Endurance Analytics

Our blog - covering sports analytics and EA products

Ride Review with Post-Event Performance Modelling

The opportunities to use power based performance modelling in a pre-event context, for Goal Setting, Scenario Analysis, Equipment Optimisation (EO), and Optimal Pacing Strategies (OPS) should by now be compelling for all serious cyclists and triathletes. For the price of a few carefully collected details describing a rider's physical attributes, power-duration curve, and bike setup, it is now possible to first predict a given rider's time on a given course and then to apply scenarios on the underling model to make the same rider go faster. Analytical intelligence equals free speed...if you want to put the work in.

While the pre-event context has always been the most compelling there are also significant opportunities to use the same performance modelling techniques in a post-event context. Nobody wins a race by looking back at the day's data but but you can, perhaps, use it to win the next one...if you know what to look for and have the right tools. We have invested a lot of effort recently creating performance modelling tools for the post-event context and this article outlines what they might be able to do for you.

It's not (just) about the bike (data)

Pre-Event modelling combines data about a rider, his or her bike and equipment, the course, and the weather. Weather can mean historical trends or the actual race day forecast. We then simulate the rider's performance on the given course, under the given weather conditions. A simple simulation, based on the maxium power a rider should be able to sustain for the course duation gives us a time-on-course prediction. Many variations of power output, sector by sector, allow us to find an Optimal Pacing Strategy. Variations of rider parameters or equipment allow us to see the effect of scenarios. Variations of weather allow us to build a complete picture of which equipment is "probably" the fastest, on a "normal" day.

Post-Event modelling replaces weather forecasts with actual, observed weather. It also takes the ride power and speed data (via a .fit. .srm. .tcx, .pwx or .gc file) as a precise description of how the ride was executed. This gives us analytical possibilities and opportunities to combine and compare "model" versus "actual" performance not possible in the Pre-Event context.

Performance Modelling is more than race planning

Some people see performance-modelling as a way to predict race-day performance and then create a plan. That is Pre-Event modelling, but there is more to offer. Post-Event modelling, with the extra second-by-second information provided by the ride file, has some powerful applications. Let's review some of them:

 

The Virtual Wind Tunnel has never been more available

In our experience the most versatible, and now most used, field test protocol available for determining a rider's aerodynamic drag is the "Virtual Elevation" method a.k.a. "Chung" method (after it's inventor Robert Chung) a.k.a. the method of the "Virtual Wind Tunnel". Some very serious operations routinely use this method, in a velodrome setting, to provide riders with drag data and aerodynamic optimisation advice. In simple terms this methodology takes ride file data (power and speed) to build a "virtual elevation" (VE) profile for the ride. If you vary CdA, or let a computer do that for you, until the VE profile closely matches the actual course elevation profile of the same ride, then the CdA you finish with can be a very good estimate of the rider's actual CdA.

Post-Event Modelling, by combining the rider data, course data, ride data, and environmeental conditions, already has all of the information necessary to estimate CdA using the VE methodology. There is one caveat - wind. Rather than assume zero wind, as velodrome operators can, making VE available for any rider requires us to factor in the wind. Since we already have weather data, at times of day covering the event, and we know rider's time's of day on all sectors of the course, from the ride file, then we can do a reasonable job of factoring in air density and effective winds. VE under observed winds isn't perfect, but if estimating CdA is important to you it is better than nothing...a lot better!

Physics + Power Data = Power & Energy Attribution

Imagine a ride file says that a rider averaged 200 watts on a short climb. Now imagine we model that rider's performance on the same climb, we take all of his parameters, then plug his time into a "power given speed" model to estimate the power required to complete that climb at the given speed. If we're using the right parameters, including VE CdA, then our estimate will be 200 watts. But we have more information... Those 200 watts are composed of some watts required to overcome gravity, some to overcome aerodynamic drag, some to overcome rolling resistance, and a few others. By combining the actual power data from a ride file with theoretical power demands from a performance model we can achieve what we call a "Power Attribution" or "Energy consumption study". (Remember that energy, in joules, is just watts multiplied by time in seconds). This is an interesting capability of Post-Event Modelling.

Pacing Quality Evaluation - How well did you execute?

The identification of theoretically Optimal Pacing Strategies is one of the most compelling applications of Pre-Event Performace Modelling. But how can we know how well a strategy was executed? Sure, we can compare race normalised power with the prediction from the plan, and we can eyeball ride power and speed data relative to the sector by sector plan, but we felt that some better tooling was required.

