Transcript
Introduction to Merlin
Today, we have an interesting one. We are gonna talk about Merlin, and Merlin is all about using AI to analyze power quality data. The paper gives a really high-level overview of how this can be helpful for utilities.
One of the real fears of using AI, and it’s a perfectly valid one, is the concept of AI, especially large language models, hallucinating, saying things that aren’t true, doing math when they’re not really good at math — they’re good at language. And so it’s really important to avoid those kind of traps when using AI for anything engineering-related, including, of course, power quality analysis. With Merlin, we’ve gone to great lengths to avoid those sort of problems.
How Merlin Works
The concept with Merlin is to provide kind of an eyes on every piece of power quality data. It’s not doing math, it’s not computing metrics and thresholds. That’s done with deterministic code. What Merlin does is takes that output — compliance reports for voltage regulation, for flicker, for harmonics, for unbalance, for sags — and puts a narrative around that information.
In addition to that, it also is using image analysis to look at waveforms, classify them, looking at strip charts, looking at sags and swells, and finding patterns. It’s not doing quantitative math. It’s doing a qualitative analysis to give an interpretation of the graphs, spot trends, and correlate information from different types of data.
This sort of pattern matching and kind of fuzzy logic and putting together a narrative around the recording is what these models are really good at. They can take a lot of information — again, it’s not doing math, but information that’s summarized for it — and giving you that sort of high-level perspective of, “What is this data telling me? What are the patterns? How is this flicker, say, related to current, or to sags, or to harmonics?” And put all that together into human-readable format so that you get a head start.
The Junior Engineer Analogy
It’s kind of like having a junior engineer that looks at every single waveform, classifies it, ranks it, and then looks at every strip chart, looks for sags, looks for swells, looks for loose neutrals. And it’s looking at every piece of that, putting together its narrative. Then those pieces feed into a higher level tier that gives an overall analysis of the entire dataset.
That’s presented to you so you have a head start. You can, of course, always drill down into the data, but the concept is that you’ve got a first pass. It’s like having a junior engineer that you send off to the corner and say, “Here, look at everything in this file.” And they give you their first-pass junior engineer analysis. What are the strip charts telling me? What are the waveforms telling me? If you look over time with the waveforms, what’s the trend across all of them? How does that relate to the strip chart? How does that tie into the compliance reports? It’s putting all that together for you.
Metadata and Knowledge Base
The more metadata you give it — for example, if you give it the customer complaint, things like the transformer size or any sort of primary site operations, like recloser operations or cap-back switches — it’ll use that information also as an analysis. It’ll think about that.
It’s using the information that PMI has gathered over the years. Hundreds of whitepapers, standards, IEEE and other standards — that information is kind of baked in. The 30 years of power quality knowledge that we guided these systems with, to have it understand how current and voltage are related together, and how different patterns — a single-phase fault versus a three-phase fault in the primary, what that looks like on the secondary and how that impacts equipment. All that power quality knowledge that we’ve accumulated over the decades, and by analyzing literally tens of thousands of files over time, is baked into the system.
So it has a really good understanding of power quality, but it’s also grounded with the standards themselves. It’s got deterministic code that computes the standards compliance. It knows what the thresholds are, and so that keeps it tied to real life.
Demonstration in PQ Canvas
Here in the paper, we explain that, but I’m gonna jump to a demonstration of Merlin. It’s in our cloud-based system, PQ Canvass. Here we have a recording. You can see the data directly. You can look at interval graphs and analyze the data yourself, just like you would a traditional power quality investigation. But you can also look at the AI output, and you can start with that or you can start with the raw data.
AI Analysis Overview
I’m gonna hit Overview and go to the Merlin output. And here we have the AI analysis. Again, this is a multilayered, kind of tiered orchestration. At the top level, we have the power quality report. We have a one-line summary of the entire recording. So in this recording, we have sustained overvoltage with brief sags and swells. That’s kind of a one-line overview of everything in this recording.
On the right side, we see the compliance standpoint. Regulation per ANSI C84 is failure, high severity. This is a score from 0 to 10. A 10 is pretty severe failure. The other compliance reports are passing, but they’re not perfectly clean, but still pretty good, threes. And then we have the top issues here: voltage regulation, sustained overvoltage from regulator step and high set point, some voltage swells, and some high distortion, 5 to 10% THD.
