Machine Learning Techniques for Performance Analysis in Formula 1

The introduction of the 2026 Formula 1 regulations brings a major redefinition of the power-unit, with increasingly complex hybrid energy states whose optimal deployment will play a decisive role in performance. This thesis addresses such developments in the context of performance analysis, focusing on automated methods to model and predict hybrid-system behaviour under varying operating conditions. Two complementary approaches are developed, taking the Audi Formula Racing hybrid control logic as reference. The first is a rule-based framework, where threshold and banded rules are discovered through mathematical optimization algorithms and validated with a Leave-One-Lap-Out procedure to ensure generalization across scenarios. The second is a machine learning framework, where calibrated ensemble models are trained on selected features to predict system states and performance metrics with probabilistic outputs. In both cases, modular pipelines were designed to guarantee reproducibility and rapid deployment to unseen conditions. Results show that rule-based scanning provides interpretable decision criteria but fails to generalize when applied to new parameter sets. Machine learning models, by contrast, generalize more robustly, making them well suited for deployment in simulation-driven performance studies where operating conditions may differ substantially from the training reference. A suite of visualization tools was also developed, including activation timelines, usage distributions, and probability traces that reflect prediction confidence, providing both quantitative and qualitative insights into hybrid-system behaviour. Overall, this work provides a methodological foundation for performance analysis in Formula 1, combining interpretability, generalization, and practical applicability to the upcoming regulatory era.