Research on Time Series Forecast of Container Throughput Based on Selective Deep-Ensemble Model

Abstract

Accurate forecast of container throughput is an important basis for reasonably planning port construction, making port operation plan and adjusting port development direction. Aiming at the complex nonlinear characteristics of port container throughput, a hybrid model based on selective deep-ensemble for container throughput forecasting (HMSD) is presented in this paper. Firstly, this model decomposes the original container throughput time series into several intrinsic mode functions and a residual by empirical mode decomposition (EMD). Considering highly nonlinear characteristics of each intrinsic mode function, the proposed model constructs three deep neural networks, namely, long short term memory (LSTM), gated recurrent unit (GRU) and convolutional neural network (CNN), as base learners to predict intrinsic mode functions. Then, this model establishes selective deep-ensemble forecasting model by improved group method of data handling (GMDH) on intrinsic mode functions and obtains their ensemble forecasting results. Furthermore, this model uses an autoregressive integrated moving average model to predict the linear residual. Finally, the forecasting result of the total container throughput is obtained by integrating the forecasting results of all intrinsic mode functions and the residual. The empirical analysis of container throughput data of Xiamen port and Shanghai port in China shows that the HMSD model has better forecasting effect than other models.