March 10, 2020 — Customer Stories
Too hot to handle: How A.I. can tackle overheating
A case study about how advanced mathematical models can help wind turbine operators to detect problems with overheating before they cause failures.
Head of Research, Greenbyte
Are your turbines burning to tell you something? With 30% of a wind farm’s lifetime cost attributed to component failure and maintenance costs, any solution that can reduce downtime is a major competitive advantage for wind turbine owners.
The good news is that wind turbines produce large amounts of data about their state and performance. This data often reveals problems, such as overheating oil, wearing gears or parts that need replacing – but you have to know how to interpret it.
Our Predict application is a temperature-based condition monitoring system that uses artificial intelligence to sift through this data to identify subtle signs of potential issues. These signs can often be missed by human operators.
When Predict spots a problem, it creates an alert so that engineers can investigate the data. If the evidence confirms that a component is showing signs of degradation, managers can schedule maintenance work at a time that makes sense – both from an operational and economic perspective. This can make a significant difference to the performance of a wind farm, as our recent work with Eneco testifies.
Our work with Eneco
Eneco is a Dutch utility with 3.5GW of energy projects in operation and almost 3,000 employees. Its portfolio is made up of onshore (1.7GW) and offshore wind (427MW), solar (294MW), as well as conventional sources and biomass. In March 2020, it was acquired by Japan’s Mitsubishi Corporation and Chubu Electric in a €4.1bn takeover.
Our work with Eneco begun in 2017, when it started using Greenbyte Platform, and the relationship has grown since then.
The company committed to use our platform on 1GW of its wind farms in early 2018, and now uses it to manage its large wind portfolio. Eneco then became one of the first customers for our add-on product Predict later in 2018, which is the focus of this case study.
It has used Predict at six onshore and offshore wind farms where it has a financial exposure for the cost of lost production. The portfolio includes projects with turbines ranging between 1.5MW and 3MW, and from four different manufacturers. This is in operation on its portfolio alongside our main Greenbyte Platform.
Eneco’s challenge is that it had service contracts with the four wind turbine makers at the six projects, but these contracts did not cover the full cost of lost production. As a result, any downtime was a costly drag on the financial performance of the portfolio.
We used Predict to analyse the turbines’ SCADA (supervisory control and data acquisition) data to anticipate problems and resolve them before components failed.
One of these projects is the 120MW Princess Amalia offshore wind farm in the Dutch North Sea around 23km from the coast of Ijmuiden. The project is made up of 60 2MW Vestas turbines and was commissioned in 2008, in waters as deep as 24m.
It is owned 100% by Eneco and is one of the longest-running offshore wind farms in operation, which means Eneco has 12 years’ experience managing and maintaining the development. This makes it a perfect test case and, as a result, Eneco has been very supportive of our work across the whole portfolio that we analysed.
Marina Tsopela, asset analyst at Eneco, said our analysis had been beneficial for its asset management work: "Predict has helped us get a better insight into our portfolio and stay on top of things that need to be done in order to improve our performance.”
How the analysis project worked
For Eneco, we proved the effectiveness of Predict to anticipate problems before they became failures. Our analysis gave Eneco the opportunity to repair or replace turbine components in a controlled manner, with minimal downtime.
Predict identified 17 wind turbine issues for Eneco before they were critical
Over a six-month period, Predict identified 17 impending component failures and, as a result, it stopped many days of downtime and kept Eneco’s turbines turning.
Our project had two distinct phases:
- Phase 1 involved a retrospective analysis of one year’s worth of past SCADA data. We could then show how and when Predict would have detected past problems.
- Phase 2 was a live six-month project, with Predict running across the fleet of six onshore and offshore wind farms in the Netherlands.
Using data to anticipate problems makes sense for businesses. However, an over-sensitive condition monitoring system can create more problems than it solves, if it encourages owners to take projects offline for investigation when they don’t need it.
