Welcome to the Pipe Liners Podcast, Episode 310, sponsored by Inter-SYS Corporation, providers of Poems, the Pipeline Operations Excellence Management System, compliance and operations software for the Pipeline Control Center to address control, re-management, SCADA, and audit readiness. Find out more about Poems at Inter-SYS Corp.com. The Pipe Liners Podcast, where professionals, bubba-geeks, and industry insiders share their knowledge and experience about technology, projects, and pipeline operations. And now your host, Russell Treet.
Thanks for listening to the Pipe Liners Podcast. I appreciate you taking the time and to show that appreciation we are giving away a customized Yeti Tom Wert to One Listener every episode. This week, our winner is Tom Correll with Northern Natural Gas. Congratulations, Tom.
You're Yeti's on its way to learn how you can win this signature prize. Stick around until the end of the episode. This week, we speak to Vicki Knott with Crux OCM about model predictive control and its application to pipelining. Hey, Vicki, welcome back to the Pipe Liners Podcast.
Thanks for having me, Doc. It's always fun. I've asked Vicki back to talk about model predictive control adapted to pipelining. So, maybe as a way to build background Vicki for the listeners.
Let's just start. If you would just tell us through my list, who you are and what you do if you don't mind. Yeah, sure. So, Vaym is Vicki Knott, CEO and co-founder of the company Crax OCM.
By background, chemical engineering was the process control. And then I trained as a controller operator on the biggest batch pipeline in the world. And then my co-founder Roger Shurk, who's done done lots of work up north. And we came through a realization and understanding that how we operate pipelines today and how as a controller operator, it's busy.
It's a lot of work. It's quite challenging. And we need the comparison to pilots and planes of autopilot software. Our pipeline operators are pilots for our critical energy infrastructure and then having to help them be as successful as possible.
So, you're doing autopilot for pipeline control? Yeah. Yeah, it's a really good way to summarize it. So, a lot of things you use autopilot for, there's a lot of things you don't use autopilot.
Exactly. And it's like, oh, it's like, you don't have any remissive. You need to be in control. You're still driving the car around there yet.
But it's an it's handy when those things get in. Yeah, absolutely. Absolutely. So, to talk about model predictive control is probably a good idea to just build.
So, most pipelines are familiar with the idea of automated control. You've got a skater system. I've got PLCs in the field. I've got pressure control.
I've got flow control. Most of us are familiar with that. So, there's a term that's out there that's called advanced process control. What is it?
And why are we not historically done it in pipeline? What is it? Start with that. So, what is it?
Actually, I bet you almost everyone who's listening to this will be like, wait, I know what that is. I've heard of it. Because they probably work at Midstream companies that own refining, right? Or overfineries.
Inside of their findings, you know, a unit like a hydrocracker or a legislation column will typically have advanced process control capabilities on inside the DCS system and it'll be configured. And what it's doing is it's targeting, you know, an optimal operation. And it's a series of complex mathematical equations, you know, differential and an integral. But what it's effectively doing is it's constantly driving to optimal, right?
So, if you think of refining in our chemical industries, you know, and if you were to do a quick Google advanced process control and refining, what is the improvement? It's about one to three or one five percent, you know, depending on the source you're looking at. So, it's okay, this is really, really proven in Pullman paper refining, you know, chemical, desk plans. So, it makes sense to me that in these process industries, we would be focused on optimizing because they're commodity industries and they operate on razor thin margins.
So, any savings to cost directly goes to the bottom line. So, there's a lot of opportunity to optimize at kind of the incremental functional level. And not just cost too, right? It's also, you know, if you think about how these assets are financed and built, right?
Like, there's an assumption on the performance. So, when you get above and beyond that, it's actually really high margin revenue for the organizations, which is quite nice. Yeah. And I think, you know, when you start talking about it's algorithm equations and differential, you know, some of those, some of those are fairly complex, because they're actually, you know, they're modeling what's occurring chemically or flow or whatever in those units.
And the reason it tends to be focused on a unit is each of those units is running a different process. So, they're optimizing every process in a chain. So, that's kind of artificial intelligence. It's kind of machine learning.
And it's, there's a lot of similarity, but how is that different than artificial intelligence or machine learning? Excellent. So, yeah. So, artificial intelligence, you know, whether those mean neural nets, right?
