Research Projects and Publications

Predicting weather-related transmission line failures using machine learning

Power Systems and Smart Grids

Author: Roddy Akeel
Year: 2019
Supervisors: Samuel Perkin, Ragnar Kristjánsson



Power systems are required to reliably deliver electricity to many different consumers. To attempt to meet this constant demand for delivery while also maintaining a level of safety, power systems operators must make informed decisions by considering a number of factors. At present, utilities are developing and implementing smarter methods of reliability assessment to allow for more informed decisions to be made. This thesis presents a machine learning approach for determining and providing a general weather-related transmission line failure probability based on weather forecasts. To develop this informational tool, a number of different feed forward neural network configurations were built, tested, and assessed. In combination with the feed forward neural network, different preprocessing methods were introduced and utilized to allow different information to be extracted from base weather parameters prior to neural network neural network training and testing. These preprocessing methods include: seasonal classification, k-means classification, and a custom risk parameter flag. Other preprocessing methods used in this study aim to address the redundancy of weather conditions which are related to both non-failure and failure scenarios. The first was a daily aggregation method which effectively reduced the number of non-failure data entries while leaving the number of failure entries intact. The second method was a failure weather condition dilution method which increased the number of failure weather entries to be both 1:3 and 1:1 with respect to non-failure weather entries. In addition to developing a neural network approach to failure probability prediction, a 3-state failure probability method was investigated using a k-means classifier to show the amount of information provided by the neural network compared to another state-of-the-art approach. Finally, three case studies were conducted on transmission lines which experience a low, medium, and high risk of failure using the best performing neural network configurations.

Link to Publication