Abstract In North China, pollen stands as a leading allergen responsible for allergic rhinitis, with climate change exacerbating allergenic pollen sensitization and posing significant health risks to residents. Despite its critical importance, pollen forecasting technology is still not sufficiently optimized. This study leverages multi-year daily pollen concentration observations and ECMWF (European Centre for Medium-Range Weather Forecasts) real-time forecast data, applying twelve machine learning models to learn perturbations separated from characteristic quantities. Specifically, it forecasts pollen concentrations in Beijing, utilizing R2 and RMSE as evaluation metrics. The findings reveal that the CatBoost, Extra Trees, and XGBoost algorithms perform well for three-day consecutive pollen predictions. Specifically, when considering a one-day prediction period, the R2 values for these algorithms are 0.72, 0.73, and 0.73, respectively. In contrast, algorithms such as Neural Network, LightGBM, and K-nearest Neighbor demonstrate weaker performance, though all models except Neural NetTorch achieve R2 values above 0.50. Notably, the prediction accuracy of Neural NetTorch significantly improves with extended prediction time, with its R2 increasing from 0.34 to 0.67 as the prediction period extends from one day to three days. The Weighted Ensemble model, which adjusts other models based on weighted optimization to mitigate excessive peaks, consistently yields stable results with an R2 exceeding 0.67. Furthermore, the study assesses the importance of feature groups within the model, indicating that pollen emission intensity and phenological characteristics are crucial for both training and testing phases, whereas meteorological factors predominantly influence pollen dispersion. Given the strong impact of meteorological conditions and nonlinear regulation on pollen, a type of bioaerosol, machine learning demonstrates substantial potential for simulating and predicting its concentrations.