Empirical Study on Traffic Flow Prediction and Route Optimization in Chengdu City

This study presents an empirical investigation of urban traffic flow forecasting and route optimization using over 1.4 billion GPS trajectory records collected from more than 14,000 taxis in Chengdu between 24th and 30th August 2014. Applying the DBSCAN clustering algorithm, a high-density hotspot region was identified, centered on the city center (approximately 104.05°E, 30.65°N), highlighting pronounced spatial clustering in taxi activity. For taxi demand forecasting, a multivariate LSTM neural network model was developed, achieving low prediction errors on the test set (MAE ranging from 0.14 to 0.17) and accurately capturing spatio-temporal demand patterns across different zones. For example, the model successfully forecasted nearly 5,500 demand instances during the evening peak in the High-Tech Zone, closely matching the observed peak of approximately 6,000 instances, while discrepancies of around 500 instances occurred in areas such as Chenghua District. Additionally, the Isolation Forest anomaly detection model identified sudden acceleration (763 instances) and high-speed driving (593 instances) as the primary abnormal driving behaviors, with a higher incidence observed during empty-vehicle states. These findings offer a reliable data-driven foundation for intelligent urban taxi dispatch, transport resource optimization, and traffic safety management.