基于Transformer-MLP的轮渡码头短期客流量预测方法  

A Transformer-MLP-Based Method for Short-Term Ferry Terminal Passenger Flow Prediction

作　　者：林哲显 LIN Zhexian(Shanghai Ship and Shipping Research Institute Co.,Ltd.,Shanghai 200135,China)

机构地区：[1]上海船舶运输科学研究所有限公司,上海200135

出　　处：《上海船舶运输科学研究所学报》2024年第6期36-42,共7页Journal of Shanghai Ship and Shipping Research Institute

摘　　要：为提升轮渡码头的客流量感知能力,促进轮渡码头的数字化发展,提出一种基于Transformer模型与多层感知机(Multi Layer Perceptron,MLP)组合的轮渡码头短期客流量预测方法,支持预测轮渡码头未来1 d内4个时段的客流量。该方法充分考虑轮渡码头业务数据的价值,通过ETL(Extract-Transform-Load)的方式将业务数据转化为模型输入的时序客流量数据,并利用NeuralForecast时间序列预测工具库构建由PatchTST模型与TiDE(Time-Series Dense Encoder)模型构成的组合模型PatchTSTiDE。对该模型与其他常见时间序列模型进行对比试验,结果显示该模型对客流量预测的平均绝对误差相比TiDE模型和PatchTST模型能分别下降约4.44%和30.59%。该研究可供轮渡码头的客流量预测参考。In order to augment the traffic perception capability of the ferry terminal and improve the overall digitalization level,a Transformer-MLP(Multi Layer Perceptron)-based method for short-term prediction of ferry terminal passenger flow is introduced,The design is capable of predicting the passenger flow in 4 predefined time periods of following day.The prediction method performs ETL(Extract-Transform-Load)and get passenger flow data in the form of time series from ferry business data.A combined time series model PatchTSTiDE is built based on the collected time series data transformed from ferry business data by means of the time series forecasting toolkit NeuralForecast.Tests show that,Compared with TiDE(Time-Series Dense Encoder)and PatchTST,the developed model is about 4.44%and 30.59%better in MAE(Mean Absolute Error)respectively.

关 键 词：轮渡码头 客流量预测 时间序列预测 

分 类 号：U696.1[交通运输工程—港口、海岸及近海工程]