Abstract: Accurate forecasting of port container throughput is of great significance for port operators and government administrations in making scientific decisions. Existing forecasting methods,however,often pay insufficient attention to short-calendar-time PCT and exhibit limited accuracy in handling nonlinear and non-stationary fluctuation series. This paper takes the container throughput of Shanghai Port as the research object and proposes a novel deep learning model based on secondary decomposition using CCVMD and STL. Using the correlation coefficient as a reference,variational mode decomposition is first applied to the original time series. Subsequently,a secondary decomposition divides the data into seasonal,trend,and residual components. An algorithm-optimized long short-term memory neural network is then employed to predict each component separately,and the final prediction results are aggregated. Experimental results show that the combined decomposition model with data preprocessing significantly outperforms other models in PCT forecasting. The proposed model achieves a mean absolute percentage error of 0. 021 703,a root mean square error percentage of 0. 026852,and a mean absolute error percentage of 0. 022 14,indicating superior overall performance compared to 12 benchmark models and several models from prior studies. Furthermore,the secondary decomposition approach demonstrates enhanced reliability in tracking extreme values,removing and reducing noise,and improving interpretability.