Application of a Novel Deep Learning Model Based on CCVMD-STL Secondary Decomposition in Short Calendar Time Forecasting of Port Container Throughput: A Case Study of Shanghai Port

Accurate forecasting of port container throughput (PCT) is of great significance to port shipping enterprises and government administrations in making scientific decisions. The current research methods pay less attention to the short-calendar-time PCT, and have limited prediction accuracy for non-linear and non-smooth fluctuation series. In this paper, the container throughput of Shanghai Port is taken as the object of analysis and forecasting, and a new deep learning model based on secondary decomposition of CCVMD and STL is proposed. The primary decomposition is achieved by combining the VMD with the correlation coefficient to screen the decomposition modes more objectively and to achieve a better effect of noise removal; the secondary decomposition is achieved through the STL, and the seasonal, trend and irregular terms are obtained, which further captures the time-series. The HHO algorithm is used to optimise the hyperparameters of LSTM, and then the optimised LSTM predicts the decomposition terms separately, and the final prediction results are summarised. The results show that the decomposed combinatorial models with data preprocessing significantly outperform other models in container throughput prediction; the model proposed in this paper outperforms the remaining 12 comparative models; and the secondary decomposition preprocessing is more reliable in tracking the extremes, removing and reducing noise, and interpretability.