Data Study Group Final Report: University of Strathclyde and Supergen Energy Networks Hub

Power grids are a fundamental part of today’s infrastructure - almost all aspects of modern life require power. Blackouts can be catastrophic events that can lead to great discomfort or even fatalities. Due to the high degree of spatial correlation of power grid networks, sudden changes in the operating conditions of a given component can cause failures in other nearby components. These failures may then affect other components causing further failures, resulting in a series of failures, known as a cascade, which can result in a blackout. Recently blackouts have become more frequent due to two primary factors: the increased complexity of grids caused by incorporating generators with a more variable power output, such as wind and solar, and the increased frequency of extreme weather events causing damage to grid components. Hence, reliability is a growing concern in power grids. Although there are many countermeasures and safety nets to avoid power outages, the vast amount of system parameters makes it difficult to predict and account for all failure modes. Furthermore, modelling of the physics of power grids requires numerically solving systems of hundreds or thousands of coupled differential equations, making an online physics-based approach to blackout prediction impossible with today’s computing power. Machine learning approaches, in particular deep neural networks, are very promising because they are universal function approximators and can therefore learn the complex relationships between the large number of system parameters and the resulting failure events.

The overall aim of this challenge is to investigate the spatial and temporal distribution of failures across the power grid, examine the relationships between the components and predict the first failure before it happens in order to prevent it. We use the simulated dataset provided by the University of Strathclyde & Supergen Energy Networks Hub and utilise machine learning methods to achieve exciting results. We also aim to identify the limitations of the data and the methods used in order to provide a road map for future research.