<p>The operation safety and stability of trains is closely related with the wind speed. However, given the intricate nature of its characteristics, which encompass linearity, nonlinearity, nonstationarity etc., accurately predicting the short-term wind speed presents a notable obstacle. To this end, this paper presents a novel forecasting approach using the hybrid of enhanced variational mode decomposition (EVMD), auto-regressive integrated moving average (ARIMA), fully convolutional neural network (FCN), and physical auxiliary mechanism (PAM). This method not only can provide the accurately deterministic prediction, but also can produce the desired probabilistic prediction. Specifically, EVMD is developed based the mode aliasing problem for performing the data decomposition and reconstruction. Then, the combination of ARIMA and FCN is used to perform linear and nonlinear predictions. Finally, PAM is introduced into the above established model for realizing the desired deterministic and probabilistic predictions where the relationship among the wind speed data recorded at various time intervals and the data variability are considered. Numerical examples, utilizing two sets of measured wind speed data, underscore the efficacy and advantage of the developed method. For example, the proposed method can realize the reduction of the average of mean absolute error from 1.08 to 0.73 in comparison with ARIMA-FCN-PAM. Hence, the proposed method stands as a viable and efficient alternative for forecasting the short-term wind speed.</p>