Weather-based Fault Prediction in Electricity Networks with Artificial Neural Networks

Predicting weather-related outages in electricity networks is an important issue for distribution system operators. In this study, we apply a data-driven approach and train artificial neural networks to predict faults in the electricity network. In our experiments, we utilize the meteorological data and fault records collected for the period of1.1.2011-31.12.2013 in central Finland. Assuming that there might be long-term dependencies between weather conditions and faults in the network, we investigate simple recurrent neural networks, long short-term memory networks, and traditional multilayer perceptrons. Taking into account the meteorological observations preceding faults and varying this period from several hours to several days, we found that 6 hours prior to faults included the sufficient information to make accurate predictions. Also, there was no need in more complicated recurrent neural networks as multilayer perceptron was able to predict events with the large number of faults more accurately. Besides, while forecasting all types of faults and wind-related faults only, oversampling allowed the model to predict rare high peaks.