融合预训练的港口吞吐量LSTM预测模型  

Integrate Pretraining with LSTM for Cargo Throughput Forecasting

作　　者：张聪 许浩然 詹炜[1] 黄岚[1] ZHANG Cong;XU Hao-ran;ZHAN Wei;HUANG Lan(School of Compute Science,Yangtze University,Jingzhou 434023,China)

机构地区：[1]长江大学计算机科学学院,荆州434000

出　　处：《科学技术与工程》2023年第32期13910-13916,共7页Science Technology and Engineering

基　　金：中国高校产学研创新基金(2020ITA03012)。

摘　　要：港口吞吐量时序变化数据量较小且变化快,传统长短时记忆(long short term memory,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,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21年的月吞吐量为实验数据,以BP(back propagation)、自回归积分滑动平均模型(autoregressive integrated moving average model,ARIMA)、传统LSTM等预测模型和目前流行的图神经网络(graph nerual network,GNN)-LSTM模型为比较基准进行仿真实验。结果表明: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%。Cargo throughput data is usually limited and varied rapidly.The standard long short term memory(LSTM)neural network is prone to overfit on such data,resulting in unsatisfactory predictive performance.To address this problem,pre-training was proposed to enhance the standard LSTM model.Pre-training was firstly performed with all available ports and then an LSTM model was fine-tuned to accurately describe the throughput variation patterns of each individual port,increasing the overall predictive performance.Experimental results on the monthly throughput data of 15 medium to large ports in China's Mainland collected in the past 21 years show that the proposed model outperforms mainstream time series prediction models,including the back propagation(BP)neural network model,the autoregressive integrated moving average model(ARIMA),the standard LSTM model and the graph nerual network(GNN)-LSTM model.Experimental results show that the proposed fused model effectively solves the overfitting problem of the of standard LSTM model,reducing the mean absolute error(MAE)value by 45.2%on average and 80%to the maximum.

关 键 词：深度学习 港口吞吐量 时序预测 长短时记忆神经网络(LSTM) 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]