I've had NP be lower than AP at times for intervals up to 10 minutes. Something to do with being able to average a high steady output from the get go versus having it weighted on a rolling 30 second delay.

NP is simply too gimmicky. I pretty much stopped paying attention to it 8-9 years ago after I figured out how to crack some ridiculous hour long NP buster sessions 30 watts over FTP with less than 15 minutes of actual hard riding.

That's normal and to be expected for the repetition interval you are using. NP wasn't designed to handle intervals in the single digit range, the bigger the difference between the ON and OFF segments, the more skewed results will be. As velorunner notes, NP uses a 30 second window, rolling window. Very short (very hard) intervals will accumulate the 30s prior to the start of the interval (at low power), and skew the NP value downwards.

However, its not a gimmick as stated by velorunner. Its simply a limitation of the mathematical model of human physiology.

Historically WKO, wouldn't even display NP for efforts shorter than 10 minutes, and its still not likely to be accurate for windows less than 20 minutes. Kinda falls under the heading of RTFM. Its not a faulty metric, its a faulty application of a metric to a circumstance it was never intended for.

NP was intended to estimate the physiological cost of an entire session, given variability in power, and non-linear costs as power approaches and exceeds threshold. So, it should give you a good estimate for the cost of the entire O/U session (as the peaks and valleys will eventually even out the errors)...but, it won't tell you anything useful about a single O/U cycle.

That's the gimmicky part.

Its not any more gimmicky, than modeling the current-voltage relationship of a diode, or the gate-voltage RdsOn characteristic of a MOSFET, or the model of the earths movement around the sun, or any of the other million things we model in the physical and physiological world. Every single model humans have ever developed is good "up to a point", and "not so good" beyond that. That's at least half of what I teach new engineers that have just graduated from college and come into the real world.

However, as with ALL models, its important to understand the nature, and limitations of the model and use it for what it was intended...and avoid the degenerate cases. Failing to know and heed those limitations is on you, not the model. They have been well documented since they were first conceived by Coggan, et al. Its still documented over on the TP site, including the exact question asked above: "Why is my NP lower than my AP?"

https://help.trainingpeaks.com/...an-my-average-power-

I think TP NP calculation is a rolling 60 sec average so you're including the work before the interval

Thanks all, makes sense.

I was going to go from my 3x12 under over to shorter steady value at new higher target and was attempting to use that to get an idea.

Though I think the point is that the Banister model and derivatives - and particularly NP with it's pretty arbitrary "4" exponent - are brutally rough models compared to Kepler's laws or classical EE models. They're the best we have, and I use them. But I think the "up to a point" doesn't get anywhere near as far as your examples.

As just one example, aside from NP, is how Chronic Training Load correlates to "fitness." The Banister model assumes humans can absorb infinite training stress. Which is of course completely false. So it's not all uncommon to have your highest-ever CTL scores and be absolutely shattered and worthless as an athlete. By comparison Kepler's laws won't *completely* get it wrong about where a planet might be (I don't think?). Though I supposed in real-world diodes and MOSFETs the classical models will blow up near the edges of where they're designed to operate. Trouble with us humans and the Banister model is sometimes it's super tough to figure out the edges of where we're supposed to operate. n=1, I blow myself up all too often.

Not that NP doesn't have some use. For long duration, it can be an indication of crappy pacing if you tried to have a constant power goal. Or it can give suggestions about how to recover for road racers who necessarily have spikey power output. But I consider it a pretty blunt hammer, not a scalpel, as a model.

All very fair criticisms. There are, in fact, multiple models of diodes and transistors. Some very simple, and some deeply complex. Which one you might use depends on the application. The same can be said of planetary motion. We are taught in elementary school that they CIRCLED the sun, then we learned it's really more of an elipse, then there are gravitational effects from nearby planets, etc. Depending on how close you look, that path is round, egg shaped, or wobbly.

However, I suppose I would counter it really depends on what you are trying to do. Are you trying to know where to point in to sky to see venus... Or are you trying to land on Mars? At some point relativy matters, and can't be neglected... Else you miss the entry window. Similar differences can be described for semiconductor physics. It's a matter of perspective, scale, and precision. Strange things happen at the edges.

For time series functions, it's important to understand the intended time span application. Np with its roughly 1 minute time scale will be susceptible to degenerate manipulation in applications below 10x that time scale (10min), and can still show appreciable error below 20x. That's not a special property of NP.... It's just the nature of impulse response functions---their error increases as you get closer to the window size. The nonlinear function makes that all a bit worse, than something more linear.

Anyway, I suppose this has drifted far from Mr sheep's question.

So sure, one could argue one detail or another about the NP algorithm, as with the Bannister model. But, a broad understanding of both models and their limitations still makes them useful for planning and management purposes....until something better comes along.

It doesn't take long to realize you can't add training load infinitely. 6 months into a build and it is dead obvious....i might be about there, actually.

If the planets took a square path around the sun, would they still be circling it or would they be squaring it? :)

are we to believe that when settlers 'circled the wagons', they did it with any sort of precision?

Hence Coggan frequently quoting "the map is not the territory"

NP is calculated using a rolling 30-second window, which will smooth out a lot of the variation in those shorter intervals. And 30W isn't that much variation to start with. If your 2' unders are at 285W and 1' overs at 315W, and you had an infinitely long set of those intervals, the NP would be 295.8W and AP 295W.

You can also get edge effects affecting AP and NP differently depending on how the software treats the first and last 30s of the workout.

TrainingPeaks article that goes into the 30s averaging

Ha. I was on the way to posting this when I got to your post. Good reading for those who haven't already.