Research On Short-time Passenger Flow Prediction Of Urban Rail Transit Based On Deep Learning

Urban rail transit,with its speedy,low-carbon and eco-friendly characteristics,is highly favored by travelers and has become an important component of modern transportation.With the rapid increase of urban population,urban rail transit system is facing huge pressure of transportation.Therefore,scientific management planning of urban rail transit system is needed to effectively improve its operation efficiency and service quality.In the era of big data,various front-end sensing facilities and automatic fare collection(AFC)systems can collect a large amount of passenger flow data.It is an important research topic to fully explore the intrinsic information of passenger flow data and make accurate prediction,to provide basis and reference for urban rail transit system operation.Short-time prediction of urban rail transit passenger flow is an important component of traffic management and planning,and is also an important application area of deep learning technology.This thesis summarizes the advantages and limitations of domestic and foreign traffic flow prediction models,and start the research on urban rail transit passenger flow short-time prediction model based on deep learning technology from the spatial and temporal characteristics of urban rail transit passenger flow data,the main work is as follows:(1)Passenger flow data preprocessing and spatiotemporal feature analysis.Firstly,this study defined the problem of urban rail transit passenger flow prediction,and clarified the specific objectives and methods of the deep learning model to achieve passenger flow prediction.Then,the Hangzhou Metro swipe card data was pre-processed and integrated to form the dataset required for deep learning model training.The changing characteristics of passenger flow were analyzed from the temporal and spatial dimensions of urban rail traffic passenger flow data,and the feature extraction capability required for the prediction model was discovered.For the characteristics of urban rail traffic passenger flow,a dual-level combination of static topology and dynamic spatio-temporal correlation of rail network was proposed to learn the complex correlation between stations and enhance the interpretability of the model.(2)Built two short-term passenger flow prediction models using different architectures based on deep learning techniques.The models utilized innovative design features such as graph convolution operations that combined adaptive and diffusion convolutions,loss functions improved by the Soft-DTW algorithm,and local attention mechanisms that extracted topological and traffic semantic features of the orbital network.The operation modules were combined with different feature dimensions and multi-step prediction approaches,depending on the model architecture.As a result,two types of urban rail transit short-term passenger flow prediction models were designed using deep neural networks:Adaptive Diffusion Convolutional Recurrent Neural Network(ADCRN)and Dynamic Spatio-Temporal Neural Network(DSTNN).Both models could extract and learn complex spatio-temporal features of data,enabling accurate short-term passenger flow prediction.(3)The proposed model effect was validated by real passenger flow data of Hangzhou Metro.A reasonable experimental process and optimization strategy were designed based on the experimental objectives and data characteristics.The model structure and parameters were trained and debugged accordingly.Various baseline models were chosen as comparisons,and the prediction performance of each model was analyzed from different prediction scales and evaluation indices under the same experimental environment.The results indicate that the proposed model has excellent and efficient prediction capabilities.An ablation experiment was conducted to explore the contribution of each basic module in the proposed model.The experiment found that the prediction accuracy improved with the introduction of each module,confirming the effectiveness of each module and the rationality of the model structure.Finally,a visual analysis of the model parameters and prediction curves demonstrated the powerful feature extraction ability and accurate prediction effect of the proposed model.