Abstract:

With the rapid development of international trade and the continuous improvement of port containerization level， its throughput has become a key indicator to measure a port's global commercial management and industrial development. Therefore， it is very important to construct an effective time series forecasting model to accurately predict port container throughput. A DKELM-based selective deep-ensemble model is proposed for container throughput forecasting （DSDE）. Firstly， the original time series is decomposed by the optimized variational modal decomposition method， and several intrinsic mode functions are obtained. Then， considering that most of the decomposed intrinsic functions are highly nonlinear， DKELM of different kernel functions is constructed as the benchmark model for prediction. Further， the new enhanced multilayered iteration algorithm of the group method of data handling （eMIA-GMDH） is introduced to carry out the selective deep ensemble on each intrinsic mode function. Finally， the final forecasting result is obtained by integrating all the forecasting results. To verify the effect of the forecasting model constructed in this paper， four evaluation indexes are introduced to carry out empirical analysis on the monthly container throughput data set of six ports. Through experiments， it can be concluded that the performance of DSDE is better than that of the three hybrid models based on a single model and the hybrid model based on the weighted average ensemble

Key words: container throughput, hybrid prediction model, deep neural network, selective deep ensemble