Novel Decomposition and Ensemble Model with Attention Mechanism for Container Throughput Forecasting at Four Ports in Asia

Abstract

Container shipping has suffered a sharp decline since COVID-19, and risks associated with container transit will persist in the future. The decrease in container transportation has caused a ripple impact on the global supply chain. However, container throughput forecasting is both critical and complicated under the circumstances of economic uncertainty and the outbreak of the COVID-19 pandemic. A novel model propounded in this paper for container throughput forecasting to assist the port management bureau and container shipping industry integrates with the variational mode decomposition (VMD) algorithm, SARIMA technique, convolutional neural network (CNN) method, long short-term memory (LSTM) approach, and attention mechanism, among others. In this model, there are three stages: (i) data decomposition, (ii) component prediction, and (iii) ensemble output. In the first stage, the original data of the container throughput time series is decomposed into several different components using the VMD algorithm. Next, from low frequency to high frequency, each component is modeled by the corresponding prediction approach. Subsequently, the prediction results of each component generated by the previous stage are integrated into the final forecasting results by addition strategy. To enhance the prediction accuracy in the second stage, the attention mechanism is adopted in the CNN-bidirectional LSTM method. Finally, six measurement criteria, the container throughput times series at four ports, and a statistical evaluation approach are applied to comprehensively evaluate the proposed model compared with seven benchmark models. The empirical analysis demonstrates that the proposed model significantly outperforms other comparable models with regard to prediction results, level, and directional prediction accuracy.