The Greenbyte recipe for Artificial Intelligence in renewable energy — Part one: The Journey
By Pramod Bangalore and Mikael Baros
This article series intends to explain how we at Greenbyte use Artificial Intelligence to provide deep insights for operators of wind and solar assets. In part one, we reminisce about our journey that got us to where we are now. Part two will explain how we make our AI services be available 24/7, continuously updated, and constantly expanding with our CI/CD pipeline. Part three will go in depth into some of the more advanced aspects and experiments we are running right now.
A humbling experience
I remember the first time I met Pramod Bangalore back in early 2016. I had just accepted the job at Greenbyte — an impressive-looking startup with dizzying ambitions. Having just come out of a meeting with my future engineering team, I was already elated about this amazing company and the raw talent of its engineers — my future co-workers. Before talking to Pramod (my future head of research) I was both excited and nervous, as I had earlier experienced companies toting “we work with artificial intelligence” and then showing only paper ideas, or even worse, a simple statistical model. It took about five minutes before I was blown away by his research and the potential of this technology, sitting there with my mouth agape while listening to the results he had already observed, on real data.
Fast forward two years.
Talking about both machine learning and artificial intelligence has become ubiquitous in many industries today, and there is a lot of hype about the possibilities. Although significant progress has been made in this area, large-scale application is still mostly limited, especially in the relatively young industry of wind power. At Greenbyte we are constantly striving to improve the value proposition for our customers and one way to do this is by providing better insights derived from the data. So back in 2016, it was only natural to turn towards artificial intelligence. We wanted to create an analysis system which would tirelessly crunch data and provide our customers with insights that are impossible to gain through human analysis, and lead to significant cost savings.
Reality check
Wind turbines are wonderfully complex machines that face the wrath of mother nature day in and day out. I mean, if we pushed our cars as hard as we push our wind turbines, they would fall apart. The average car lifespan, according to this paper, is 15,6 years. Since we, on average, drive our cars an hour a day this transforms into (with some incredibly simple math and assumptions) 5616 hours of lifetime¹. We expect turbines to operate 24 hours a day, 7 days a week. If we did the same with a car it would only last us 8 months! Hence it is not surprising that these poor turbines fail (too) often. Frequent failures in wind turbines cause issues for the operators and owners, as they must send in technicians to fix the faults and lose money due to lost production. It is estimated that up to 30% of the total life-cycle cost of a wind farm is due to failure and maintenance activities.² An alert about impending failures in the turbine components can go a long way towards saving money and time for the wind turbine owners and operators. Imagine if someone could tell you three months in advance that your car’s rear axle will bend out of shape and break; what if you knew you should hammer it straight then and there?
Like modern cars, wind turbines are equipped with many sensors which measure and record various operating parameters. A small wind farm of 20 turbines creates about half a million data samples in one day. It would be unreasonable to expect a human to be able to make sense of all this data. Sadly, all this juicy data goes unused and any analysis performed is, mostly, retrospective. However, the human brain is a miracle of nature and is capable of amazing computations in a matter of milliseconds. So, the question is, can we utilize the human brain like characteristics while not asking a human to do the data crunching.
This is where artificial intelligence and neural networks come in.
At Greenbyte we have developed an Artificial Neural Network (ANN)-based system which monitors the health of various components in their wind turbines and informs the owners and operators about issues that need attention. After the initial research efforts, it took us over 2 years of late-night development and rigorous testing, but it was worth it; as we have now achieved a 95% system accuracy! This means that the current system has a maximum of one false alarm per year per 60 wind turbines; and we are still not satisfied.
Teaching for success
An ANN model starts out empty, with no inherent knowledge. It needs to be taught the task that it needs to perform. Just like a good teacher always finds creative ways of making information stick in our brains, the quality of learning is important for an ANN model as well. Unfortunately, information delivered by the wind turbines is not always a good teacher. The data often has gaps and garbage values that we do not want influencing the model’s learning. An ANN model trained with incorrect data is as useless as an ejection seat in a helicopter. We realized this early on in the development phase, and at the time it seemed like an insurmountable problem. So, we set out to create the best learning experience for our newborn ANN. All the brilliant people in the Order of the Dolphin (that is the name of our research team) dove in with enthusiasm and came up with an elegantly simple three-pronged approach. It is funny how solutions to difficult problems always seem simple once you know of them. We ended up with a concoction of hard thresholds, deep statistical analysis, and mathematical data clustering to remove unwanted data and create a good teacher for our ANN model. As we had thought, a good teacher really does make for a better learning process.
Learning to fail
Once our neural network model child had matured into an intelligent youth, it was time to send it out into the real world to perform at the highest level. However, we soon realized that the perilous real world is very different from comfy classrooms. The quality of data again caused issues which, disappointingly, led to poor performance. The tiny AI brain needed to be able to generalize well, which meant it should be able to create reasonably good estimates for a new set of inputs, something it had not seen before during training. To achieve this we created a ranking system where several models are trained and only the best performing model in the group is selected for the application; a selective breeding if you will. Finally, after long days of painstaking code modification and strict test conditions, we were ready with our well-trained ANN model 2.0. This time around we overcame our expectations in achieving the desired performance level for our models in a real-world application, and we couldn’t have been more proud.
We now had our intelligent and trained ANN models eager to start processing huge amounts of data. However, we still needed to address the issue of conveying the right information to our users. Simply reporting the difference between the ANN model output and the measured value is not sufficient to provide the required level of insight. It is like saying that there is a problem in your car without being specific (like that little yellow engine light), you will have to do more analysis to find the real problem (visit a mechanic). The Dolphins came up with a post-processing approach using a denoising technique called wavelet analysis. With it, we filter out unimportant information and only present relevant insights to our users so that they don’t have to spend time on analyzing the output. Now we had our artificially intelligent system completely specialized for its job - crunching wind turbine data and reporting back when a component in the wind turbine was deteriorating.
The here and now
Today, we have verified the performance of the AI based condition assessment system for about 200 wind turbines and a wide variety of wind turbine makes. We call the feature “Predict”, short and sweet. The entire business has shown tremendous interest in our new offering and we went live on April 1st 2018 — not a joke. Our initial offering will show the list of alerts in the turbine and tell the users which component is showing potential issues. The next step with our Predict feature is to make it further permeate our entire wind farm management system. Everything from linking ANN alarms with Tasks, and improved visualization, to taxonomy information and Breeze API support for integration with other systems. As we operate a SaaS service, all future improvements to Predict will be continuously delivered to all customers that have added Predict to their subscription, which means we have to work hard on our distribution and deployment pipeline. This is a topic for the second part of this series.
All of us are delighted to send our child out into the world of renewable energy, eager to see it succeed. To paraphrase the late Dr. Stephen Hawking, the future only exists in a spectrum of possibility; we are staring down the future right now, and what we see is very exciting.
References
- American Driving Survey 2015–2016
https://aaafoundation.org/wp-content/uploads/2018/02/18-0019_AAAFTS-ADS-Research-Brief.pdf - Operation and Maintenance Costs of Wind Generated Power
https://www.wind-energy-the-facts.org/operation-and-maintenance-costs-of-wind-generated-power.html
