Accurate prediction of blood glucose levels is essential for effective diabetes management, and explainability is equally crucial to help clinicians and patients trust and understand model decisions. Diabetes exists as Type 1 and Type 2, each with unique characteristics requiring personalized forecasting approaches. This study proposes an explainable CatBoost framework for glucose forecasting with prediction time horizons of 15, 30, 45, and 60 min on Shanghai diabetes datasets, addressing both Type 1 (T1DM) and Type 2 diabetes mellitus (T2DM). The method uses linear interpolation, sliding window techniques, and temporal-aware feature engineering, with model parameters specifically optimized for each diabetes type. Among all horizons tested, the 15-minute horizon performed the best, outperforming previous works, with patient-specific models showing strong performance with average Root Mean Square Error (RMSE) values of 11.83 mg/dL for T1DM and 6.99 mg/dL for T2DM. Furthermore, through SHapley Additive exPlanations (SHAP) analysis, it was observed that for T1DM models, recent glucose readings and rapid short-term changes contributed with large influence to predictions, while for T2DM models, recent glucose features also played key roles but with more moderate individual contributions. Finally, the study concludes with future directions, such as incorporating insulin dosing and meal intake data, and including patients from more diverse geographic areas to enhance prediction performance.