Abstract: In view of the fact that wind noise simulation is time-consuming and resource-intensive, and that the current intelligent prediction models—based on geometric deep learning—are overly complex and require large sample sizes (≥107), a wind noise prediction method combining characteristic parameters with machine learning was proposed. For the A-pillar, characteristic parameters were comprehensively identified and decomposed. A total of 14 characteristic parameters were selected; it was verified that these parameters uniquely and accurately reconstruct the external geometry of the A-pillar and are suitable for engineering wind noise analysis. These A-pillar characteristic parameters served as the input dataset, while the in-cabin 1/3-octave band sound pressure level spectra obtained from numerical simulations constituted the output dataset for training the machine learning model. Results show that the proposed method achieves highly accurate predictions relative to simulation benchmarks (R² ≥ 99.47%) using only a small sample size (~10²), making it applicable to automotive A-pillar wind noise prediction. Furthermore, a multi-parameter optimization algorithm was employed to achieve a 2.3% reduction in A-pillar wind noise (measured as AI—Articulation Index—or equivalent loudness metric) under realistic engineering constraints. Finally, design recommendations for key parameters were provided based on sensitivity analysis and practical engineering experience.