Predict avoids this risk by using a network of AI-based models that are trained to learn from SCADA data. This intelligence reduces the occurrence of false-positives and ensures you only need to respond to genuine issues.
A performance assessment of Predict, using a test data set of 162 wind turbines that are operated by multiple companies including Eneco, achieved a precision of 93.8% and a recall rate (percentage of faults that can be detected) of 76.2%.
Predicting the failures
During the two phases of implementation, Predict identified 17 component failures that were verified by maintenance records and inspections by maintenance teams.
The graphic below shows the range of component failures identified by Predict:
Figure 1: Failures anticipated by Predict
Three issues identified by Predict
The value of Predict rests on the potential for avoiding unplanned downtime, and in being able to make minor repairs before major repairs are required. Over the next three pages we will look at three examples of the issues we identified for Eneco.
Example 1: High temperatures in the converter system
Predict detected high temperatures in four elements of a turbine’s converter system.
The chart below shows a comparison of actual and model-estimated temperatures for the one component. You can see that Predict was able to identify an anomaly that is hard to spot with the naked eye.
We notified Eneco of the issue and engineers conducted an inspection of the turbine, but they did not complete a detailed assessment.
Predict alerts continued, and temperatures began to rise. The turbine was eventually stopped, and engineers discovered a clogged IGBT fan that required cleaning.
Figure 2: Estimated and actual temperatures for the converter system
Benefits of early detection:
- Reduced downtime (approximately 9 hours)
- Controlled maintenance
- Reduction in stress and wear through frequent stop/starts
Example 2: High temperatures in hydraulic oil
Predict indicated high hydraulic oil temperatures in one of the turbines.
The actual temperatures were far higher than those estimated (see chart below). Predict had detected a similar issue in another turbine, which was verified using historical data. Eneco was able to notify its maintenance team about the probable cause, which was a problem with the rotating union in the central hydraulic system.
The maintenance crew verified the issue and scheduled proactive maintenance to remedy the problem.
Figure 3: Model-predicted and actual temperatures in the hydraulic oil system
Benefits of early detection:
- Prevented up to six days’ downtime, which was the case for a similar failure
- Avoided breakdown and additional stress on other components
Example 3: Overheating gearbox
Predict detected raised temperatures in a turbine gearbox on 9th August 2018.
This slight deviation was too small to be detected by the SCADA system (see chart below), but Predict had previously detected a similar issue on another customer’s fleet. In the other case, a thermal bypass valve was replaced and the temperatures returned to normal.
Eneco checked the maintenance record for this turbine and found that, in this case, the issue was also connected to a thermal bypass valve.
Figure 4: Comparison of model-predicted and actual gearbox temperatures
Benefits of early detection:
- Prevented around 46 hours of downtime
- Reduce stress on gearbox components
- Extend the life of components and lubrication oil
This case study shows how Predict uses advanced modelling to help owners keep turbines running and boost their returns. If you protect parts, you protect profits.
Rather than responding to breakdowns, Predict gives you the chance to be proactive in your O&M strategy. Instead of rushing to get stopped turbines back into action, you can schedule repairs and maintenance at a time that is safe and cost-effective.
And by learning from other wind farms using Greenbyte platform, Predict can anticipate many of the issues that typically reduce profitability. This means that Predict is particularly effective at early detection of chronic issues – the kinds of problems that are easy to fix, but recur again and again, and account for a considerable portion of lost production.
While SCADA systems typically produce alerts, there are occasions when these are not activated.
For example, during periods of low wind, the turbine may not reach full power, and therefore the temperature may never exceed the limit set in SCADA. This illustrates how issues can go undetected for long periods, which causes projects to operate at reduced efficiency; reduces the lifetime of costly components; and eventually leads to a long downtime due to unplanned maintenance.
However, these would be picked up by Greenbyte Predict.
Do you want to know how Predict could improve the productivity and cost-effectiveness of your wind farm? Contact us for more information.