[thumbnail of Liu-etal-2025-Multi-modal-machine-learning-prediction-of-fleetwide-switchgear-SF6-escape]
Preview
 Text. Filename: Liu-etal-2025-Multi-modal-machine-learning-prediction-of-fleetwide-switchgear-SF6-escape.pdf
Accepted Author Manuscript
License: Creative Commons Attribution 4.0 logo
Download (713kB)| Preview

Abstract

Sulphur hexafluoride (SF6) has demonstrated exceptional performance as an insulator in switchgear and other substation equipment. However, its technical capabilities have lately been overshadowed by its potency as a greenhouse gas, with severe regulatory penalties incurred from its escape. From an asset management perspective, escape of SF6 is undesirable as it requires an additional level of maintenance and oversight with a level of uncertainty which can be resource intensive; from an operational perspective, the primary concern is the escape of SF₆, as avoiding this is what leads to the asset management activity. Escapes challenge network resilience, as they can render an asset inoperable, with switchgear locking out if gas pressure drops below a particular level. Assets containing SF6 are still highly prevalent in high voltage applications to the extent that wide scale monitoring is not yet feasible to support asset management. Condition assessment is often conducted by proxy through gas top up records from which gas escape mass can be inferred. As a utility company may have several hundred, if not thousands, of such assets of various ages, from various manufacturers and under various duty cycles, manual assessment of these is impractical and prone to inaccuracy. Accordingly, the work described in this paper develops a machine-learning, domain-knowledge and physics-informed predictive modelling framework for anticipating SF6 gas escape. This framework automatically learns trends from historical maintenance records to operate at the asset, asset family and asset fleet level over various time horizons for both operational and planning purposes. Domain specific knowledge has been harnessed to select predictive covariates that can be related to gas vessel escape rates based on geography, past asset behaviour and network application. Additionally, short to medium term behavioural analysis has been investigated from limited sets of online gas density monitoring data to understand the pathologies of leak development and incorporate these findings into predictive models for unmonitored assets. The framework has been developed and tested on the fleets of the three transmission network asset owners of Great Britain with results from the individual and combined predictive models presented. In use, this informs both asset replacement strategies as well as targeting of maintenance effort for increased network resilience.

ORCID iDs

Liu, Ting ORCID logoORCID: https://orcid.org/0000-0003-0604-4939, Irwin, Fiona, de la Barba, Luis, Holton, Allan, Jones, Dan, Terret-Hensman, Rob, Wilson, Gordon, Riccardi, Annalisa ORCID logoORCID: https://orcid.org/0000-0001-5305-9450, Brown, Blair David ORCID logoORCID: https://orcid.org/0000-0002-4734-9985, McArthur, Stephen ORCID logoORCID: https://orcid.org/0000-0003-1312-8874, Stewart, Brian ORCID logoORCID: https://orcid.org/0000-0001-8084-573X and Stephen, Bruce ORCID logoORCID: https://orcid.org/0000-0001-7502-8129;
CORE (COnnecting REpositories)