考虑波动特征相似度的地铁客流预测模型迁移学习方法OA

[Background]The continued expansion and optimization of urban rail transit networks has led to a marked growth in the number of metro stations has increased substantially.This growth demands the development of individualized passenger flow prediction models for hundreds of sta-tions,each with distinct ridership patterns.Conventional station-by-station training methods are time-consuming and fail to meet the stringent temporal efficiency requirements of contemporary transit management systems.[Objective]To address this issue,we explored transfer learning techniques for passenger flow prediction models,training them exclusively on select prototype stations.Subsequent-ly,the learned parameters are transferred to the remaining stations through transfer learning,enhanc-ing the prediction efficiency throughout the network.[Method]We propose a transfer learning-based prediction model that considers similarities in passenger flow fluctuation characteristics.By extract-ing each station's passenger flow fluctuation characteristics,we compute similarity scores with other stations and assign each to the most similar station as its transfer learning target,achieving a"one sta-tion,one solution"strategy.This method requires extensive prediction model training for only a few stations while allowing others to leverage transfer learning.[Data]For analysis,the Chengdu Metro served as a case study,utilizing passenger flow data of selected stations on Lines 1,2,3,and 4 from June to August 2022.[Conclusion]We evaluated four deep-learning passenger flow prediction mod-els.Transfer learning reduced training time by an average of 62.68％,with some stations experienc-ing over 80％reduction.Despite the decrease in training duration,prediction accuracy slightly im-proved across all models,confirming the efficacy of this method.