To address the problem of low prediction accuracy caused by the non-linearity, volatility, and multi-source influencing factors in the carbon emission data series of the transportation industry, a carbon emission prediction model based on a hybrid feature selection (RF-MIC), Improved Variational Mode Decomposition (IVMD), the Animated Oat Optimization algorithm (AOO), and a Bidirectional Long Short-Term Memory network (BiLSTM) is proposed. Firstly, a hybrid feature selection strategy based on Random Forest (RF) and the Maximal Information Coefficient (MIC) is constructed to screen out key influencing factors and eliminate redundant features simultaneously. Secondly, a data decomposition theory of Variational Mode Decomposition (VMD) based on the Escape Optimization Algorithm (ESC) is developed. The original carbon emission sequence is decomposed into a series of stationary modal components, by which its non-linearity and volatility are weakened. Then, a parameter optimization theory for BiLSTM based on AOO is established, where the optimal hyperparameters of the BiLSTM network are searched for by AOO to prevent the parameters from falling into local optima. Finally, prediction models based on AOO-BiLSTM are constructed for each modal component respectively, and the prediction results of all components are integrated and reconstructed to obtain the final predicted value. The proposed model is validated using the carbon emission data from China’s transportation industry for the period 1990-2023. The results show that, compared with the best-performing comparison model, the RMSE, MAE, and MAPE of the proposed model are reduced by 35.77%, 40.48%, and 59.52%, respectively. It is demonstrated that the carbon emissions of the transportation industry can be predicted effectively by the proposed model.