POA-CNN-BiGRU模型下的港口货物吞吐量预测

马赜湫，周兆欣^*，常育苗

山东交通学院航运学院，山东 威海  264200

摘要：为准确预测港口货物吞吐量，提出基于鹈鹕优化算法（pelican optimization algorithm，POA）优化卷积神经网络-双向门控循环单元（convolutional neural network-bidirectional gated recurrent unit，CNN-BiGRU）的POA-CNN-BiGRU组合预测模型。通过POA对CNN-BiGRU模型的关键超参数进行自适应寻优，将得到的最优超参数组合输入CNN-BiGRU模型进行模型训练后用于预测港口货物吞吐量，选取青岛港2001—2023年港口货物吞吐量季度数据进行实证分析。结果表明：POA在迭代过程中具有较强的全局搜索能力和较快的收敛速度；POA-CNN-BiGRU模型预测结果曲线的变化轨迹与实际结果曲线吻合度较高，能准确捕捉港口货物吞吐量的动态变化趋势；与BiGRU模型、CNN-BiGRU模型、POA-CNN-BiLSTM模型相比，POA-CNN-BiGRU模型的平均绝对百分比误差、均方根误差、平均绝对误差均最小，拟合优度最大；采用该模型预测上海港和天津港的货物吞吐量，预测精度仍较高，预测性能较稳定，说明其普适性和鲁棒性较强。POA-CNN-BiGRU模型能有效捕捉港口货物吞吐量的特征，预测误差较小，预测精度较高。

关键词：港口货物吞吐量；预测模型；POA；CNN；BiGRU

Port cargo throughput forecasting based on the POA-CNN-BiGRU model

MA Zeqiu, ZHOU Zhaoxin^*, CHANG Yumiao

School of Navigation and Shipping, Shandong Jiaotong University, Weihai 264200, China

Abstract: To accurately forecast port cargo throughput, a hybrid forecasting model is proposed that integrates the pelican optimization algorithm (POA) with a convolutional neural network-bidirectional gated recurrent unit (CNN-BiGRU), denoted as POA-CNN-BiGRU. POA is employed to adaptively search for key hyperparameters of the CNN-BiGRU model; the resulting optimal hyperparameter set is then used to train the CNN-BiGRU, after which the trained model is applied to forecast port cargo throughput. An empirical analysis is conducted using quarterly throughput data for Qingdao Port from 2001 to 2023. The results indicate that POA exhibits strong global search capability and rapid convergence during iterations. The POA-CNN-BiGRU model accurately captures the dynamic evolution of port cargo throughput, with predicted curves closely aligning with the actual series. Compared with BiGRU, CNN-BiGRU, and POA-CNN-BiLSTM models, POA-CNN-BiGRU achieves the smallest mean absolute percentage error, root mean squared error, and mean absolute error, and the highest goodness of fit. When applied to forecasting cargo throughput at Shanghai and Tianjin ports, the model maintains high accuracy and stable performance, demonstrating strong generalizability and robustness. Overall, the POA-CNN-BiGRU model effectively captures the characteristics of port cargo throughput, yielding small forecasting errors and high predictive accuracy.

Keywords: port cargo throughput; forecasting model; POA; CNN; BiGRU