"Pacing Quality Evaluation" is a software based comparison of actual, sector powers and speeds, from a ride file, verus a theoretically optimal plan calibrated to the rider's power-duration curve as exhibited in the ride it'self. Having looked for some magic number to describe pacing quality we eventually settled on the one number that counts...time! This analysis can also highlight something very interesting - good riders can be very good at pacing. Often the way a rider executes, with the freedom to vary power on every little hill or under every little gust of headwind, results in a time that is better than what we might predict from an "optimal" strategy simple enough to understand and execute. Nevertheless, Pacing Quality Evaluation is a fantastic gross error check and has some great applications in learning lessons and executing better in the future.

Rider Benchmarking with Virtual Competitor

We have always emphasised the application of performance-modelling in the key coaching disciplines of rider profiling, performance benchmarking, and goal setting. It is a powerful thing to be able to compare a rider's actual performance against some theoretical competitor, a better version of the rider's future self, or even a worse version of the rider's past self. A virtual competitor study, in the post-event context, makes this possible.

Real World What-If Analysis

Have you ever agonised about the race you or one of your riders lost, perhaps by a matter of a few seconds? Wouldn't it be great to know the size of the winning (or losing) margin, not in terms of time but in terms of parameters you can work to improve such as power, weight, or drag? Post-Event what-if analysis does exactly this because we can re-simulate performance given certain improvements. No more agonising.

Post-Event modelling is now available at our usual Performance Modelling link. Simply select "Post-Event Modelling" near the bottom of the page. CdA estimates via the virtual wind tunnel are open and free. The full Post-Event report, with all of the above analytical features, can be configured below.

3D Visualisation, applications in Performance Modelling

Anybody who has ever been fortunate enough to pilot an aeroplane has probably experienced an eerie realisation that the world is far more 3 dimensional than most of us ever really think about. Up there, a couple of thousand feet or more above ground level, there is almost a sensation of diving somewhere in the middle of an aquarium while the place that once represented the 3D world feels like nothing more than the gravel and the weeds down at the very bottom of the tank. The opportunity to see the world like that has a lot to do with why many of us stop, gaze and enjoy the vista on high mountain passes.

Modelling cycling courses relies on what we call “triplets” of GPS data – latitude, longitude, elevation – and we have recently been looking for better ways to visualise data in these 3 dimensions. We were incredibly fortunate to hire a guy whose recent projects included some pretty serious geospatial work – software used to remap the bed of the Panama Canal – and the image with this blog is an example of the capability he came up with. In this image we’ve rendered the gradient of every road sector as a kind of “4th dimension”, now quite a familiar technique, but the possibilities go way beyond that.

Hotspots

CPL predictive race modelling typically includes projections of riders speed given power, variable power levels required to execute an Optimal Pacing Strategy, and even cadence ranges given that power, rider attributes and gearing. We can now map any of these projections onto 3D visualisations of a course and we think this opens up a world of possibilities in terms of understanding race planning and execution. Imagine being able to look at the course of a forthcoming event and see the “hotspots” where a rider will be travelling particularly slowly – often the best candidate sectors for pacing improvement – or where a particular range of gearing may be inadequate.

Impacts

There is more too. One of the big benefits of Pre-Event Performance Models is scenario analysis. What is the impact of a rider’s time on course given more power (CPL Power Delta), less weight (CPL Mass Delta), less aerodynamic drag, or changes in equipment? Currently we can illustrate these impacts for the 4 quarters of a course, and for the course in aggregate, but imagine if these impacts could be visualised through every single sector. Where does a heavier but more aerodynamic wheel cost time and where does it buy time? To the best of our knowledge the cycling world has not seen this sort of analytic detail in 3D before.

We expect to incorporate more and more 3D visualisations into what we do. The image with this blog is from a rather special pacing project that came across our desk this week. In the meantime, if you the reader should have any particular visualisation ideas or problems with which we may be able to help, then we are always happy to hear them.

Sustainable Power is a moving target

The trouble with rider’s power output is it falls over time. Not in the course of a season, when we hope it increases due to training stimuli, but as a response to the duration of a ride. Whether we’re interested in short efforts, when sustainable power decreases as a function of anaerobic work capacity, or relatively longer ones, when fatigue is the problem, there is an inescapable reality that we become weaker as the clock ticks over. This reality can be visualised as a rider’s “Power Duration Curve”.

In some way stronger riders enjoy a double advantage. As well as going faster from the outset they also arrive at the finish line way before weaker riders, putting their effort on a very different section of the curve. Indeed pro level ironmen face a completely different race to the typical age grouper who will be out on course for hours more. What may be a tempo ride for someone can be a test of endurance for another. A layman might conclude that we should all ride full gas from the off, thus completing whatever course asap and aiming for some shorter (but higher) point on the curve. Unfortunately the realities of physiology quickly undermine such a strategy.

