融合预训练的港口吞吐量LSTM预测模型

港口吞吐量时序变化数据量较小且变化快，传统LSTM神经网络在此类数据上易出现过拟合，导致模型预测性能不佳。针对此问题，本文提出融合预训练与LSTM时序模型，通过预训练捕获任务领域的全局信息，再用LSTM模型精确描述各个港口的吞吐量变化规律，以提升模型对全部港口吞吐量预测的准确性。以天津港等15个中大型港口过去二十一年的月吞吐量为实验数据，以BP、ARIMA、传统LSTM等预测模型和目前流行的GNN-LSTM模型为比较基准进行仿真实验，结果显示本文所提出的融合预训练的LSTM模型能有效解决LSTM神经网络的过拟合问题，整体预测准确率高于所有基准模型。与传统LSTM模型相比，基于预训练的LSTM的MAE指标平均降低45.2%%，最多降低80.0%。

Integrate Pretraining with LSTM for cargo throughput forecasting

For the cargo throughput forecasting task, traditional methods usually built models from historical throughput data collected on a specific port. Such data was usually limited and varied wildly due to complex geographical and economic factors. Recent advances in neural networks such as the LSTM network for time series prediction require abundant training data, and thus are prone to overfit in this task. To address this problem, pre-training was proposed to enhance the standard LSTM model. Pre-training was performed with all available ports and then an LSTM model was fine-tuned to accurately describe the throughput variation patterns of each individual port. Experimental results on the monthly throughput of 15 medium and large ports in mainland China in the past 21 years showed that the proposed model outperformed traditional time series prediction models, including BP, ARIMA, standard LSTM model without any pre-training and currently popular GNN-LSTM model. Specifically, the overfitting problem of standard LSTM was observed on several ports, and experimental results showed that the fused model could effectively solve overfitting, reducing the MAE index value by 45.2% on average and 80.0% to the maximum.