How artificial intelligence will improve O&M

Artificial intelligence is being applied to almost every industry in efforts to improve operations and trim costs. Here’s how early efforts are already benefitting the wind industry. 

Rob Budny / VP, Reliability Engineering

Sandeep Gupta / CEO

Ensemble Energy, ensembleenergy.ai

The world is entering the early stages of a technology revolution called artificial intelligence (AI). It is showing an impact in many different fields such as image recognition, fraud detection, and self-driving cars, to name a few. Machine learning techniques have resulted in remarkable performance improvements in each field to which it has been applied.
The best definition of artificial intelligence is that it is set of methods or algorithms that use a large amount of data to learn rules or patterns. Another aspect of AI is that it continuously improves with additional data and without being explicitly programmed to do so. Even though AI is considered the broader concept, machine learning, and artificial intelligence are often used interchangeably.
The use of artificial intelligence requires three basic elements:

  • Learning algorithms,
  • Large datasets, and
  • Large-scale, inexpensive computing

AI dates back as far as the 1970’s but has only recently become practical due to low cost, powerful, cloud computing. The wind industry is well suited to benefit from this technology revolution. The best way to do so combines wind-turbine engineering and operations expertise with the latest artificial-intelligence methods.

However, AI cannot fully deliver the technology’s potential benefits without a complete understanding of wind-turbine loads, control strategies, and component-failure modes.

The shortcomings of conventional O&M
Today’s wind turbine operations-and-maintenance work is less effective than it could be because it is:

  • Reactive − Maintenance actions are taken in response to faults or failures.
  • Static − Turbine behaviors have fixed upper and lower fault limits, which are not situation dependent, and are not customized to the known behavior of each turbine.
  • Labor Intensive − Developing and running SCADA queries takes many hours of time which strain resources. Useful insights in the data are overlooked, resulting in lost opportunities to increase production or reduce costs.

Wind farm O&M with input from artificial intelligence will pay off several ways. For instance, AI will be:

  • Predictive − Anomalies will be identified in early stages and addressed before faults or failures occur.
  • Dynamic − Turbines will have dynamic fault limits that will be situation dependent and can be customized for each turbine.
  • Automated – A machine learning platform continually analyzes data in the background, and provides automated notifications to operators, along with suggestions for the most effective corrective actions. Skilled human operators will be freed from tedious data analyses, and their time will be spent on higher value activities.

Improving O&M

The cage segment comes from a failed pitch bearing.

A few examples can show how our company is using the best combination of machine learning and physics to improve wind-turbine operations and maintenance. For example, the premature failure of pitch bearings is just one issue facing many wind operators. Several of their different failure modes include false brinelling, macropitting, cracking of the bearing’s outer ring, and failure of the cage that separates the bearing’s rolling elements. An accompanying photo shows the cage segment from a failed pitch bearing.

Repairing these failures is extremely expensive and leads to downtime and lost energy. They can also be dangerous because they have resulted in the loss of a blade. One client company experiencing pitch-bearing failures had been relying on time-consuming and costly visual inspections to identify failing units. We developed an alternative detection method by combining our wind-turbine expertise with the latest machine learning techniques in a model of expected pitch-bearing behavior under all turbine operating conditions.

The AI platform we have developed continually monitors turbine operation in the background, so it needs no operator input. Recently, the platform identified a deviation between the expected pitch-bearing behavior and actual behavior which triggered the sending of a notification. When the pitch bearing was inspected, its cage failure was spotted. For this AI system, the average early notification time for pitch bearing failure has been about four months prior to the bearing needing replacement.  This lets the operator plan the replacement work with a similar task on another turbine, thereby saving one crane mobilization charge, almost $100,000.

Preventing main bearing failures provides another example of how machine learning benefits the wind industry. Premature failure of main bearings is, unfortunately, a widespread problem in the wind industry, and is one of the most expensive unplanned maintenance events, costing up to $250,000 per failure.

While there are several root causes of main-bearing failure, including bearing design issues and excessive rotor thrust, poor lubrication is one of the most important root causes. Poor lubrication comes from by an insufficient amount of grease in the bearing, a lack of grease in the correct locations, or by grease that has thickened and is no longer capable of lubricating the bearing. An accompanying image shows an example of a lubricant related, main-bearing failure.

This main bearing failed because its insufficient lubrication went undetected.

Our proprietary methods were used by another wind-farm owner to confirm that the lubrication anomaly in a main bearing would have been detected almost six months prior to failure. The methods include a combination of expertise in wind-turbine loads, bearing operation and lubrication, and advanced data analysis. Had our predictive analytics platform been in use, the operator could have performed a simple, inexpensive maintenance action that would be prevented, or at least significantly delayed the failure.

The bar chart Health scores shows those for a main bearing compiled over almost a year. When the health score reaches a pre-predetermine value due, the system sends an alert along with a recommendation to add grease to the bearing.

Health scores for a main bearing

If adding grease to the bearing corrects the condition, the alert is cleared. However, if adding grease does not correct the condition, the alert escalates so an operator can purge the bearing of hardened grease and replace it with fresh lubricant.

In this example, notice that an abnormal condition (indicated by falling health scores) was identified at least three months before significant main bearing damage occurred. No fault was generated by the turbine, even though the behavior was abnormal. Predictive actions such as these can effectively prevent failures, or significantly extend the life of critical components, resulting in large cost avoidances for operators.

Anomaly detection provides another example of how machine learning applies to wind turbines. Detecting underperforming turbines is notoriously difficult using the power curve that OEMs provide. Although turbine manufacturers publish reference power curves, the actual power produced by a turbine is affected by factors other than wind speed. Such factors include site elevation, seasonal factors, wind shear, turbulence intensity, and more. Furthermore, the power curve for similar individual turbines from the same OEM can vary by as much as 10%.

Peer-to-peer analysis provides another performance evaluator. Although somewhat more effective than comparing to a reference power curve, the turbine peer-to-peer method requires a careful consideration and understanding of turbine positioning and wind regimes. The Ensemble Energy approach allows for both peer-to-peer comparison and individual turbine performance deviation.

The curve in Predicted power shows a power curve created by our machine-learning platform. The model incorporates many factors in addition to wind speed and accounts for their effect on power production. Unique models are created for each individual turbine. This model allows identifying a deviation in power production as soon as the same day the deviation begins, letting owners take immediate corrective action, and limit lost production.

 

Machine learning techniques, when created and applied in combination with domain expertise, are resulting in increased energy production and reduced maintenance costs. Furthermore, these improvements are being made by making better use of existing SCADA data, with no additional sensors required. The use of machine learning to increase energy production and reduce costs will be the standard in the near future.

Speak Your Mind

*