Pioneers Talk #2: Meeting EAF Performance Targets Using Data Analytics

Thanks for inviting me today.

When we go to the plant, we make the startup and everything, we have to meet the performance values. For these performance values we create the optimal steel making practice which is based on the materials we have at that moment. We have scrap, dolo, lime and all these things and they can change over time. So, if these things change because you have more scrap qualities, your supplier for your electrodes changes, you have different qualities of lime or dolo, you also need to tweak a little bit the steel making practice to stay at the optimum and from our experience this might not always happen. The responsible operator for it is usually loaded with other work. For example, they are dealing with quality of scrap, quality of the steel, of the final steel and such things so it’s just a small part which can make a big difference on the outcome.

We got some sample data and it was really interesting. We found that, for example, if you have the first basket just a few minutes longer and also the second basket a few minutes longer, it was actually first basket was one and a half minutes long and the second basket was half a minute longer, in total two minutes and then at the third basket we could save four minutes. For us it was kind of surprising that good timing really makes such a difference in the tap-to-tap time or the power-on time, depending on what you’re focusing on.

I think the operators are aware of that. Everybody is aware that the good timing of the baskets is really crucial for the success of a heat. But every heat is kind of different, you have different scrap recipes and all that and so it’s also for the operator pretty difficult to find the perfect spot when to charge and when not to charge. Timing is crucial even for gas, oxygen, and the slag builders, so yes, timing is crucial.

Usually for other aggregates, most of the time you know what is coming. If you have for example an EAF route followed by a ladle furnace, you have a sample of the EAF and you know what’s going into the EAF. There’s no surprise there. On the other hand, if you put scrap into an EAF, there might be some surprises, so it’s very difficult to get a real model there which is always working perfectly and that’s one of the things we thought we want to solve.

The diagram you can see shows that at the beginning we are clustering this scrap, it’s not really rocket science, we cluster the scrap according to quality and weight. We get this cluster where we know that we have certain scrap qualities. It’s just clustering of what we know is going in to the EAF and we then find for each cluster the reference heat and we think reference heat is one of the most important steps because it defines the outcome of our search. For example, if you want to increase productivity of your EAF this reference heat should also reflect that productivity approach, it should be a highly productive heat. On the other hand, if you want to improve for example costs or something like that the reference heat should be a very cost-efficient heat and the other steps also three and four are usually there to check if this heat is repeatable. At step three, we look at the aggregated data, one would say level two data or reporting data something like that, so we find these potential heats where these gold nuggets or these similar heats are and with that potential heat or with that potential pool of heats, we really go into the time series of data. We check when did the operator push the carbon button or when was the gas or oxygen activated and we then compare it to find similarities, where we find this last cluster and then we can see that this reference heat was produced 10 or 15 times and it’s realistic that we can reproduce that. From that we extract the best practice and at the end it’s more or less just a table which you can hand to the operator and try it and check the outcome.

You always have to prove what you’re doing with the data, if you can prove that the steel producer gains an advantage it’s usually pretty easy. But when you say, I want data to just have data and nobody knows what is done with the data it really gets difficult and so we need a transparent approach that we really can check what is done with the data step for step so everybody can follow it.

Usually we don’t care how we get the data, if we connect to a data lake or to an SQL database where everything is really nicely structured. We have a very flexible approach. We need to get the data in some way, then we transform the data ourselves or the customer can already do it if you can identify what they are interested in and if they think there are some data signals which are very important, they can spare it out and that’s it. The only thing we really want is that the data quality should be okay and the second thing is it should be really comprehensive, most of the data should be in there so that the results can be used.

Sure, you have to talk. There’s no way around that.

There are some manual steps, for example defining the reference heat, but the search steps are more or less automated. We can tweak it a little bit, for example if electric consumption is more important, we want to focus a little bit on that and put some weight on it. The last search steps then are automated.

We know how to build plants and our customers are good at operating plants, so we don’t want to have some kind of model in that direction which is predicting something with an accuracy of 95 or 99 percent, which it’s pretty good. But that’s not what we want because you have a five percent failure rate. We think that our customers are really good at operating plants and so we just look into the data and check what happened in the past. What we extract from there is not something which we fantasize about but it’s just a little help for our customers to see what they have done in the past and what they have done well in the past and how they can keep this up in the future.

That’s steel making practice, you’re still at level two and steel making practice is in the background and it tells the operator when to push the buttons, it just tells to push the buttons and the operator can decide otherwise. The thing is that steel making practice is often not adapted.

Yes, because it worked in the past, so why shouldn’t it work now.

In my opinion, EAF must have some kind of targets. For example, if steel prices are high you need to produce as much as possible to get a lot of profit out of it. You want to focus on productivity and that’s what we do then with the reference heat. On the other side, if the steel prices are low, you want to produce as cost-efficient as possible to get at least some profit out of it and you don’t care about how much you produce because steel prices are low and you just want to be cost-efficient to still have some profit. That’s what we try to achieve with our reference heat and what we can achieve with our tool. We can easily switch that, so we can extract these SMPs, steel making practices, within one or one and a half days and then we can say hey here’s your new plan and you can switch the target which adds some flexibility.

Seasoned operators will retire at one point and new operators will need to learn this over the time. Operators in my experience just get better over time and with experience and so if this one operator you talked about is really good, we will find his heat in the data and we will make it a best practice. We will extract his knowledge on how to process with certain steel grades and deliver it to all the operators, not just one.

That’s a good question, probably yes. If there are good heats of one operator who was good at cost-efficiency, who retired. So, if you change the target of your EAF from the last three years where you focused on productivity and you find this data from five years ago, then you can renew it and put it back into operation again.

Fuchs Finger Shaft was the predecessor of the EAF Quantum.

I know that the finger shaft is still used but I’m not sure how much hot heel it has. The thing with hot heel is that it influences the heat, the outcome of the heat a lot and if there is a constant hot heel, we can apply that.

Unfortunately, we haven’t managed to put it into practice yet. We got some real data, we made our studies on that and one of the reasons we are doing this today is to find EAF operators who would be interested in working together with us to see how much it really takes or how much you can really get out of it?

It depends on the data. Usually electrodes are weighted, depending on the operational practice, once a week or once a day maybe. It’s pretty hard to really say which heat had which electrodes consumption. I wouldn’t say it’s impossible but with our method it’s probably not that feasible because we’re checking the heats and if you just have an average electric consumption, I don’t think it will work.

It’s really hard to have an automation concept for the reference heat because it depends on the target of the EAF. You can also have the clustering done automatically and we have automated the search algorithms as well as the data extraction. We finished that three weeks ago and it worked really well. For the reference heat though, it’s one of the things where you still need a human brain and an expert who evaluates the path we want to go with.

If you look at level one data for three years, for every two seconds you have one row. So, you have millions of rows that you have to automate because otherwise it’s not possible. We also have these two steps, because to compare the time series with tens of thousands of sheets of time series data you need a pc cluster and big pc machine and so we have this two-search approach to get the potentials and then the time series.

Neural networks create a model with which they predict something and that’s not the path we wanted to go. If you have a model for predictions, you always have this error margin and in situations, with which they are not familiar, they’re predicting sometimes absurd things. That is why we have statistical models which are based on the past and we extracted data from those and then we use machine learning algorithms.

As I said whether it’s a data lake or an excel file, we can deal with it. We have to put it in one place at some point. That’s not a problem if we have different data sources because we just put them in one place and then we continue to work with that.

We will put it into a data warehouse, our own data warehouse, and then we put our solution on top of the data warehouse.

Thanks, goodbye.