Power and food consumption predictions

[aguimorefran]

Hi guys, continuing my work from here:

Finish time prediction tool

Hi guys, I come to present a little project that I have done. As a data geek computer scientist, I saw the opportunity to learn from my data and make predictions on the GPX files given by the organizations before the races. That’s why I’ve made a tool that collects all the Strava data you have, trains a ML model, for now with 94% accuracy, and tells you how long it would take you to finish a given track in GPX. The result it throws up is a map with the GPX plotted, with the sections for each s…

I am developing a new feature which is a power and food consumption prediction program. I am training a ML model with my freerides (no workouts = no target power), and depending on the track and training stats (ATL, CTL, weight, etc), it learns how many watts I will give in any route section.

For now I am getting roughly 0.85 of accuracy for 3s avg power which is quite good.

I once saw that Garmin units had some sort of plugin that tells the rider when to eat or something like that, and since I have a “simple” Wahoo, I wanted to get something similar.
As I can predict the power, I can predict the kcals consumption. From that, I developed an algorithm that takes your favourite foods (gels, bars, fruits), with its carbohydrates macros, and this kcals prediction, and calculates when is the best time to consume each food to match the kcal consumption.

I am working on the “health” formulas, as the theory says you have to eat roughly 70% of your kcal output, also the CHO consumption, CHO level in blood decay, etc.

I will be posting the results once I have a nice output and charts. Share your ideas if you have some.

Cheers

[aguimorefran]

Here are some advances I achieved:

I am using the following data to train the model:

• grade (compared to previous point)
• distance_meters (total distance)
• ascent_meters
• distance_diff (distance from previous point)
• atl_start / ctl_start (atl and ctl before the ride)
• watt_kg
• temp

The training process looks ok, but I need to tweak the model a bit more, maybe use 5s avg power, as power is quite noisy.

[nfkb]

Hello, it would be interesting for you as a way of digging into energy expenditure to have a metabolic cart test done to draw fat/carb consumption for each power brackets.
I have done this and a little CIQ app in which I give these brackets now tell me how much carbs/fat I have burnt in real time.

It’s a bit old on my blog but it might give you some ideas : Xert Carb:fat model from Baron Biosystems is not for me ! – hic et nunc

[aguimorefran]

Wow, thanks! That’s very interesting. I will be checking it soon, as I am so lost on “health” formulas.