Executive Summary and High-Level Report
Now we can click on this and go to the high-level report. Again, this was all generated with Merlin. We have an executive summary. This is an overview of everything that Merlin saw at a really high level. This is the top of the orchestration, where all those sub-agents, so to speak, were analyzing their piece of the data — strip charts, compliance reports, waveforms. And that all percolates to the top here in an overall summary.
This is a 21-day recording and consistently high service voltage with long renewables above ANSI C84 Range A. So that’s the overall narrative, and there’s some other issues — some swells, some disturbances and not many disturbances, one or two sags, low flicker, some moderate voltage distortion. This is an overall summary of the entire recording.
PQ Issue Matrix and Attribution
We have the PQ issue matrix that kind of summarizes each of the power quality problems, what the ranking for all of these in terms of severity and how often they happen, any sort of patterns, its estimate of the attribution of the problem. Is this a utility side problem or a customer side or is it mixed? In this recording, it’s pretty much all utility side. This is steady state voltage regulation, so that’s gonna be mostly on the utility side.
Compliance and Voltage Regulation Details
We can scroll through the report. Here’s a compliance standpoint, so for each of the compliance reports, here’s its overall narrative. We can also just jump directly to the sections that are important. In this recording, steady state voltage regulation is the most important, so we can just jump right down into that.
Here is an overall picture of the overvoltage situation at this recording. Only about 61% of the time is it within ANSI C84 Range A, and it’s actually going above Range B 16% of the time. So this is a pretty severe overvoltage. The customer has a legitimate complaint here, and it’s not really changing with load current. This is an upstream regulation problem. It’s not on the customer side at all.
Equipment Compatibility and Root Cause
We can look at the customer equipment compatibility assessment, and this gives you some guidelines for whether this is likely to pose a problem for the customer. Here, the fact that we’re going above ANSI C84 Range B is a likely issue. So it’s not surprising that it’s gonna say you might have some issues here.
The utility contribution and customer contribution narratives discuss, based on the data it’s seen and based on the relationship between current and the power quality issues, how much of this is likely on the utility side versus how much is likely to be on the customer side. A root cause discussion, key events and some guidance if you’re gonna continue the investigation. Some different types of sections here in this report are gonna be useful or less useful depending on the actual problem and what the complaint was. But it’s all available here for you as an overall Merlin analysis.
Data Quality and Confidence
Merlin has analyzed this data and put together this narrative. It’s not computing thresholds, it’s not computing compliance — that’s all done with deterministic code, that’s outside of the AI piece. It’s putting together the picture for you, putting it into words and explaining what’s happening.
Another useful piece here is the data quality and confidence. It’s ranking not only its own confidence in what it’s saying, but also in the data itself. Is the data complete? Is it missing channels? Did you not record something that would be useful? That can be helpful to put some perspective on these claims, how can it back it up?
It does back up what it’s saying with these snapshots. Whenever it has some sort of claim, it tells you when these problems are and it’s giving you, in a snapshot list, where the proof is. It shows its receipts for anything it claims — here’s where in the strip charts you would go to see this, so that you can do a deeper dive and verify this yourself.
Questions for the User
Another piece that can be useful is the questions for the user. These are questions that are geared towards the loads, geared towards the utility side, geared towards the customer. These can be helpful in kind of drawing out more information. Based on what it’s seen and based on whatever you tell it the complaint is, it’ll suggest questions that might tease out more clues in the investigation.
In this situation, it’s fairly straightforward where it’s overvoltage, it’s certainly on the primary side, and that’s fairly straightforward. In a more subtle investigation, these questions might turn up some useful information that would be helpful in continuing to find a root cause.
Waveform Classification
Now you can also dig deeper. The waveforms have been classified. Here is every waveform capture with a severity score and a classification, a narrative around every single one. These are sorted by severity. Here we have a lot of rapid voltage changes. We have a voltage swell and we can click on this and see that indeed it was a voltage swell, but Merlin’s already done that for us, we don’t need to do that manually.
