A Novel Progressive Dual-Attention Residual Network with Bidirectional Long Short-Term Memory for Interpretable Traffic Flow Prediction

Traffic flow prediction constitutes a fundamental operational requirement for both Intelligent Transportation Systems (ITS) and urban planning systems. This research provides a detailed methodology integrating deep learning together with optimization features along with extraction methods and enhanced preprocessing techniques to enhance traffic flow prediction precision. An enhanced asymmetric Gaussian filtering method during preprocessing achieves successful reduction of noise while sustaining important patterns of traffic behavior. A VGG19- based Deep Sparse Auto encoder serves as a tool to extract features which makes it possible for deep learning to reveal complex traffic flow patterns. The prediction model implements its functionality through PDR-BiLSTM and its architectural components. Through its residual network structure and BiLSTM units, the model retains long-term dependencies in data patterns for stable predictions to be achieved in addition to utilizing dual-attention methods that help focus on essential spatial and temporal traffic characteristics. The Augmented Snake Optimizer uses parameter adjustments for model performance optimization to achieve prediction accuracy optimization of PDR-BiLSTM models. The experimental research indicates the proposed method achieves outstanding results by maintaining the lowest error rate with 99.5% accuracy and 99% precision recall F1-score.