Air pollution continues to be a critical public health and environmental challenge, particularly in fast-growing urban areas. This study presents an interpretable, multi-horizon forecasting framework for PM2.5 concentrations in Prishtina, the capital of Kosovo. Using hourly observations from 2018 to 2024, the study evaluates the predictive performance of five machine learning models: XGBoost, LightGBM, Random Forest, Support Vector Machine, and Linear Regression. Feature engineering, incorporating pollutant lags, rolling statistics, and cyclical time encoding on model performance, was investigated. The results show that among the selected ML models, XGBoost achieves the best one-hour forecast with R² of 0.862, MAE of 3.524, and RMSE of 6.513, while maintaining reasonable accuracy, with R² of 0.50 even at 24-hour horizons. To promote transparency, the study employs SHAP (SHapley Additive exPlanations) to quantify feature importance across different forecast horizons. Key drivers include recent PM2.5 lags, wind speed, and meteorological indicators. The proposed framework offers a robust, scalable, and interpretable approach for predicting air pollution, thereby supporting efforts to reduce emissions in Prishtina and similarly affected urban environments, enabling real-time alerts and data-informed environmental policy planning. Scientifically, this study uniquely integrates multi-horizon forecasting using advanced ML models with detailed temporal feature engineering and SHAP interpretability to reveal temporal shifts in feature importance, previously unaddressed systematically in air pollution modeling literature. These insights significantly enhance the understanding of dynamic air pollution interactions and are broadly applicable to urban environments globally with analogous pollution and meteorological dynamics.