<p>Accurate predictions of the air quality index (AQI) are crucial for urban decision-makers to formulate appropriate environmental management strategies. Despite significant progress in AQI prediction methods, many existing models face challenges in addressing the non-linear and non-stationary characteristics of AQI time series data, which limits their accuracy and interpretability. To address this gap, a hybrid prediction model based on Empirical mode decomposition (EMD), Bayesian optimization algorithm (BO), and XGBoost considering interpretability through Shapley additive expansions (SHAP) was proposed, which improves the accuracy and interpretability of AQI prediction by revealing the relationship between influencing factors and AQI. Firstly, we integrate the air quality data, and analyze the nonlinear and non-stationary time series characteristics. Secondly, EMD is performed on the AQI sequence to suppress the impact of white noise on the reconstruction error. Subsequently, the XGBoost model is constructed to predict the subsequences derived from the EMD decomposition, and the core parameters are optimized using BO. Finally, interpretability analysis is conducted on the prediction results of the optimal-hybrid model from both global and local perspectives. Compared with the sub-optimal model, this model reduces RMSE, MAE and MAPE metrics by 29%, 9% and 4%, respectively, while also reducing the runtime by 41%. In addition, the effectiveness and data adaptability of the proposed model were validated on the dataset from another region. This study provides valuable data support and theoretical basis for air quality assessment through the prediction and explanatory analysis of air quality index.</p>