Hybrid traffic flow prediction model for emergency scenarios with scarce historical data

Accurate traffic volume prediction under emergency conditions is critical for enhancing traffic control, disaster response, and route optimization for both government agencies and individual travelers. However, current prediction models often assume homogeneity between historical and future data, rendering them inadequate for anticipating unprecedented traffic disruptions during emergencies. To address this limitation, this study introduces a hybrid predictive model that integrates Prophet and DeepAR (Probabilistic Forecasting with Autoregressive Recurrent Networks) to improve traffic volume forecasting in such scenarios. Prophet captures long-term trends, seasonal variations, and holiday effects, while DeepAR is adept at modeling abrupt nonlinear changes induced by emergencies. This hybrid framework synergizes the strengths of both models, facilitating adaptive learning of anomalous patterns and striking a robust balance between regular traffic dynamics and emergency-induced disturbances. Through modular decomposition, integration of exogenous factors, and dynamic weight adjustments, the hybrid model offers enhanced flexibility and predictive accuracy compared to existing models, particularly under unprecedented conditions. The hybrid model was validated on highways across Luxembourg, Heilongjiang Province, and Minnesota. Results showed the hybrid model outperformed its components and other competing models under different emergency scenarios. In the Luxembourg pandemic scenario, the model reduced mean absolute error (MAE) by 73.60%, root mean square error (RMSE) by 74.22%, and coefficient of variation (CV) by 31.52%, with a 98.55% improvement in R². Comparable enhancements were achieved in Heilongjiang's snowstorms and Minnesota's adverse weather. These results underscore the robustness of the hybrid model and its potential for real-time traffic management in emergency situations.