Background and objectives

Our client in one of the largest power distribution companies in the UK with more than 10m customers. To deliver its aspiring and inspiring cost reduction journey, it has set ambitious targets to reduce the fault rate and related repair costs of underground cables and overhead lines. To better understand the underlying reasons for its network failures, our client has recognised the potential of predictive analytics / machine learning to better predict where and when the network will fail. With that prediction, it can better allocate its replacement budget and reduce network fault rates accordingly.


Our approach

With representatives of the operational and IT teams, in about a month time we developed a baseline model to predict network faults and identified there was predictive power in the data. In doing so, we firstly identified various correlations before we would dive into the causality which the machine learning would address.



As a spin-off benefit of many Proof of Concepts, we typically identify operational issues when conducting our feature and correlation analysis. In this case we identified a counterintuitive correlation between the number of faults per secondary sub-station and the number of faults per secondary sub-station: the more a secondary sub-station switches to its back up network, the less often the network fails. Independent from the predictive analytics work we conducted, we fed this correlation back to our client. In response, they are adjusting tolerance levels and therewith significantly reduce the fault rate in their network.