Abstract: To address gradient vanishing or explosion, inadequate capacity for capturing long-term dependencies, and lack of interpretability in deep learning models when using recurrent neural networks for extracting features from long sequential student behavioral data, a method for student performance prediction and analysis is proposed based on dilated causal convolution (DCC) tailored for long sequential data. First, employing a generative adversarial network (GAN) to generate new samples of the minority class conforming to the distribution pattern of original student behavioral data, these new samples were merged into the student dataset to achieve class balance. Second, a prediction model based on DCC was introduced, wherein the combination of DCC and gated recurrent unit (GRU) enhances the model􀆳s capability to extract dependencies from long sequential data. Finally, the Shapley additive explanations method combined with three-factor theory was employed to analyze and interpret the importance of factors influencing student performance. The experimental results on public datasets indicate that the proposed method outperforms the baselines methods in student performance prediction by raising the weighted F₁ score by approximately 6% , and further validates the effectiveness of key modules proposed in the method and the generalization ability of the model. Additionally, the comparison of the learning characteristics between high-achieving and at-risk students show that factors such as good study habits, proactive classroom engagement, and different behavioral environments exert significant impacts on students􀆳 academic performance.