Like, however you're training the, the, the, the massive AI model, when you're training the, when you're training an AI model, you're typically looking at, you know, trying to automate or create a way to eliminate a human task, right? So, we think about the process industry and those operations, probably think about how humans operate them. You know, we can look up many studies on the inconsistencies of control operators, right? So, we would put all the information into AI.
We will have created a fantastic average before, right? So, that's great. You know, that's fine. But then if you were spending significant, you know, capital on that investment to create those models, you want to see hard ROI returns, right?
Like, you want to see that revenue uplift. So, in order to do that, you actually need to be driving towards optimal operation, not average. And that's where advanced process control comes in, because that's what it does. It's not just recreating an average of humans.
Yeah, you go down, yeah, I think another way to think about that is what you're really doing is you're returning to the base science. Yes. And you're starting to get to the base science, right? And then going from there, like, you know, what does the science say the best to do is, and then you're just running that math real fast over and over and over again?
Way faster than a human can, right? Like, you can't run it that fast. And so, because of that, you know, you always have this time leg and the time leg is where the value is lost, right? Because the tire you can make that, which all the folks from refining know, that's where the opportunities are.
Right, right, exactly. So, what is model predictive control and how is it distinct or different than advanced process control? Oh, it's a classification of. So within a vast process control, it's considered like a group of methodologies, right?
So within the group of methodologies of using more advanced types of process control, which is, you know, common in plants that are like in mechanical plants and refining. So you'll see like the IEEE lecture angle of control. So inside the control, inside that triangle, you know, would be a subset of advanced process control is model predictive. It's the most commonly used in refining effectively, you know, internally taking first principles, driving towards an optimal operation and then continually just reiterating that as the process is changing around it.
And so that's how it's able to execute more commands at a higher frequency than what a human can do and that's how you're getting that lift. Okay, so yeah, so that's interesting. You know, when I think about where we are as an industry and how we think about things like AI and machine learning and what we're hearing about what we ought to be doing, right? Where I go is we're probably not thinking deeply enough, right?
Because for example, if I have a scale system and I've got automation in the field and I'm doing the things that I need to do, then I can apply AI to the pressure control algorithm and have it work more smoothly, more effectively, right? But if I take what a human does and I try to just automate what a human does and I don't go back to first principles, then I'm missing something. So what would you be your commentary about kind of the whole mess of stuff I just said there? So I think the commentary is, or I think the easiest way maybe for, you know, pipe liners to think about it is like, okay, you know, we've always done things this way.
We think of it as a logistics system, not a process. You know, let's look at our friends that are probably our organization that are in refining and like, huh, interesting. Like the way that they treat it is a process, not a logistics problem. So if you take that thinking and you think of a pipeline network as as a process problem and our logistics problem, it becomes very different in terms of how you grasp, you know, what would be useful with regards to AI versus, you know, certain like algorithms that try to achieve optimals, right?
Does that summarize a little or weird? No, it's really good because it makes me think about how I would take and break down a pipeline. And when you think about a pipeline as a process, that's a different way to think about it than I'm just trying to get product from point A to point B, right? What I'm really trying to do is I'm trying to move product from point A to point B and trying to do that without exceeding any of my safety constraints.
And I'm trying to do that by minimizing my power consumption to support that movement and minimize the labor hours required to support that movement. And you want confidence that you want more selling to your shippers that you're able to sell them more and you want the measurable confidence that that volume is actually there. Oh, yeah, that's so true. Because if I have a pipeline and I'm running at 80% of optimal, yeah, and you think you're at optimal, but you could run at 83.
The first principles and I reevaluate what I'm doing. Oh, no, you're not really running. You could be moving another 10% of your product, you could reduce your electric consumption by 20%. Yeah, so you can do it some days, like, how do you think goes right to the bottom line?
Yeah, yeah, yeah. And that's like in our production installations of this, it's exactly that. Like we are driving high draw max recalculated every 10 seconds. And at the same time, operating at a lean profile, which is showing about 10% power cost savings.
So it's like it can be done. It's just that humans right now, we have a human in the loop constantly computing their best and they kind of their head and we just can't keep that best. So let's use a try and drew algorithm that's in industries that we're comfortable with, right? So I think we kind of, so my next question is what's the value of taking a proven refinery technology and implementing in a pipeline?
I think we've kind of addressed that, but have we missed anything? I mean, what is that value beyond moving to optimal operations? Moving to optimal, I think it's also the big reduction, definitely. We have reduction in control or operator overload.
