Diabetes mellitus remains a critical global health issue, affecting nearly 10% of the adult population. Machine learning methods are increasingly used to forecast glucose levels as an early warning indicator. However, existing models often rely solely on historical glucose data, limiting their integration with physiological parameters and thus reducing forecasting reliability. This research proposes integrating physiological modelling with neural network architectures to enhance accuracy and relevance in predicting glucose. By using a differential equation-based model of post-prandial glucose response and embedding it within a neural network, this study aims to create a robust forecasting system that aligns closely with physiological processes. Our method reduces forecasting error scores in glucose value prediction, addressing the limitations of current models and contributing to more personalized and effective diabetes management.