Abstract: The metro door actuation system is a typical multi-speed electromechanically coupled actuation system that frequently undergoes state transitions during operation. It exhibits pronounced nonlinearity and time-varying dynamics and is highly sensitive to environmental disturbances and passenger behavior，making accurate short-term prediction of key behavioral variables challenging. To enhance the prediction capability of short-term transient behaviors and the anticipatory control capability of the system，a behavior digital twin model is developed that integrates physics-based mechanism modeling with data-driven learning，enabling prediction of key behaviors—such as motor current and door displacement—within a 0-0.5 second horizon. First，an electromechanically coupled dynamic model is developed based on torque balance，transmission efficiency，and load response，providing a high-fidelity representation of the system’s motion behavior. Second，a hidden state identifier based on a pattern search algorithm，together with an action-cycle-level model-parameter updating mechanism，is proposed to achieve equivalent inverse estimation of the state parameters. Third，a physics-guided deep learning prediction model (PG-DLPM) is constructed by integrating a temporal convolutional network with a transformer encoder，improving both prediction accuracy and model interpretability. Finally，a door-motion-based consistency metric is developed，and the digital twin is driven by monitoring signals (e.g.，motor current and lead-screw rotation angle) to generate high-fidelity state-parameter trajectories，which are fused with measured data as inputs to PG-DLPM for high-accuracy short-term transient prediction. Experimental validation using field-measured data collected from a metro line shows that the proposed twin model exhibits high consistency with the physical system in dynamic behavior. Compared with state-of-the-art time-series forecasting models，the proposed PG-DLPM reduces the average root-mean-square error and mean absolute error by 37.45% and 46.05%，respectively，while their corresponding standard deviations decrease by 63.13% and 79.01%. These results demonstrate superior predictive accuracy and robustness，supporting intelligent control and decision-making for metro door actuation systems.

Key words: multi-speed electromechanically coupled actuation system, behavior digital twin, short-term transient behavior prediction, physics-guided deep learning, hidden state identification