We see those reports. We see McKinsey, Fims, all right? We've got to do what we can to help these folks who actually operate our critical energy as I do. We have to help them do that, right?
And so if they're overloaded and they've got too much work and too many displays, telling them to do too many things, they're not able to achieve and get that value that commercially needs to be sold to clients, right? So by introducing and getting them those tools that are tried and true in refining or firing operators are used to seeing them, I don't think it's a big shift to the pipeline industry. If we think about it that way, and we are a risk-inverse industry as we should be, right? But if we think of it like, this is tried and drew in just as risk-inverse industry, like that's about it, right?
Yeah, I would think the challenge is getting people to think about a pipeline as a process. Yeah, yeah, yeah. Because that's what the control is. Can you elaborate a little bit because how, I mean, I think I know what that means, but I don't know that I could speak it very well.
Like in my head, I'm like, okay, that makes sense, but I don't know how to articulate it. Yeah, it's... Why is it what is it about process that's missing in the way that pipelines typically think about things? I call it, so for me it's dynamics.
So when you're sitting in a control room, you feel the dynamics of the liquids and you can feel like as your operating, you can feel like, okay, this batch has a different density because these pumps behave slightly different, right? So all of that is, you know, as you said, like first principles to get us going, of course we can use better tools like AI and other augment. But that's really, like as you're sitting there, like that's what you're feeling, right? You're like, okay, like I've got to tweak this here because I know that this pump has this impeller and from my pattern matching of operating for the last 10 years, I know that this density with this batch is a little less.
So you use, you create more effort to control your head as a control operator to manage all of those dynamics along the line to shoot for that high operation. Yeah, so that's a very liquid-centric kind of... Which is fine, helps illustrate. That makes sense.
But if you break down what's going on with the pipeline and really get to the details, then you start seeing where the opportunities are. You use a process problem. You know, a particular pump with a particular impeller type is probably optimal for a particular density of fluid. And you're not switching to it, then you're not switching.
Yeah, right. And as a human controller running that pipeline, I'm not going to be watching for every one of those switches. It's just too much workload. But a model could do that.
Yeah, it's not working. Modeling every pump. And then it becomes more like the advanced process control that's implemented and refining because now I'm putting that on every pump. I'm putting that on every separation process.
I'm putting that on every intake process and so forth. And the feel when you're a control operator in pipeline and advanced process control, what we call pipe-by is working for you. The feeling of that as a control operator is like, oh, okay, cool. It's just doing it for me.
It's doing what I want to do. Right. And it feels to you like you feel like the pumps are talking to each other, the pump stations can talk to each other because you can see how, okay, this change resulted in this change over there and I didn't have to go do it, right? Right.
But the display is changing. Yeah, so it means that the pipeline controller is going to become a lot more familiar with the hydraulics and details of the process. Versus just keep it running. Versus just keep it running because they're going to watch how the advanced control is operating at and they're going to be like, oh, that's kind of cool.
So, you know, then in their next training cycle, which is going to go back to aviation, extremely important for control management standpoint. They may have learned some really cool new tricks for the next time they're going in. So how is the technology like this applied? If somebody wants to evaluate this type of thing, because when you start talking about modeling every pump?
It does vary. Yeah, that's scary. That's exactly right. It's like, oh my God, that's crazy.
I don't want to do that. So essentially they're all saying, I'm going to get the value. Yes. So to get the value, right?
It's essentially it's also looking at, okay, well, how is value measured in an industry that we're all familiar with? Well, we're fine. Like, it's been measured, right? Okay.
It was your baseline for a storeable average, your baseline now is storeable new. So it's actually that simple. So with, yeah, with our one, one of our clients who has some new stuff that will be coming up public with us, we took two years of their historical data before implementing. And we took the average, but we also made sure to comb the data and pull out, like, okay, if the pipeline wasn't a portion that month, right?
So we wanted to make sure we're getting a really good apples, apples, and carrots. So the algorithm, that was the historical flow rate, you know, we're viewed with the client, we've already had it. We put it on PyPOT and it just case mass opt and we're running it. And we read it for about three months and we do like, okay, what's it looking like?
You know, we ended up at about a hundred data points or something and we're seeing, you know, we asked you to do two methodologies. This is our first time really betting this. We hired internal client doing their methodology. We did ours.
