21252055-Xu Kaimin-Research on Integrated Prediction of Multi-Modal Transportation Demand for Airport Landside Driven by Multi-Source Data.docx

Undergraduate thesis (design)

Research on Integrated Forecasting of Multimodal Transport Demand on Airport Landside Driven by Multi-source Data

Integrated Forecasting of Multimodal Transport Demand on Airport Landside Driven by Multi-source Data

School: Transportation

Major: Transportation (Sino-Foreign Cooperation)

Student Name: Xu Kaimin

Student ID: 21252055

Supervisors: Huang Ailing, Gao Jie

Beijing Jiaotong University

2025Year5Month

北京交通大学毕业论文（设计）版权使用授权书

Bachelor's Thesis Copyright Authorization Letter

The author of this bachelor's thesis fully understands the regulations of Beijing Jiaotong University regarding the retention and use of bachelor's theses. Beijing Jiaotong University is hereby authorized to compile all or part of the thesis content into relevant databases for retrieval, provide reading services, and use photocopying, microfilming, scanning, or other reproduction methods to preserve and compile the thesis for reference and borrowing purposes.

(This authorization statement applies to confidential dissertations after declassification)

Signature of Thesis Author: SupervisorSignature:

Signing Date: YearMonthDaySigning Date: YearMonthDay

北京交通大学毕业设计（论文）论文诚信声明

Bachelor's ThesisIntegrity Statement

I hereby declare that the submitted graduation thesis (design), with the title, represents the results of my independent research work conducted under the guidance of my supervisor. To the best of my knowledge, the thesis does not contain any research results that have been published or written by others, except for those specifically noted and acknowledged in the text. It also does not include materials used for obtaining degrees or certificates from Beijing Jiaotong University or other educational institutions.

If there are any inaccuracies in the application for the degree thesis or materials, I shall bear all related responsibilities.

Signature: Date:

北京交通大学毕业设计（论文）中文摘要

Chinese Abstract

Abstract:Large airport hubs serve as convergence points for multiple transportation modes,connecting air transport with various landside transportation options.Addressing the complexity and dynamics of multi-modal landside transportation demand forecasting at airports, this paper proposes a deep learningcombinedmodel (CNN-LSTM), aiming to achieve relatively accurate predictions based on multi-source dataand make forecasts in a shorter time tooptimize traffic resource allocation. Taking Beijing Daxing International Airport as a case study, the research integrates multi-sourcepublictransportation data such as flight arrival schedules, rail transit, taxis, and buses to construct an integrated forecasting framework. The model employs Convolutional Neural Networks (CNN) to extract local spatiotemporal features from multi-source data and combines Long Short-Term Memory networks (LSTM) to capture long-term dependencies, achieving collaborative demand predictions for three transfer modes: taxis, subways, and buses. Experimental results show that theCNN-LSTMmodel achievesR² values of0.75,0.91, and0.94for bus, taxi,and subway demand predictions, respectively, with overall performance surpassing baseline models such asARIMAandGRU. Additionally, the model significantly improves training efficiency, with a training time of only0.9minutes, enabling faster responses compared to traditional methods and facilitating managerial decision-making. The study validates the effectiveness of multi-source data fusion and deep learning techniques in airport transportation demand forecasting, providing theoretical support and practical references for enhancing hub operational efficiency and passenger experience.Figure18 pieces, 6 tables, X references.

Keywords: Multi-source data, multimodal transportation, airport landside traffic, CNN-LSTM

北京交通大学毕业设计（论文）英文摘要

ABSTRACT

ABSTRACT:Large-scale airport hubs serve as critical junctions integrating multiple transportation modes, connecting air transport with various landside systems. To address the complexity and dynamic nature of multimodal transport demand prediction on airport landside, this study proposes a deep learning hybrid model (CNN-LSTM) based on multi-source data, aiming to achieve accurate predictions and optimize resource allocation with rapid response. Taking Beijing Daxing International Airport as a case study, the research integrates multi-source public transport data, including flight arrival records, rail transit, taxi, and bus operations, to construct a unified forecasting framework. The CNN component extracts local spatiotemporal features from heterogeneous data, while the LSTM captures long-term dependencies, enabling collaborative demand prediction for taxi, subway, and bus services. Experimental results demonstrate that the CNN-LSTM model achieves R² scores of 0.75, 0.91, and 0.94 for bus, taxi, and subway demand predictions, respectively, outperforming baseline models such as ARIMA and GRU. Additionally, the model significantly enhances training efficiency, requiring only 0.9 minutes for training, thereby enabling real-time decision-making for management. This study validates the effectiveness of multi-source data fusion and deep learning in airport transport demand forecasting, providing theoretical and practical insights for improving operational efficiency and passenger experience. The paper includes 18 figures, 6 tables, and references.

