Back to search

ENERGIX-Stort program energi

EarlyWarn - Proactive detection and early warning of incipient power system faults

Alternative title: EarlyWarn - Proaktiv deteksjon og tidlig varsling av ustabilitet begynnende svikt i kraftsystemet

Awarded: NOK 6.1 mill.

Project Number:

268193

Project Period:

2017 - 2022

Funding received from:

Subject Fields:

The EarlyWarn project has developed AI-based monitoring systems to detect and identify potential problems in the power grid, before the problems escalate and cause blackouts or destruction of equipment. The power system is equipped with many sensors that continuously register the power and current flowing through the system. There is a trend that ever more instrumentation is being deployed, and that more data is being collected near real-time. Many of these sensors have very high time resolution (some up to 50 kHz). At this resolution, the sensors may register small disturbances in the system, caused for instance by equipment that is damaged or weakened, but has yet to fail completely. The detected of incipient faults prior to occurrence would be beneifical, so actions may be taken before situations such as blackouts occur. At the same time, such sensors generate enormous amounts of data. Manual interpretation of such data streams is unfeasible, and EarlyWarn has therefore applied techniques from machine learning to automatically and continuously monitor the sensor data, in order to alert system operators of instabilities or disturbances that would otherwise have gone unnoticed. The project has identified several error modes for which it is reasonable to assume there exists precursors that can act as signatures for specific errors. Discussions with domain experts has given insight into the required notification horizon needed for implementing mitigating actions relative to faults of various types. These time constants will be used as a basis for the generation of the training data sets and algorithm development. Since manual error identification on such large datasets as those intended for machine learning in this project is unfeasible, automatic algorithms for error detection and labelling in time-series have been utilized. In excess of 250 years of high-resolution power quality timeseries collected from more than 45 locations in Norway has been analysed, and in excess of 10000 events has been identified by type, location of measurement and time. Since many of these events are close in time, a set of training events has been generated with 2000+ events for training. Based on these event lists, and a selection of time intervals representing normal operations for the power grid, a dataset is generated that is suitable for application of machine learning algorithms and inspired by the timescales outlined in the notification horizon mentioned above. Work is carried out in order to investigate which methods and features is most suitable for training data-driven methods (feature selection). There has been a number of publications from the project in this field, this includes: Work on clustering algorithms of unwanted events for balanced and non-balanced data sets. Work on statistics for the temporal distance between unwanted events and the distribution of these. Discussion on the implication for prediction of unwanted events. Work on the suitability and performance of a selection of machine-learning methods and a comparison of their performance on a selection of data sets. Quantification of the predictive capabilities for each method is given. Work on the seasonal dependence of the algorithms on the training data, including advantages and disadvantages of training on seasonal specific data. Work on the added benefit from including several complimentary data sources to the basis for the predictions. Quantitative measures for the evaluation of the added value has been proposed. In addition to the above-mentioned work a pilot study has been carried out in cooperation with a DSO in order to evaluate any potential barriers against deploying these methods in an operational environment. The pilot has implemented a full data stream from the sensor through to results visualisation for an operator, and delivers continuously updated predictions of the probability of unwanted events in the next time interval. The results from the pilot study is under publication in a peer-revied journal. The results demonstrates that it is possible to get a statistically significant predictive capability from analysing power gird data using datadriven methods, however the accuracy and false/negative ratio of the developed methods and models is not sufficient for operationalisation as of now. The project has demonstrated that improvements can be made though the inclusion of multiple data sources, and this is identified as the most productive route for increasing the accuracy of the models. The project has included the work by 4 master students at NTNU and NMBU that has delivered 3 Master Thesis on selected topics in the project. Furthermore, the project has included 2 PhD-candidates that has worked in the project and published results from these works. However, the PhD period for both candidates exceeds the project duration, and thus both are yet to finalize their PhD.

EarlyWarn har hatt som mål å øke kunnskapsbasen for automatisk analyse og utnyttelse av overvåkningsdata fra strømnettet. En sterkere grad av automatisering og digitalisering av strømnettet er nødvendig for å kunne realisere fremtidens smarte nett som ryggraden i Norsk energiforsyning. EarlyWarn har gjennom sitt arbeid innen utvikling av modeller og metoder, bruk av datadrevne metoder og felttesting hos DSO bidratt til kompetanseheving innen dette feltet. Gjennom utstrakt publisering av resultater i fagfellevurderte kanaler åpen for alle interesserte er denne kunnskapen blitt tilgjengelig for allmenheten. Dette bidrar til at videre arbeid innen automatisering og digitalisering av strømnettet vi kunne nå lenger.

The EarlyWarn project addresses the need for improved real-time monitoring and state assessment in the transmission and distribution grid. The project will develop models for prediction and identification of incipient component failures and system instabilities, exploiting recent advances in machine learning and Big Data technology to continuously and in real-time monitor and analyze large streams of grid sensor data from synchrophasors and power quality analyzers. The project will bring together expertise in power systems and ICT from SINTEF and NTNU, as well as from grid operators at both transmission and distribution level. The consortium partners are continuosly collecting the sensor data that will drive the development of algorithms and tools, thus ensuring real-life relevance and value of the research. The project will contribute to the long-term plan for science and education through the education of two PhDs cooperating on cross-disciplinary and strategically important topics.

Publications from Cristin

No publications found

No publications found

Funding scheme:

ENERGIX-Stort program energi