It looked to us after about four or five months of runtime, you know, the values which converge. So it was really interesting to see, like, okay, in terms of like a project and ROI, you can actually put this thing in and measure exactly how much volume you're going to get. And we've seen it now in production. What that means is that your commercial team can go take that and can confidently sell it.
And this isn't something that engineering tools can achieve, right? Because you don't know that that volume is actually moving, but with an implemented closed loop system, you know that there. Does that make sense for like how it's measured? It's actually kind of simple.
Yeah. Well, and then as you build confidence in your capability, you start applying it in other places, you start, you know, dividing and say, well, what can I do with this pump station? Because this pump station is a substantial asset. And how could I apply this model predictive control there?
So we don't think of it that way. Actually, we think of it in different needs. So because if you think about like optimizing something, if you optimize that one pump, but the pump next to it is an optimizer, they don't talk, then all your work here at this one point doesn't matter, right? So that's how I like to think of it.
So when it comes to implementing advanced process control on a pipeline network, you're implementing from the human's perspective. So you want to implement it in the data and have it talk to every single pump station so that then when you go to start up the pump station and start up a pipeline instead of executing all of your pressure set points or pump starts, you select your target flow rate or your math setting and you select on. And then the software will execute every single one of those intermediaries on your piece. Well, I guess what I'm driving at is with more time and more competency, you have the opportunity to implement this on smaller segments of where you want to get optimal, right?
And I mean optimizing the system, I don't mean optimizing the pump optimizing the system. And that's a different way of thinking about things, right? Because if you go back to your part diagramming or something like that and you diagram this out, well, there's all these different places in the process where I have the opportunity to optimize, but I'm not trying to optimize, I'm trying to make the whole system optimal and that's different. Yeah.
Trying to optimize a point. Yeah, no, good point. I just, I had to say, yeah, no, it's okay. It's a clarification.
Yeah. It's a clarification. Yeah. It's a clarification.
Yeah. I really want to. So where do you, where do you think this tech is headed? I mean, you're, you guys are kind of on the leading edge of control room technology, particularly talking about this predictive control type stuff, autopilot pipeline controllers, auto control.
We call it Python. Yeah, I know. I'm just making up other names for it. Yeah.
So where do you think this is headed? So as of right now, it's, it's, it's fun. Like honestly, like when control room operators get their hands on it, like it's folks actually really, really love it because like, just, just think about, you know, as a human, like you, you, you've all said you buy a car with a rapid cruise control. Like it's cool, right?
Like this is fun. Like, yeah, I have this new tool. It makes my life more joy, a lot easier, right? So that's, you know, in terms of where it's going, I think the remembering fundamentally that, that making these folks life easier, they're the ones on the short bend.
They're actually pushing the buttons that, that generate the revenue, right? And so anything we can do like their lives easier, I think as employees and shareholders of all of these wonderful energy companies, like that, that, I think is extremely important. And so I think what I'm hoping in terms of where the technology is going is that there becomes a better understanding of what the actual capabilities are of the technology and what the point is, right? Like the point is moving, you know, oil and gas safely and efficiently, right?
So how do we do that? It's not about the AI. It's not about the AI. It's about moving oil and gas safely and efficiently.
Yeah. No, I, I, I'm with you. I, you know, we, I want to kind of segue a bit. Might be a little clumsy segue, but we'll see.
We were talking before we got on about how do you apply and how do you get to optimal? And you were talking about the, the first thing you need to do is you need to look at your system and understand what the safety constraints are. Yeah, right? And then once you have the safety constraints, like, don't get outside of this, then you can start looking at applying artificial intelligence machine learning to do some of these highly computational, highly repetitive tasks.
But you've got to start with getting clear about what are the safety constraints and make sure the overall system model doesn't go outside the safety constraints. Exactly. I'm just going to ask you to elaborate on that because that idea of how to build a pipeline control systems radically different than how we have historically done it. Right.
So, what we mean in terms of like APC? Well, what I mean, what I mean by that, you know, being a guy who's done a ton of skated projects, right? When you go to do a skated project, basically, that's all about understanding all the IO and, you know, all the things like all the buttons I can push and just making all those buttons reliably available to a person. When I start trying to run on autopilot, the first thing I got to do is make sure because we rely on the person to keep it inside the safety constraints.
If I'm going to really optimal, I'm going to run the system in an optimal way and optimize the human input, then I've got to first build my model around, well, what am I safety constraints? So now if I'm building a safety, I'm building a control system. First thing I got to know is I got to know, well, what are the safety constraints? What are the pressures and flow rates that I can't get outside of?
