Modelling faults on power distribution networks

This is the title of the case study that should take up no more than a few lines.

nasa
Weather disrupts power supplies - knowing where and when this will happen allows you to plan 
This is the space for a nicely written and engaging synopsis that will entice people to read.

Client: SSE Plc
Date: 2020

Client: The London Fire Brigade
Date: September 2018

The power distribution network in the UK is a regulated monopoly - good performance isn't ensured by competition, but by regulatory oversight. In short, the power distributor will be fined for poor performance, which includes power interruptions to its clients. Given much of the high-voltage infrastructure is above ground, it is prone to damage from weather conditions. Predicting where/when faults will occur based on forecast weather data allows planning to minimise interruptions, and associated regulator penalties.

What we've done about it

We have built predictive models that predict the numbers and locations of faults on the UK's SSE distribution network. Weather forecasts for up to 5 days may be used, with predicted regional fault numbers returned along with measures of uncertainty. 

Some technical bits

Unlike many predictive modelling problems, the uncertainty around the predictions is of particular interest here - planning is made against best-/worst-case scenarios. The predictive modelling methods currently gathered under the machine-learning banner, tend not to provide inference. We've evaluated a large number of model classes, favouring those more statistical in nature so we can give probabilitistic predictions. 

The current favoured models are variants of Generalised Additive Models, combined with boosting, and allowing for flexible modelling of the error distributions. Separate models are fitted for each 24-hour weather forecast period i.e. 1 day into the future, 2 days into the future etc. Models have been fitted and validated against a decade of fault and forecast data, and are naturally validated every day through their industrial use.