This project will use machine learning techniques to improve forecasting capability for SF6 circuit breakers taking nameplate, environmental and operational factors into account. Building on existing work the following will be undertaken:
- Understand the performance of the existing model when applied to NGET assets
- Develop the model taking NGET assets into account and then test to determine accuracy
- Collect online density monitoring data and use it to improve the forecasts
- Demonstrate how forecasting may be used to automate planned interventions.
Benefits
This project aims to improve the forecasting of SF6 circuit breaker leaks and will provide benefits in the form of reduced SF6 leaks / emissions through automation of and better planning for interventions.
Learnings
Outcomes
Data from GDMs has been used to create monthly correction factors to offset the potential for a forecasting bias based on when leaks happen caused by when top-up events occur in practice.
A state-based approach to predicting leak rates shows promise for predicting when top-ups are likely to be needed based on historic top ups.
Recommendations for further work
None at this stage
Lessons Learnt
Identification of relations between environmental and condition monitoring data was undertaken through consideration of the distribution of pairs of variables. No obvious relations were found.
- Density: It would be expected that densities would co-behave similarly across assets on the same site e.g. increase and decrease at the same time, but this was found not to be the case in the three decommissioned assets considered. The observed independence of density between the three could result from shading or sheltering from prevailing wind. The absence of any load going through the assets would remove any diurnal trend from load profile induced heating.
- Density vs. Temperature: The density and temperature display a more dispersed scatter pattern for each asset demonstrating independence again, albeit with more variation for one asset. As most asset heating comes from load and these are decommissioned assets, this was to be expected.
- Density vs. Relative humidity (RH): No real relation seen with two of the assets possibly skewed by outliers.
- Temperature vs. RH: A more obvious environmental trend observed where extreme temperature drives down humidity.
Extending the analysis to the larger dataset from 16 monitors revealed some issues with calibration and outliers. As above density and RH showed little relation, while a strong relationship between density and pressure was observed as expected. Some monitors showed a relationship between pressure, density and temperature but this was not seen in all datasets. This suggests that pressure inferred where not measured may not be reliable.
The escape behaviour of SF6 across different circuit breakers was observed to vary considerably between assets. The assets examined as case studies to show how well the linear regression prediction model worked, had a mixture of results, not obviously linked to asset type or severity of SF6 leak (as given by number of top ups and gas volume topped up). MAE ranged from 0.08 to 1.56, although most were below 0.50. RMSE values were very similar to the MAE values, indicating that outliers are not significant as the errors do not vary much. MAPE values were sometimes below 10% indicating an accurate prediction model. The remainder were in 10-20% range which shows some accuracy but that improvements might be necessary.
Analysis of SF6 top ups have a bias towards winter events – this may be a result of the colder temperatures on the pressure causing more alarms. This does not mean that SF6 losses are greater at those times and a misleading picture of gas loss could be gained.
The Markov-chain model was compared using NG data only and data for all GB (Great Britain) transmission assets against a naïve model (assumes leak rate is always constant) and a linear regression model (fitted trend line to predict future leak rate) using historic top up data. The Markov “state-based” model shows promise. MAPE values applied to the NG and GB fleets of assets were around 30-40% after some degree of filtering (to remove catastrophic events).
Retrofitting GDMs onto in-service equipment can prove difficult. On this project the following observations were made:
- Filling points, where they may be connected, can have a variety of connectors necessitating a range of fittings available for the GDM to be installed
- Connection points may not be accessible from ground level, meaning that an outage would be required for the installations
- Filling points may serve three phases of one asset making it impossible to monitor the density in all three SF6-filled tanks individually
- Where filling points are inside cabinets, it may not be possible to connect to a device and close the cabinet, or space may be restricted.
Dissemination
The project results are shared with SPEN and SSEN-T as they carried out the previous work that led to this project and are invited to attend project update meetings.
A paper entitled “Multi-modal Machine Learning Prediction of Fleetwide Switchgear SF6 Escape from Historical Maintenance Records, Online Monitors and Domain Expertise” has been submitted for publication and presentation at the CIGRE Symposium in Quebec in September 2025. This is a joint paper with SPEN and SSEN-T.
