A Temporal Attention-Based SARIMA–BiLSTM Residual Learning Model Tuned by Grey Wolf Optimizer for Parallel Urban Traffic Forecasting

Accurate and timely traffic forecasting is essential for ensuring the efficiency, reliability, and safety of modern transportation networks. However, urban traffic exhibits high levels of temporal variability, spatial complexity, and nonlinear dependencies, posing challenges to traditional time series and machine learning models. In this study, a novel hybrid residual learning framework is proposed that integrates Seasonal Autoregressive Integrated Moving Average (SARIMA) with a Bidirectional Long Short-Term Memory (BiLSTM) network enhanced by a temporal attention mechanism. The SARIMA component captures underlying seasonal and linear patterns in traffic flow, while the residuals—representing unmodeled nonlinear dynamics—are learned by the BiLSTM module. To improve the model’s generalization and predictive accuracy, a Grey Wolf Optimizer (GWO) is employed for automated hyperparameter tuning of both SARIMA and BiLSTM components. Furthermore, the entire pipeline is adapted for parallel execution to improve scalability and reduce computational latency, enabling faster processing of large-scale traffic data. Experimental evaluations on real-world traffic datasets demonstrate the effectiveness of the proposed model, achieving a processing speedup of 10.11× and a forecasting accuracy of 99.31%, outperforming conventional single-threaded and baseline hybrid models. The results indicate that the proposed attention-enhanced SARIMA–BiLSTM model, optimized via GWO and accelerated through parallel processing, provides a robust and scalable solution for real-time urban traffic forecasting applications.