Learnings

Outcomes

Since March 2025, HV cable sensors and Precision Event Timing Units (PETUs) developed through the HV Pinpoint project have been deployed on a live section of NGED's 6.6 kV network. The sensors were successfully installed across four different HV cable types without interrupting network operation, demonstrating that the technology can be deployed using a non-invasive approach.

The system has now detected and recorded over 200 potential cable defect events. Analysis shows these events are associated with early-stage cable deterioration (micro-arcing), with most occurring during or shortly after periods of heavy rainfall. The trials have demonstrated that the technology can successfully identify and locate events on both two-ended and multi-ended HV circuits to a level of accuracy sufficient to support the next stage of the HV Pinpoint process using the PIG validation system.

The monitoring data has enabled event locations to be analysed using advanced clustering techniques, which have proven more effective than reviewing individual events. This approach helps distinguish genuine defect locations from isolated events and network artefacts, such as signal reflections caused by circuit characteristics. Current location accuracy is typically around 1% of circuit length, which is sufficient for validation activities, while recent improvements in GNSS timing performance are expected to deliver further increases in accuracy.

Testing at NGED's Midland Training Centre has also provided end-to-end validation of the overall approach. Controlled trials on both two-ended and three-ended circuits successfully located known defect positions, confirmed event locations using the PIG time-of-flight validation method, and demonstrated that a prototype sensor mat could accurately identify the physical location of buried cable defects.

In addition to locating cable defects, the technology has demonstrated the ability to identify and analyse a wide range of network events, including switching operations and power quality disturbances. The project has therefore established a strong foundation for future development, with several additional applications and opportunities identified for further exploration beyond the scope of the current project.

The next stage of the project is focused on validating and precisely pinpointing detected defects. While initial methods and prototype technologies have been successfully demonstrated, further development is required to deliver a fully integrated, user-friendly operational solution. Completion of the remaining validation activities is essential to verify end-to-end system performance, build confidence in the technology, and ensure the project achieves its intended outcomes and benefits.

Lessons Learnt

During the live trial phase, HV Pinpoint equipment has been successfully deployed on a live 6.6 kV circuit in Coventry. Although a significant number of events were captured and analysed, no confirmed High Voltage (HV) pecking events associated with a known HV network defect have yet been identified. The targeted circuit had been selected because historical data suggested the presence of developing HV faults; however, detailed analysis of the captured waveforms indicated that many events initially believed to be HV pecking activity were actually originating from the Low Voltage (LV) network. This highlighted the complexity of distinguishing between HV and LV events using waveform data alone and reinforced the need for improved classification techniques.

To address this challenge, a machine learning work package was developed using both HV and LV waveform data. The work package demonstrated that automated waveform classification is technically feasible and can be used to differentiate between a range of distribution network events. A major obstacle was the lack of labelled real-world fault data. Whilst more than 28,500 NGED waveforms were available, very few had verified fault classifications. To overcome this limitation, a synthetic dataset comprising 2,400 labelled HV and LV fault waveforms was created and used to train supervised machine learning models. By systematically varying fault magnitude, duration and inception angle, the synthetic dataset provided a broad range of realistic fault scenarios that enabled the models to learn the characteristics of known fault types before being applied to operational data. Model performance reached approximately 78% accuracy using synthetic data alone and improved to approximately 81% when trained using a combination of synthetic and selected real-world waveforms. The resulting models were then used to analyse previously unclassified NGED waveforms and provide an indication of the likely distribution of HV and LV events within the dataset.

The project demonstrated that synthetic data can successfully aid supervised machine learning where labelled operational data is scarce. However, one of the key lessons learned was that future improvements are unlikely to come primarily from more sophisticated machine learning algorithms. Instead, greater benefits are expected from improving the quality and availability of labelled data, enhancing the realism of synthetic models, improving network observability through better sensor placement and visibility, and integrating engineering knowledge and known network event information more closely into the machine learning workflow. These factors are likely to have a greater influence on overall performance than increasing model complexity alone.

The Coventry trials demonstrated the importance of establishing robust event timing, data analysis and validation processes from the outset. Small timing errors were shown to have a significant impact on fault location accuracy, highlighting the need for effective system synchronisation and GNSS configuration. Analysis techniques such as event clustering and heatmapping proved more effective than assessing individual events, improving confidence in identifying genuine defect locations. The work also emphasised the need for robust noise management and multi-sensor validation to reduce false positives and improve data quality. Successful testing relied on thorough preparation, commissioning and validation activities, while near real-time visualisation and automated analysis were identified as key enablers for operational deployment. Finally, the influence of environmental conditions, network characteristics and circuit reflections reinforced the importance of adopting an iterative development approach, allowing lessons learned to be incorporated rapidly into future hardware, software and analytical improvements.

The project also highlighted a common challenge associated with deploying low Technology Readiness Level (TRL) solutions directly onto live operational networks. Equipment limitations, site constraints and the unpredictable nature of network faults reduced the likelihood of capturing the specific fault events required to fully validate the technology within the available project timeframe. As a result, achieving the original project objectives proved more difficult than anticipated and validation activities were delayed. Nevertheless, the experience has provided valuable insight into both the strengths and limitations of the current approach and identified several opportunities for future development.

A key recommendation for future projects is that early-stage technologies requiring validation against real fault conditions should first be assessed within controlled test environments. Such environments would provide repeatable and predictable fault scenarios, enabling faster validation, improved data capture and more efficient system development prior to live network deployment.