Professional cycling calendar optimization using Transformers and evolutionary algorithms

This Final Degree Project proposes an innovative methodology to optimize competition calendars in professional cycling teams by combining Transformer neural networks and multi-objective evolutionary algorithms (NSGA-II). The system aims to maximize the accumulation of UCI points and to balance the competitive workload among riders throughout the season. The predictive models, trained with historical data (2021-2025) on stage profiles, rider attributes, and results, achieve 77.9% Top 50 prediction accuracy.

The NSGA-II algorithm explores the solution space by generating a Pareto front. The solutions improve the estimation of UCI points and enhance calendar equity compared to the actual calendar used between January and May 2025. Although empirical validation does not confirm statistically significant improvements over the team's actual calendar, the results show high potential as a strategic decision-support tool. The system identifies opportunities in underestimated races (e.g., 2.Pro/2.1 categories) and prioritizes riders based on their profiles. Developed with open data and reproducible techniques, this solution provides a scalable foundation for data-driven decision-making in complex sports contexts.