Predicting fault events in the Norwegian electrical power system using deep learning : a sequential approach

Abstract

The demand for energy is steadily increasing. The global community is working towards a society supplied by clean energy from renewable sources. This leads to new requirements for the electrical power system. To solve the transition, digital technologies need to be implemented in the system through smart grid solutions. This master’s thesis explored one central aspect of smart grid; prediction of fault occurrences in the power grid. These questions were attempted to be answered: Is it possible to predict the faults before they happen? How long in advance can they be predicted, if the first question turns out positive? What kind of physical parameters are most suitable to use as features for prediction? What kind of method should be used to predict the faults? Will it be possible to implement this in a real-time monitoring system?

Knowledge from the electrical power system domain and the data science domain were combined to obtain a qualitative evaluation of the tests conducted. Measurement data of the fault types interruption, voltage dip, and earth fault, as well as data gathered from nominal operated power grid were be used in the predictions. The samples were collected from Norwegian power grids operated at voltage levels ranging from 15 kV to 300 kV. The assessment was be divided into multiple tests. The objective in focus was to build and compare results from three different recurrent neural network (RNN) architectures trained on time-series data acquired from power quality (PQ) measurements.

A sequence-to-sequence Autoencoder was proposed for use in signal feature extraction. Various tests were conducted from investigating the raw data, to analysing the output of the model. Results have shown a prediction horizon up to seven minutes is possible. It was proposed that even longer horizons may be plausible. Further investigation into the harmonic components was proposed, related to signal analysis and statistics, for better feature extraction. The most promising features were the harmonic components of voltage and current. The fault types may have various composition of harmonic components giving the different fault types an unique signature. Improvements to the model have been proposed, focusing on anomaly detection. In combination with other monitoring equipment, a fault event prediction system can be used as a tool in decision making.

As part of a competence building research program this thesis contributes to the foundation of further research on the area outlined. Placing the research in a broader view, the results may lead to increased security of power supply, reduced operation and maintenance (O & M) costs, and indirectly reducing the impact on the environment by enabling a safer integration of sustainable energy sources.

Etterspørselen etter energi er i stadig økning. Verdenssamfunnet samarbeider om en fremtid forsynt av ren energi fra fornybare energikilder. Dette fører til nye krav for det elektriske kraftsystemet. For å løse overgangen må digitale teknologier implementeres i systemet gjennom smartnettløsninger. Denne masteroppgaven utforsket et sentralt punkt ved smartenett; nemlig predikering av feilhendelser i strømnettet. Disse spørsmålene ble forsøkt besvart: Er det mulig å predikere feilene før de oppstår? Hvor langt inni fremtiden kan feilene bli predikert? Hvilke fysiske parametre er mest egnet til bruk som features ved predikering? Hvilen metode burde brukes til å predikere feil? Vil det være mulig å implementere dette i et sanntidsmonitoreringssystem?

Etterspørselen etter energi er i stadig økning. Verdenssamfunnet samarbeider om en fremtid forsynt av ren energi fra fornybare energikilder. Dette fører til nye krav for det elektriske kraftsystemet. For å løse overgangen må digitale teknologier implementeres i systemet gjennom smartnettløsninger. Denne masteroppgaven utforsket et sentralt punkt ved smartenett; nemlig predikering av feilhendelser i strømnettet. Disse spørsmålene ble forsøkt besvart: Er det mulig å predikere feilene før de oppstår? Hvor langt inni fremtiden kan feilene bli predikert? Hvilke fysiske parametre er mest egnet til bruk som features ved predikering? Hvilen metode burde brukes til å predikere feil? Vil det være mulig å implementere dette i et sanntidsmonitoreringssystem?