KEYWORDS： multi-source data, multimodal transport, airport landside traffic, CNN-LSTM

北京交通大学毕业设计（论文）目录

Table of Contents

1.1Research Background1

1.2Research Significance1

1.3Literature Review1

1.3.1Single-Mode Prediction Models1

1.3.2Multi-Mode Prediction Models3

1.4Research Content4

2Analysis of Current Airport Passenger Transfer Situation5

2.1Analysis of Arrival Flight Data5

2.2Analysis of Multimodal Departure Data5

3Methodology7

3.1Problem Definition7

3.2System Framework7

3.3Feature Extraction Module7

3.4Encoder Module8

3.5Attention Layer8

3.6Decoder Module8

4Model Validation Experiments and Results9

4.1Data Description9

4.1.1Data Preprocessing9

4.2Experimental Configuration9

4.3Evaluation Metrics9

4.4Baseline Models and Model Parameters10

4.5Experimental Results Analysis10

4.5.1Baseline Model Comparison10

4.5.2Ablation Study10

4.5.3Model Performance Evaluation10

4.5.4Prediction Results Visualization10

Table of Contents Notes:

(Small 4, bold, left-aligned, 1 character space)

(Small 4, Song typeface, first line indented by 1 character)

(Small 4, Song typeface, first line indented by 2 characters)

北京交通大学毕业设计（论文）正文

Introduction

Research Background

Large airport hubs serve as convergence points for multiple transportation modes, such as aviation, high-speed rail/urban rail transit, taxis, ride-hailing services, buses, and private cars. These modes collectively form a complex multimodal transportation network, which not only meets diverse passenger travel needs but also poses significant operational management challenges. Passenger travel demands at airport hubs often involve transfers across different transportation modes and are influenced by flight dynamics, operational environments, and other factors, resulting in highly dynamic and complex traffic demand patterns.

With the continuous growth of passenger flow at airport hubs, accurately predicting multimodal transportation demand to enhance operational efficiency and passenger experience has become a key research focus in transportation management and planning. However, traditional prediction methods exhibit clear limitations in addressing complex cross-modal demand, struggling to adapt to the dynamic and variable characteristics of airport hub traffic. Therefore, constructing an integrated traffic demand prediction model based on multi-source data can more comprehensively capture the dynamic interconnections among different transportation modes, which holds significant importance for improving airport hub traffic management capabilities.

Research Significance

From a theoretical perspective, research on integrated multimodal transportation demand prediction at airport hubs will advance the application of multi-source data fusion technologies and deep learning models in complex transportation systems, deepen the understanding of cross-modal transportation correlations, and further enrich the methodological framework of intelligent transportation systems research.

From a practical perspective, accuratetransfertraffic demand forecasting can help managers allocate transportation resources efficiently, optimize intermodal connectivity, alleviate congestion, and enhance passenger transfer experiences. Moreover, this research can serve as a reference for other types of transportation hubs, providing theoretical support and practical guidance for optimizing and developing urban integrated transportation systems.

Literature Review

This sectionwill explore the research fromairport transfer passenger flowandtrafficforecasting models, particularly multimodal forecasting models, from these two perspectives..

