#Digital #Energy

Preliminary results show that machine learning can predict faults and disturbances in the power system

In the EarlyWarn project, predictive models (using big data, machine learning and domain knowledge) are being developed based on historic power quality data.  Preliminary results show that it is possible to predict (with a relatively high accuracy) events in the power system purely based on the development of power quality data in the period before the actual fault event occurs.

It is essential that the Norwegian power system reliability be high since Norway is one of the most electricity-dependant countries in the world. Picture is from Oslo harbour. (Photo: Shutterstock)

Figuring out how to reduce downtime in the power system

On a global scale, Norway has one of the most reliable power systems. In 2017, the continuity of supply in the Norwegian power system was 99.988 %.

Norway is one of the most electricity-dependant countries in the world, as heating in Norwegian households is predominantly electricity-based. Due to this, it is essential that the Norwegian power system reliability be high.

If there is a power outage in the Norwegian high-voltage transmission and distribution system, grid companies are required to cover the cost of energy not supplied. (Cost of energy not supplied is a measure of the calculated value of lost load for the customers. More info here.) Due to this, grid companies have increased their focus on continuity of supply and are investigating new ways of reducing the downtime in the power system.

Pictures: Insulation failure is a type of event than can cause downtime in the power system. (Picture: Shutterstock).

Power quality analysers have gathered big data for over 15 years

In order to increase the knowledge about grid condition and causes of faults in the power systems, many grid companies have installed a number of power quality analysers in their concession area. Power quality analysers are measuring devices that can log the voltage and current waveforms with high accuracy.

These measuring instruments measure voltage and current up to tens of thousands of samples per second. 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.

Some of these power quality analysers have been logging the power quality in the Norwegian power system continuously for over 15 years.

A 10-year time-series of the line voltage in a 22 kV medium-voltage grid is shown in the figure below. In normal operation, the line voltage should be around 22 kV. This means that any deviations from 22 kV (over- or undervoltages) in the figure are caused by some kind of event in the power system. Thus, the power quality analyser has logged a high number of faults and disturbances in the given measurement period in the figure below.

Figure: 10-year time-series of the RMS line voltage in a 22 kV grid. Logged with a power quality analyser.

Combining knowledge about fault signatures with machine learning

New knowledge about fault signatures and their development, in combination with new methods within machine learning, could enable early detection of problems in the power system.

This could in turn enable the grid operator to initiate preventive measures to avoid outages and other problems for the consumers.

The EarlyWarn project aims to predict faults in power system

EarlyWarn is a research project that aims to develop a monitoring system to detect and identify disturbances in the power system before they escalate and cause faults, such as power outages and other issues.

This will be done by exploiting the large amounts of measurement data that are gathered continuously from the power system. EarlyWarn is applying techniques from big data, machine learning and domain knowledge to automatically and constantly monitor the sensor data, in order to alert system operators of instabilities or disturbances that would otherwise have gone unnoticed. Thus, the objective in the EarlyWarn project is to develop machine learning models that can form the basis in such a monitoring system.

Developing predictive algorithms

When developing such machine learning models, datasets with tens of thousands of historical data-series (observations) from both fault events and normal situations are needed to train and test the functionality of the algorithm. Thus, in its first phase, the EarlyWarn project has focused on developing methods for generating user-customised datasets with data-series based on observations from power quality analysers.

Based on these datasets, predictive algorithms are being developed using different types of machine learning methods.

Snow weighing down a power line. If allowed to develop further, a potential fault may occur. (Photo: Shutterstock)

Preliminary results from EarlyWarn are very promising

As mentioned in the beginning, EarlyWarn results show that it is possible to predict events in the power system purely based on the development of power quality data in the period before the actual fault event occurs.

In addition, the preliminary results have shown that the accuracy of the predictive model was higher for some fault categories (e.g. full interruptions) than for others (e.g. earth faults, voltage dips). They also showed that the predictive ability of the model was close to constant with increasing forecast horizon (a predictive horizon of up to 40 seconds before the fault event has been tested so far). This means that there seems to be a potential for predicting fault events more than 40 seconds before the actual fault event, maybe up to several minutes or hours, but this is yet to be tested in EarlyWarn. The aim is to be able to predict fault events at least a couple of minutes in advance, so that a grid operator can initiate a preventive measure in the operation centre. The preliminary results from EarlyWarn have been published in two scientific papers, in the AMPS and CIRED (to be published – June 2019) conferences respectively.

The preliminary results are promising and prove that further work should be done on testing different machine learning methods on power quality data, with the aim of increasing the performance and forecast horizon of the predictive models.

EarlyWarn is a KPN project (part of the ENERGIX programme) that is partly financed by the Research Council of Norway, Statnett, Haugaland Kraft Nett, NTE Nett, Lyse Elnett, Nettalliansen, Hydro Energi and NTNU. R&D partners in the projects are SINTEF Energy Research (project lead), SINTEF Digital and NTNU.

0 comments on “Preliminary results show that machine learning can predict faults and disturbances in the power system

Leave a Reply

Your email address will not be published. Required fields are marked *