An Adaptive Equilibrium Estimation and Uncertainty Decomposition Method for Traffic State

The present invention discloses an adaptive equilibrium estimation and uncertainty decomposition method for traffic state, which pertains to the technical field of traffic management. The method primarily includes the following steps: first, constructing a dual-output neural network capable of simultaneously outputting the mean and variance of traffic state, and designing a hybrid loss function that integrates data and physical constraints. During training, the weights of loss components are adaptively adjusted by calculating the variance of their gradients to balance the optimization process. Furthermore, parameter updates are modeled as an underdamped Langevin dynamics process, where noise is injected for posterior sampling to obtain multiple sets of model parameters. Finally, multiple inferences are performed using these parameters, and through aggregation and statistical analysis of the prediction results, aleatoric uncertainty arising from data noise and epistemic uncertainty stemming from insufficient model cognition are separated and quantified. The invention achieves high-precision probabilistic estimation of traffic state under sparse observation data, providing a quantitative basis for traffic monitoring optimization and risk assessment.