Diabetes affects approximately 14% of the global population. Accurate blood glucose forecasting (BGF) is crucial for diabetes management and helps prevent dangerous glycemic fluctuations. Nonetheless, significant challenges arise due to the inherent heterogeneity among patients and the disparities in data distributions, which hinder model personalization and reduce training efficiency. This study proposes the Directional Representation Encoder-Decoder (DRED), a novel personalized time series model, to address the challenge of personalized blood glucose forecasting (PBGF). The DRED framework leverages directional representations (DR) to capture individual glycemic dynamics. The model leverages an advanced encoder-decoder (ED) framework that incorporates a DR module, capturing complex temporal dependencies and patient-specific traits. Experiments on two datasets show that DRED outperforms twenty recent methods. It achieves a mean absolute error (MAE) of 6.79 mg/dl and RMSE of 11.74 mg/dl on KDD18, and an MAE of 6.74 mg/dl and RMSE of 10.18 mg/dl on CDD.