Modeling Cycling Performance

As part of my Sports Engineering master’s program, my cohort travelled to Delft for an intensive two-week course focused on cycling performance. Our main challenge was to model the real-world performance of one teammate and predict two lap times and a final coast-down distance at the Alkmaar Velodrome. The difficulty of the task came from the need to combine physiological capacity, aerodynamics, rolling resistance, and track dynamics into one predictive model.

To build this model, we started by characterizing the physiology of our chosen rider. Using VO₂max and Wingate tests, we measured mechanical power output and estimated both aerobic and anaerobic contributions. We derived critical power, fitted an aerobic asymptote, and used the shape of the Wingate curve to calculate anaerobic capacity from the area above the asymptote and aerobic capacity from the remaining area. This gave us a personalized power-duration relationship that represented the rider’s ability to generate and maintain power.

Next, we needed to understand the mechanical resistance acting on the rider. We determined frontal area using image-based pixel analysis and conducted coast-down and power-meter tests at different postures and tire pressures. However, the Cd and Crr values calculated from these experiments did not behave as expected and did not align with published values, so we ultimately used validated literature values for drag and rolling resistance in the final simulation.

With the physiological and mechanical components defined, we built a full track simulation to model the two laps and the coast-down. Using the processed Wingate power curve sampled at 10 Hz, literature-based Cd and Crr, and a segmented track model with cornering speed limits, we simulated the rider’s acceleration, velocity, and position through time. Lap one represented an all-out acceleration phase, lap two held the rider at the maximal speed achievable under the power curve, and the coast-down modeled passive deceleration under drag and rolling resistance alone.

When we compared our predictions to the real test at the velodrome, our lap times were within 4 percent of actual performance. In hindsight, this accuracy came from an unexpected cancellation of errors: the rider’s straight-line speeds were about 5 km/h slower than our model predicted, while their cornering speeds were about 5 km/h faster. These opposing differences produced an average lap speed that happened to match our model closely. The coast-down distance, however, was significantly off. The higher-than-expected cornering speeds allowed the real rider to lose velocity more slowly than our simulation assumed, and inaccuracies in the track dimensions amplified the error.

Overall, the project demonstrated how sensitive performance modeling is to the assumptions and input values behind it. Even with a detailed physiological model and a carefully constructed simulation, inaccuracies in drag, rolling resistance, speed limits, or geometry can shift results dramatically. The experience highlighted the need for models that are specific enough to represent the athlete but robust enough to withstand real-world variability.