With the development of mobile Internet and digital technologies, online education platforms transcend time and space constraints to provide ubiquitous learning environments. However, high dropout rates and low pass rates pose a great challenge. Predicting student performance enables early identification of academic failure tendencies, facilitating timely intervention by educators. The continual development of online education generates data characterized by high dimensionality, dynamics, and nonlinearity, posing challenges to the predictive modeling of academic performance. This study proposes a Learning Behavior Time Window and Type Information (LBTT) model. Unlike traditional LSTM and CNN models that process sequential data uniformly, LBTT employs a novel dual-attention mechanism that analyzes both temporal patterns within specified time windows and correlations between different types of learning behaviors. The model computes temporal relationships by tracking similar behavioral patterns across multiple time windows while simultaneously analyzing correlations between different behavior types within each window. Model validation was performed using the Open University Learning Analytics Dataset (OULAD). The results show that LBTT achieves 93.6% accuracy and 95.6% F1 score, outperforming traditional models by an average of 3.30% in accuracy and 4.41% in F1 score. Ablation experiments on the time-window mechanism for learning behaviors and the behavior correlation computations validate the effectiveness of the proposed approach.