Summary of the project, product, framework
Western Power’s (WP) Pole Top Fire Model combines an engineer’s view of risk on WP assets with the weather conditions that trigger pole top fires.
WP Fault Crews are always available, but a spike in the number of incidents can put resources under pressure. Predicting the number and general location of pole top fires can help our scheduling team have enough crews ready, where and when they’re needed. This helps reduce risks to public safety and reliability of supply, without over resourcing fault crews.
Description of project or framework addressing the assessment criteria
Pole top fires are a major risk for electricity network operators but difficult to predict. Pole top fires usually occur on the 22kV or 33kV distribution network and, while rare, are considered high impact events.
Pole top fires affect a tiny fraction of WP’s network each year, but occasionally cluster geographically and temporally. Outages following pole top fires tend to be long and require complex repairs. Further, due to the localised weather that contribute to the fires, one area can be affected by multiple fires and power restoration is further delayed.
The cause of pole top fires is attributed to dust build up on insulators that leak current during misty rain, estimated to be a mere 0.2 mm of precipitation per hour. Typically, the insulators and cross-arms also have been impacted by normal wear and tear.
Modelling of pole top fires has traditionally been for proactive maintenance – washing power lines and applying silicone to insulators. While these techniques can be effective, it is not long lasting. Anecdotal evidence showed some leading indicators for pole top fires, however a formal short-term forecast could allow for more fault crews to be available when needed.
WP has existing risk models that evaluate the probability of asset failure for each asset on the network. These models leverage asset specific data to predict the long-term risk and are included in the new models to assist short-term prediction.
In addition to the asset data, geospatial observations and short-term forecasts were acquired for weather and particulate pollution. Collectively, this information can be used to identify the level of pollution and moisture that will naturally fall or accumulate on the pole tops.
A range of predictive modelling techniques were investigated using a fail-fast methodology. Random Forest, a popular machine learning technique, was chosen due to its ability to elegantly model complex non-linear relationships and the speed at which it trains. Historically Random Forests have had a reputation as a black-box technique, however key innovations in the last two years (particularly LIME and Partial Dependence plots) allow practitioners to explore the inner workings of the model in detail.
The models are run every six hours through our enterprise automation system, with automatic alerts sent to staff when there are any heightened risks of fires.
The output of the model is displayed in a dashboard that provides a visual representation of the day ahead view of risk, colour coded by region, and an hourly view of risk by region over the coming week. The interactive nature of the dashboard allows our crews and scheduling teams to engage with and explore the output of the models.
The dashboard includes a historic view of forecast risk and observed fire events, this page is crucial to the acceptance of the model by the business as it provides rapid feedback on how well the model has been performing.
Inspiration came in mid-2017 following a day of multiple pole top fires that affected the Perth Metro area. WP Executive Manager, Seán McGoldrick, made the comment, “if only we could have known, we would have been ready.”
Peter Condon developed the first iteration of the model using the 950 feeders across Western Power’s network as natural grouping based on network topology. The feeders were then aggregated to the nearest of 38 weather stations that offer sufficient historic observations and forecasts.
Air quality data from the Department of Water and Environmental Regulation was also overlayed to identify regions that may be impacted by recent dry conditions. Within a few days of launching, the model predicted a spike in pole top fire risks over an upcoming weekend. By Monday, a series of localised pole top fires had started and demonstrated that the model worked!
Most pole top fires occur between December and March, so Peter could make several iterations to the data inputs in the model over the summer period and has seen the reliability of forecasts improve. In May 2018, far outside of the season in which we typically see pole top fires, there was a spate of pole top fires in the mid-West, which was successfully predicted by Peter’s model.
The model is already driving change within Western Power. Crews are mobilised when the model predicts an increased likelihood of fires and outages duration has been improved following these incidents.
One year after Peter was inspired to develop a model that could predict pole top fires, there are multiple fully automated independent models being scored each day with a single ‘champion’ model selected in November 2018 to see Western Power through the 2018/19 summer.
Opinion as to specific contribution made by the nominated individual/team/organisation
Perth’s desirable climate creates a perfect storm for pole top fires – and the ideal place to be able to test and trial the predictive model.
The model is quickly becoming a valuable asset to ensure appropriate staffing of field crews and to minimize the impact of redeployment and cancellation of planned work. Western Power has already benefitted financially from the development of this innovative model.
Pole top fires affect electricity networks around the globe and Peter’s model is well placed to be able to improve the safety and reliability of these networks.
Identifying the performance of specific network assets and its impact on pole top fires offers other networks unique insight to their networks and has potential to drive investments and maintenance plans. It goes far beyond just a technical application of information – it touches on asset management and planning to help networks become even more effective and efficient.
General comments you may wish to add
Coincidently, Peter and his team members were taking a class to become Certified Analytics Professionals, which would give them a shared problem-solving framework and language. He was looking for a case study to work through analytics for the course and, with so much data available, this seemed like the perfect mission. He set out to predict the risk of incidents per hour in each of the 38 weather station catchments that cover the WP network.
Peter’s colleagues reacted with both intrigue and scepticism. Pole top fires were thought to be random, but if Peter looked at the right data, perhaps he could identify some leading indicators that would allow for more fault crews to be available when needed?