This study explores the application of Tabular Neural Networks (TabNet) for forecasting student academic performance, addressing the limitations of traditional predictive methods that often rely on linear models or decision trees, which struggle to find complex patterns in tabular data. TabNet, a framework of deep learning designed specifically for tabular data, leverages attention mechanisms to effectively handle and understand multi-dimensional data features. The research involves collecting diverse student-related data, including demographic information, attendance records, and previous academic performance, to predict final grades and identify key factors influencing academic success. We implement TabNet to make predictions, that are then used as input to Extreme Gradient Boosting (XGBoost), creating a novel hybrid model approach. The performance of this combined approach is evaluated through precision, recall, accuracy, f1-score and cross-validation, demonstrating superior accuracy and interpretability compared to baseline algorithms. Findings show that in addition to improving prediction accuracy, the hybrid model offers valuable insights into significant predictors of student performance, offering implications for educators and policymakers in developing targeted interventions to improve academic outcomes.