<p>Accurate dynamic modeling of racing drones, characterized by high speed and maneuverability, is challenging due to model uncertainty stemming from personalized modifications and frequent in-flight collisions. Although deep neural network-based methods have shown some effectiveness, they struggle with online adaptability as the system and environment change, and they present difficulties in analysis. To address these challenges, we propose a novel semi-adaptive spectrally normalized neural network (SASNNet) to characterize model uncertainty. SASNNet learns long-term features representing inherent operational dynamics through offline training, while online training enables it to capture short-term features reflecting system changes, enhancing its adaptability. Additionally, spectral normalization is integrated into the training process to improve SASNNet’s Lipschitz stability, and an adaptive parameter update rule is designed to accelerate the model response. Building on this uncertainty characterization approach, we develop a control compensation method for trajectory tracking in racing drones. We validate the proposed method through physics-engine-based simulations, with results demonstrating high modeling accuracy, enhanced adaptability, and fast response speed.</p>