How Collaborative Modeling with Subject Matter Experts Delivers More Robust Insights
On the Wind Energy team at Uptake, we frequently examine 1000+ data signals in order to create models that help wind turbines run more efficiently and prevent failures. While having many signals is great, some can be deceptive. As a result, it can sometimes be inefficient to rely solely on raw machine learning to understand what these signals are trying to tell us.
Instead, we work side by side with internal and client-side wind industry experts to create machine learning models that raise actionable insights. One of the models that came from this approach was able to catch a wind turbine that was operating 45 degrees away from the wind, incurring a massive power loss!
Let’s take a closer look at how we did this.
A closer look at defective wind vanes
Wind turbines generate the most power when they are facing directly into the wind. Even turning a few degrees away from the wind will cause the power generation to plummet. Typically this is caused by a defective wind vane.
What is a wind vane?
A wind vane is a device that sits on top of the turbine and measures where the wind is coming from so that it can face the wind and generate power. If the wind vane becomes defective, it can lie and tell the turbine that the wind direction is different from actuality. This, besides resulting in enormous power loss, can also put unnecessary strain on the rest of the turbine, reducing the life-span of its components.
Below we depict an aerial view of two wind turbines. The turbine on the left has a good wind vane and is facing directly into the wind. The turbine on the right has a defective wind vane which is causing it to think it is facing into the wind, but in fact, it’s turned away and losing power.
As we set about creating a model to detect turbines that struggle to point into the wind, we faced several challenges.
- We had almost no known examples of this issue occurring in the past. This is not because they don’t happen, but rather because they are difficult to catch. A turbine experiencing this issue usually appears to be operating normally, since it believes that it is actually pointing into the wind. Because of this, these problems are typically uncovered by chance. For instance, a wind technician could be driving around a wind farm for a different reason and notice that one turbine is pointed in a different direction from its neighbor.
- Data that comes off of a wind turbine’s sensors tends to be inaccurate in a variety of ways. One of which is that its internal compass — which sets its true north and guides its direction — is frequently miscalibrated, and can even change erroneously over time.
Complex challenges in data quality require more robust machine learning models
Here we show simulated data from two turbines with properly calibrated internal compasses.
What is plotted is the sensor-reported direction of two neighboring turbines. Each point represents the turbine’s direction at a given point in time, and the y- and x-axes correspond to turbines A and B respectively.
Since these turbines are close to each other, they should be experiencing the same wind coming from the same direction. Therefore we would expect both turbines to face the same direction, and thus in the plot, form a line with a slope of 1.
That is an ideal scenario, however, we often observe data that looks more similar to the image below.
Here we show simulated data of two turbines where at least one has a miscalibrated compass.
Let’s first talk about those horizontal and vertical lines. Those indicate that a turbine is not moving when its neighbor is moving. There are many reasons why this happens. For example, sometimes a wind vane gets frozen in place due to ice and will tell a turbine not to turn.
Now if we focus on the claw-like diagonal lines, we see that they still roughly have a slope of 1, however, instead of one line there are now multiple. This is due to shifts in the internal calibration of the compass over time. One reason for these shifts might be due to maintenance or repair work. This suggests that the turbine is moving in the same way as its neighbor and actually has an okay wind vane. However, our notions of normalcy in these two images are drastically different.
While we could potentially pursue reconstructing the turbine’s true direction signal, our wind experts encouraged us to look for solutions that didn’t rely on having an accurately calibrated sensor. Our SMEs knew that while some wind farms had more reliable sensors, we shouldn’t expect all farms to be equipped with those sensors.
By following their advice and then utilizing a few of their battle-hardened heuristics in our AI Engines, we were able to create a more robust model. This model incorporates a feature which represents the change in direction of the turbine, rather than the absolute direction of the turbine. Furthermore, given our lack of labels, our SMEs quickly validated our model outputs and gave us a higher degree of confidence that our insights made sense and were actionable.
In conclusion
As Data Scientists, we usually search for a way to deduce the ground truth by using the data at-hand to recreate exactly what is going on; in our case, within a wind turbine. The expectation is that, with this crystal clear image of what is happening at every moment, we can more easily train our models to figure out what is normal and what is not normal. However, by collaborating with subject matter experts, we realized that this wasn’t just difficult, it was counterproductive. Trying to recreate a turbine’s actual yaw direction would have taken us down an unnecessarily long path to our original goal. Instead, we re-calibrated our approach to be more robust rather than more refined to account for the often unreliable data that is ubiquitous in the industrial internet of things.