We have also a waveform capture report. This is a summary of just the waveforms. This is a narrative around the waveform captures. You can get a lot of event-based information from here, especially if you have a lot of transients, it’ll talk about that in detail and how that’s likely to affect customer equipment. But this information percolates up to the top level report.
Strip Charts and Drilling Down
In many cases you can stay right here on the top level report and everything you need is in here, but if you want to drill down, you can drill down into the waveforms, you can drill down into the individual strip charts. Here’s the sag strip chart, for example. Here’s the sag piece of Merlin that only is looking for sags. Here are the sags it found, here are the regions of interest, these are where the worst sags were and so you can dive into this and verify what it’s saying is true, or if you need to dig deeper.
In this recording, again, steady state voltage was the most significant piece, and so we could focus on the strip chart that is for voltage regulation. These are ranked in severity also, so voltage regulation is a seven and the swells are also seven. These are the ones to look at. The others are fairly low in severity.
If we look at the voltage regulation strip chart, it’s telling us the trends in the strip chart itself. Where did it see high voltage or overvoltage? Where was it okay? What sort of narrative can it put around that? Same thing for the focus areas — these are different power quality topics: flicker, harmonics, sags and swells. Here, for example, sags are pretty low, so there aren’t really any high severity sags, so that’s what it’s telling us. Sags aren’t the problem here.
Second Recording Example: Sags
We could look at a different recording and find one where, indeed, sags were the problem. In this situation, everything was good except for sags. Here, voltage regulation is nearly perfect. We have some sags though. It violated 1668, which is basically the same as CBEMA 47. We have some severe sags.
Let’s look at the SAG PQ report. We have one severe sag that dominates this. This one sag is probably an upstream event. This is what caused a problem for this customer. We can look at the attribution here. Utility side, because of the fact that the current didn’t really drive that.
We can go back to the overall utility contribution narrative here, and it’s strong for utility side. We can even just start at the PQ issue matrix. And in some cases, this is really where you can stop. The severity is eight for sags, utility leaning confidence 0.92. That’s a zero to one score of how confident it is in its attribution. That’s a pretty high confidence. That’s almost one, and it’s saying single phase, deep dip with cross phase swells.
That’s a single phase fault on the primary side, which causes voltage swells on the other two phases. That’s the classic pattern of a single phase fault on the primary, and Merlin knows that, and that’s why it’s saying this is utility leaning. And it’ll say more about that in the contribution narrative.
Verifying Merlin’s Analysis
So this is a very quick way of assessing the power quality of data, giving a root cause estimate, and really a first pass. Merlin has looked at everything in this recording and pulled out the most important things, summarizing them in a table with scoring and allowing you to dive deeper into it. You can always, of course, go into the recording itself and take a look and do your own deeper analysis.
Here, the traditional way would be RMS voltage and current, and we would find that same deep sag that it found, and we would basically find the same thing. So here we can see we got a deep sag, we have a swell. This is what it’s talking about. We have a swell on two phases. We have a deep sag on the first phase, and of course, we have a waveform capture here. This is the highest severity sag, and it’s got it right.
Of course, this is a classic single phase line-to-ground fault. If you look at the RMS graph here, we have a sag in one phase, swells on the other, and that’s exactly what Merlin said.
Getting Started with Merlin
Well, that’s a really quick overview of Merlin. There is a longer introduction to Merlin that’s a one-hour class. Go to our website and sign up for that. It’s a free class. You can also try the demo of Merlin. Just email us for an account, and you can try Merlin out on these recordings and see what it says.
You can read the power quality reports, you can dive into the details of what it’s saying, compare that with the recording itself, and really see for yourself how powerful Merlin is and how it can save you so much time without having to go through every single piece of data yourself, when in many cases, there’s only one or two important things, you’re just not sure where those are. And so it’s a great first pass triage for looking at power quality data.
Now, if you have any questions, give us a call anytime, 1-800-296-4120, or send an email to support@powermonitors.com. Especially if you’d like to have a demo of Merlin and try it out, send us an email or give us a call, and we’d be happy to set you up. Well, thanks for attending everyone, and everyone have a great afternoon.
