Long-range dependency processing is the core challenge of sequence modeling, and traditional sequence modeling algorithms suffer from information decay and low parallel computing efficiency in ultra long sequences. Therefore, a bidirectional composite long short-term memory network model integrating dynamic attention mechanism is proposed to optimize sequence modeling algorithm and enhance long-range dependency capture capability. Firstly, the study enhances feature extraction capability through forward backward feature concatenation. Secondly, a dynamic attention module is introduced to dynamically calibrate key feature weights using global pooling, significantly reducing redundant calculations. Finally, a feature fusion mechanism is used to further integrate long short-term dependencies and strengthen information correlation across time steps. The experimental findings reveal that in the test dataset, the training loss approaches zero after 200 iterations. The accuracy of long-range data recognition reached 98.43% and 98.27%, respectively. The accuracy of long-term data recognition remains stable at 96.14% -97.63%, and visual weight analysis confirms its ability to accurately capture key time slice dependencies. The outcomes reveal that the research design method can significantly improve the efficiency of long-range dependency processing and model robustness. The research provides high precision and achieves a good balance between high accuracy and computational efficiency for scenarios that require long-range dependency processing, such as power load forecasting and traffic flow analysis.
Lijie Li (Tue,) studied this question.