What are the RPMs and, you know, the horsepower and the current draws on these pumps that I can't get outside of? That's where I have to start. And that's a different way of thinking about how to build control systems. Yeah.
It is. It's I think it's some of it's the same in terms of like, in terms of how we do it because for us specifically, right, like we're very aware of all the critical safety constraints that are at the PLC level in the state level. So I think that when thinking about that and that being important, like you want to architect the system in a way that you're effectively making it into fragile, rightly you're taking away any opportunity that could mess up. So, you know, so that's also leads to, you know, you have a human in the loop.
If you can minimize human in the loop, you're also minimizing the opportunities that these can get messed up. So, you know, if we're thinking about that, it's going to be know that our settings are correct. We know that our PLC things are correct. Why would we mess with it?
Right? So we're going to implement a technology like this. You know, some books might think it'd be a great idea to do like NQTT go straight to the end of the race. I completely disagree because that goes against like the IEEE, your typical safety standards, right?
If you've got your layers. So, so I think when you're thinking about safety of implementing this technology, just be very aware of like, is there a possibility that you could compromise any of your existing safety layers? And if that's no, that's fantastic. If it is, you probably are protecting it wrong because you shouldn't.
And then when you're within that, you know, so now you know, you know, you know, you're bound and you've also set up multiple layers. So, you're inside of those bounds, which are defined because they're very important. And if you had an open AI model, it knows it might not respect those bounds. It doesn't know they're there.
So now you've got them. And you know that, okay, like I can calculate theoretically what my hydraulic mass line is on a spike line, right? We've all done it. We know, but we know that we never actually get that home.
So, it's, yeah, like I guess, I guess, I just don't know if I just start battling there. No, no, I think it's a great answer. I think one of the things that you can, you know, you can take this level out of less than out of aviation. I use aviation as an illustration lot because I even though I'm not a pilot, I'm told a lot about it.
But you know, this is one of the things that they've done with the flight computers in the aircraft, all to minimize fuel cost, right? So they know what altitude, you know, what throttle setting, all that kind of stuff to minimize fuel consumption and maximize my ability to ride on time. So that is like, yeah, so, yeah, it's a similar kind of idea, right? Once I have it in place the right way, then I can start looking at, okay, how do I make this optimal?
Yeah, and I think a common misconception in pipeline is that that level of interconnectivity between pump stations to be able to have some sort of control that behaves like control operator is possible with the current tools. And it's actually not, right? Like, unfortunately, the data doesn't have those libraries. It just doesn't, right?
So, so when I hear folks are like, oh, we already do this, like, the libraries just aren't there. So you can't. It doesn't exist. Yeah, no, it's fascinating.
I mean, one of the things about what you guys are doing with pipe-bought and what you're doing in the control room space, I just find it fascinating. I'd love the opportunity to play with the tech and probably never gonna have it because I just don't have the bandwidth for such thing these days. But oh my gosh, it sounds like fun. So this could be a fun result is a we are we need better sales and implement tools, you know, as a company, we've definitely like a dis demonstrated value.
But we're working on, I might take a few months, but we are working on like a very nice, like demo. So it could be a fun episode to have you like, try to play with it. Oh man, we'd have to film that one, make sure we could get the, yeah, that would be fun. I'm up for that.
Let's do that. We don't wrestle versus pipe-bought, see what happens. I think that would be way cool. I would dig that.
All right. Well, look, this is great. Vicki, it's always a pleasure to have you. Thanks for coming back and wish you all the best in your sailing adventures.
Yeah. Oh, yes, yours as well. I'm coming to join you. Yeah.
All right. Well, thank you. I appreciate it. Thanks.
I hope you enjoyed this week's episode of the Pipelineers Podcast and our conversation with Vicki. Just a reminder before you go, you should register to win our Pipelineers Podcast Yeti Tumbler. Simply visit pipelinepodcastnetwork.com slash win and end yourself in the drawing. If you'd like to support us, the best way to do that is to leave us a review.
And you can do that on Apple Podcasts, Google Play, Spotify, wherever you happen to listen. If you have questions, ideas or topics you'd be interested in, please let me know either on the contact us page at pipelinepodcastnetwork.com or reach out to me on LinkedIn. Thanks for listening. I'll talk to you next week.