Airport Transfer Passenger Flow

Early research on airport transfer passenger flows was constrained by limitations in real-time data collection technology and distribution management conditions, leading scholars both domestically and internationally to primarily rely on operations research theory and simulation modeling for exploration. Overseas research took the lead due to earlier airport construction.Monterier and Hansen[3] used San Francisco Airport as a case study to systematically analyze the impact of landside transportation systems on overall airport traffic. They proposed improving operational efficiency by optimizing traffic organization and adding transfer facilities, providing a practical template for subsequent research.Gorstein and Mccabe[8] further defined the boundaries of landside transportation systems, using simulation software to model dynamic interactions between pedestrians and vehicles. By quantifying simulation results, they derived a traffic capacity calculation model, offering key parameter support for airport planning.Stephane's team[17] broke through the single-entity perspective by integrating decisions from airports, airlines, and passengers into a unified framework. Based on survey data from the San Francisco Bay Area, they constructed a two-level nested Logit model, revealing the complex relationships between fare strategies, flight schedules, and passenger choices.Gulsah[18] focused on micro-level individual behavior, conducting field research at Columbus International Airport and using a binomial Logit model to quantify passengers' acceptance thresholds for private car alternatives (e.g., shared transportation). The study found that travel cost and time reliability were core decision-making factors.

Domestic research was constrained by the lag in airport infrastructure construction and only gradually emergedin the early21st century, but developed rapidly. Yao Yanbin[19]pioneeringly combined the Analytic Hierarchy Process (AHP) with multipleLogitmodels in2006, using the Capital Airport as a case study, to quantitatively evaluate for the first time the diversion effect of rail transit access on the landside transfer system, demonstrating that increased rail分担率 could alleviate road congestion during peak hours. Wen Yifan[20]further refined the model structure by differentiating the transfer preferences of business and leisure travelers through a hierarchicalLogitmodel, predicting market shares of different transportation modes, and constructing a demand elasticity model to support dynamic pricing strategies. Guo Zhenyi[21]targeted the diversified transfer scenarios at Baiyun Airport, improved the traditional nestedLogitmodel by introducing variables such as terminal layout and future expansion plans, and verified the model's adaptability in long-term planning. The Bao Danwen team[22]based on fused stated preference (SP) and revealed preference (RP) data, compared the performance differences between mixedLogitand nestedLogitmodels, finding that mixedLogithad greater advantages in capturing individual heterogeneity. Zhang Lanfang et al.[42]used theNLmodel to reveal that business travelers are highly sensitive to time while non-business travelers are more sensitive to costs, proposing recommendations for differentiated transfer service designs.

With the maturity of intelligent sensor deployment at airports and real-time passenger flow statistics technology, research focus has gradually shifted to data-driven prediction models. Li Xinyue[43] focused on the real-time decision-making mechanisms of departing passengers, analyzing the influence of multi-source information such as display screen guidance and mobileAPP prompts on choice behavior. However, Hu Xiaobo[13] and Sun Zhiqiang[14]'s research on Capital Airport remained limited to local transportation connection issues (e.g., taxi dispatch optimization), lacking in-depth exploration of the overall synergistic effects of the transfer system. Xia Wei[16] summarized general patterns of traffic flow organization but did not propose quantitative evaluation tools.

Single-mode traffic prediction

Data-driven models have achieved mature research in other transportation modes, especially road traffic. Single-mode traffic prediction has evolved from statistical methods to combinations of various machine learning approaches over years of development, with many established theories.

Early single-mode predictions relied on traditional statistical methods, such asARIMA (Chen et al. 2020), which smooths short-term fluctuations through time series, logistic regression (Apronti et al. 2016), and Bayesian networks (Zhu et al. 2016), which fit probability distributions using historical data. However, their linear assumptions struggle to capture the dynamic nonlinear relationships between passenger flow and external factors like weather or events. The historical average algorithm (Smith and Demetsky 1997), while simple and practical, fails to adapt to sudden flow changes such as holidays. Kalman filtering (Kumar 2017) can update predictions through state equations, but its high sensitivity to model parameters results in insufficient robustness.

The introduction of machine learning models partially addresses the aforementioned limitations.KNNalgorithm(Xu et al. 2020; Tak et al. 2014)achieves non-parametric prediction through similar-day matching,but faces the "curse of dimensionality" in high-dimensional data;neural networks(Wei and Chen 2012; Zheng et al. 2006)capture nonlinear features through multilayer perceptrons, but shallow structures have limited ability to model spatiotemporal correlations;SVM(Tang et al. 2019; Jiang et al. 2014)maps high-dimensional spaces using kernel functions, but relies heavily on hyperparameter selection, and computational complexity increases sharply with data volume. Although these models outperform traditional methods in specific scenarios, they still lack the ability to finely characterize the implicit spatiotemporal heterogeneity in traffic data (e.g., differences in passenger flow patterns between morning subway peaks and evening bus peaks).