Completing any course as quickly as possible comes down to intelligent pacing. Even if we discount the opportunities presented by a computer optimised, variable power strategy, there is a problem to be solved. A rider needs to know the highest power output he can expect to be able to sustain for the duration of the course, to leave everything on the road and get the best ride out of himself. But what is his expected duration on course?! Only predictive analytics, performance modelling, can do a good job of revealing that.

There is a chicken and egg problem here. To estimate time on course we need to know (average) sustainable power, something to feed our model. But to identify sustainable power we need to know time on course! Thankfully computers are pretty good at solving these kind of problems…we can “iterate” a power value until we “solve” for a rider’s best sustainable power. Best means the number that leads to the fastest time on course, but rests no higher than the rider’s curve at that duration.

CPL Event Models have always solved for best sustainable power based on your inputs, but we haven’t done a great job of explaining this. That’s why, as of now, where you see our power prediction, you can click on “why this number?” to open an insightful chart. You’ll see two lines 1) Power Duration Curve… based on observed power inputs & 2) Power For Time on Course. Where these lines cross (and a rider’s fitness would have to be critically unsuited to the event if they don’t) is best sustainable power.

Best sustainable power is clearly a key number for any rider wanting to go as fast as possible, even if a variable power pacing strategy has been discounted. So of course we’d encourage everyone to use the free analytics here on CPL in uncovering that number. No rider can out-perform his fitness but the number comes with two guarantees. First, said rider will finish about as fast as he can, no guess work. And second, the suffering will end about as soon as it can do. Smart riders should like the sound of both.

Evaluating Optimal Pacing Plans

Performance Analysis & Modelling in sport probably began in field games such as football and baseball where the objective is to turn analytical intelligence into real tactical advantage which increases the probability of victory. In cycling the highest objective or “holy grail” is free speed - faster race times for equivalent physical effort - and no application promises more of this than the development of Optimal Pacing Plans.

At CPL we’ve been developing optimal pacing algorithms since 2010, when we provided the first web based tool for experiments in variable power pacing highlighting the importance of sector-by-sector constraints on rider's ability to deliver power. This has always been a “beta” tool (further development intended) but we have recently had the opportunity to do a lot more research and development in this area. Since we are no longer the only resource for optimal pacing plans some of our conclusions may be of even greater interest in light of what others are offering.

Beware Simplification

In the same way that climbing and descending a mountain is not the same as riding an equal distance on the flat (though both propositions equal zero net elevation gain) any efforts to simplify a course take us a long way from the realities of rider-on-course modelling. Smoothing away even the smallest ups, downs, and changes of direction can seriously compromise and result in overly optimistic pacing plans. We have always done performance modelling at the highest resolutions made possible by GPS profiles and observe poorer correspondence to reality when that resolution is dialled down. So beware simplification.

Compute Time

The Optimal Pacing Plans included in our Popular Event Models leverage the full resolution of the course profile for the best achievable accuracy. In fact it takes a lot of computing power to come up with a plan from this kind of model. So much so that we don’t pretend to offer plans you can calculate in a couple of seconds on the internet. Our plans take several minutes of compute time to generate so sending you a pdf is the only way to keep things honest and useable. The ten sector summary we show you is just that - a summary – hiding the intricacies of the real model.

Execution

Can you control your power output to the nearest watt? Can you even hit an exact average for a certain interval or stretch of road? Probably not. Most people would struggle to hit this accuracy on a turbo trainer, never mind in the varying environment of road cycling. So beware pacing plans that expect you to execute in terms of single digit resolution. We work to the nearest 10 watts - just about executable – with insignificant effects on best time on course predictions. The only exception to this is in the final sector of the course where there is only one number that will solve for an overall target power.

Methodology

One way to solve the optimal pacing problem is to look at the impact of small changes in sector powers on 1) time in sector & 2) target power at the course level. Logically the sectors which offer the greatest reduction in time-on-course for the least impact on target power (dt[s]/dNP if you’re into maths) may be the ones where a rider should “power up”, and vice versa. Various optimisation techniques can exploit this kind of theory but we have found better results using simulations. The way we develop Optimal Pacing Plans is conceptually similar to having a rider tackle the event 10,000 different times. Every time this rider tries a different pacing strategy – sometimes he blows up, sometimes he’s too conservative. We simply keep the details of the fastest ride where he doesn’t blow up and that is your plan.

Want to try a performance model with Optimal Pacing Plan? Simply choose an event here.