K. The Tobin, Germany, Executive discussed on Data Engineering Podcast
And up to now or maybe the past up to the past two years. Nobody really analyzed the data. These these huge data sets so another thing that we found out this that sometimes the turbine is. Operating in a throttled mode at that nobody knows about so. Sometimes because of regulations because of noise regulations, the Tobin should not produce much power, or is producing less power than actually could end sometimes. These towers go into these noise modes. Without anybody knowing it, and so we figured out. Okay, let's let's do some general anonymisation of the normal behavior of turbine, and let's look if there's something that we can find with. Tobin is not behaving in a normal way, and that's like that's totally a data analytics problem. We don't really maybe have all of the domain knowledge of one particular turbine how it should turn on how to behave, but we can look at the data and see and look for a non APP normalities, and with that example that I was talking about. You can understand like if the tone is producing half the energy that it could. Then, of course this is a huge factor economic sector. If you find these data points in these these that are not producing not energy than than you have value that you can give to your customers, and so you mentioned that at least up until the last couple of years that a lot of this data that was being collected with the systems that are embedded into the turbines, is being ignored or not analyzed any great detail on wondering what the current state of the art is as far as being able to analyze the performance of the turbines and correct for. And do any sort of preventive maintenance to reduce downtime. So up to now their standard, at least in Germany, to have ten minute average values of different measurements at the turbine for instance wind speed. Power output of the turbine, and so that's the standard and basically did. The state has been locked in the past, just because of regulations for instance like if if the Tobin is shut down because of too much energy grid than this in in this case, you have the data to to see and locate. There has been such amount of wind before this event, and you would have generated zone so much energy because of this grid shutdown. That's why basically maybe people were logging data, but now. People also under start understanding that you can, you can do more with data so also more data's locked in newer turbines, and there are more sensors, and the census potentially can not only locked ever values, but also maybe second values or sub second values. So potentially you can get more data than you could get maybe in the past. And yachts like it's a physical system. The turbine is a is a machine and you can. You can grab a topic and then look into detail in look at the into the data and see if you can optimize something there, so yup, so. Just to give that example again to where you can reduce the the power. Whether whether the power of the Tobin is reduced because of some regulations, or because nobody's noticing it. Yeah, maybe I can explain it more how we do it so. We basically try to find data sets that we definitely know about that. The trump is behaving in a good way, so we built out these data sets and. We call them our training data set. And then we train networks on this data set, and we have to think about okay. What physical system makes sense like? What is the input of debt blackbox formula in what's the output and the input for the power output can be of course, the wind speeds, but the energy that's contained in the in the wind is also dependent on density of the air and density is dependent on the temperature for instance, so if you have a value. Time series may be off wind, speed and temperature of the outside air than you can use these two values as an input to generate the power to to simulate the power output. Data said way no K. The tobin is behaving correctly. You can train a neural network on that behavior, and then you can simulate with new data sets how that Toan should have behaved in that scenario and that physical scenario, and then you can make comparisons you can add some more inflammation, status logs and other. European detente service data from the from maintenance companies and makes everything. Everything together and create a value out of that, and then as far as the type of data that you're able to access from the sensors and the control systems in the turbine, and what are some of the challenges that you're dealing with as far as just the data collection, and what is the level of variability between different turbines and different manufacturers, being able to account for? Eighty in the levels of accuracy or the representations of data, as it's coming out of those control systems, yeah. For Good Right, there have been companies on the market that have had specialized executive for that problem because. Every turbine, somehow prototype. Because if maybe you, let's say you buy a turbine from manufacturer a and you put it in your site. Specific site and then you have an additional contracts were data management with another company, and so you can imagine how many potential variations of combinations of manufacturers data data-collecting computers there are on the market, and the means the huge variety of of the data sets, so we also had to learn that in the beginning, and you cannot assume that that if you wanted to type, the data is looking always the same because he knows if it's been generated by the same type of system, so the best way to deal with the promise to to look at each turbine as one system and not make cross correlations too early with with. Let's say if you have one turbine. Type and you want to make cross collisions with many other turbine types. Of the same model. Sorry so the same tobin-type INCR-. Make Cross collisions over that. You you better you. You better set if you have likes each and every time. A specific model end that also means again that you have a lot of. Machine learning models that you have a lot of data that you need to train. There's a lot of skill ability problems. Let's say that you.