The breakthrough progress in deep learning has propelled prediction accuracy to new heights. Recurrent neural networks(Ma et al. 2015) and their variantsLSTM(Liu et al. 2019; Yang et al. 2021; Zhang et al. 2019) capture the temporal dependencies of passenger flow through gating mechanisms, but their unidirectional propagation structures struggle to model long-term (e.g., weekly, monthly) patterns;CNN utilizes convolutional kernels to extract local spatial features, excelling in grid-basedOD matrix predictions, yet forcibly rasterizing transportation networks leads to the loss of topological information (e.g., road connectivity). Graph neural networks (GNN) model the native transportation topology through node-edge relationships, emerging as a new paradigm for spatial feature extraction:Guo et al. (2019) designed a spatiotemporal graph neural network (STGNN), integrating graph convolution (GCN) withGRU, achieving the first dynamic prediction of city-wide bicycle demand;Chen et al. (2020) proposed a multi-taskGCN framework, simultaneously predicting taxi demand and vacancy rates, enhancing fine-grained prediction accuracy through road cascade relationship modeling;Geng et al. (2019) constructed a multi-graph convolutional network, encoding inter-regional distance, traffic flow similarity, and functional complementarity, significantly optimizing ride-hailing dispatch efficiency.However, single models often face performance bottlenecks, such asGCN's over-smoothing issue when layers deepen, and LSTM's delayed response to abrupt patterns.

To overcome the limitations of single models, hybrid architectures have become a research hotspot.Ke et al. (2017) combined LSTM with CNN in parallel to separately extract temporal trends and spatial hotspots of ride-hailing demand, improving short-term prediction robustness through feature concatenation; Zhang et al. (2020) designed a GCN-3DCNN hybrid model to capture spatial correlations between subway stations while using 3D convolution to mine spatiotemporal cube features of passenger flow; Zhang et al. (2021) proposed Res-LSTM, which introduces residual connections to alleviate gradient vanishing and combines GCN to model station topology, reducing subway passenger flow prediction errors by 12%.The rise of attention mechanisms has further enhanced model interpretability:Vaswaniet al. (2017)proposed theTransformer, which captures global temporal dependencies through self-attention.Liet al. (2019)'sLogSparse Transformeradopted an exponentially decaying sparsity strategy to reduce computational overhead;Zhouet al. (2020)'sInformerachieved efficient modeling of long sequences through probabilistic sparse attention mechanisms, improving subway passenger flow prediction accuracy by23%compared to traditionalLSTM;Yaoet al. (2019)designed the Spatio-Temporal Dynamic Network (STDN), which filters noise using local flow gating and weights historical periodic data through attention mechanisms, effectively addressing the challenge of predicting sudden holiday passenger flow fluctuations.

The aforementioned models all focus on single-mode transportation prediction. However,in the real world, multiple transportation modes form a dynamic interconnected network due to passenger behavior choices and system synergies.Therefore,jointly considering multiple transportation modes rather than ignoring their correlations becomes particularly important.

Multimodal prediction models

In the real world, multiple transportation modes form a dynamically interconnected network due to passenger behavior choices and system synergies.Irawan et al. (2019) empirically found significant complementary and competitive relationships between motorcycle ride-hailing services, motorcycle taxis, and public transport. For instance, motorcycle taxis provide "last-mile" connectivity for subways during peak hours, while directly competing with buses during low-demand periods. With the integration of multi-task learning and deep learning techniques, joint multi-modal prediction is gradually becoming feasible.Ke et al. (2021) proposed a multi-task multi-graph learning method, constructing independent graph structures for taxis, shared bikes, and buses. They utilized multi-graph convolution to extract spatiotemporal features of each mode and designed an inter-task knowledge-sharing mechanism to predict ride-hailing demand. However, such methods do not explicitly model cross-modal influences (e.g., subway delays causing surges in taxi demand), relying instead on implicit parameter transfer for correlation, resulting in weaker interpretability.

Early multi-modal prediction studies had significant limitations.Zhong et al. (2017) simply aggregated subway, bus, and taxi passenger flows into regional total passenger volume for prediction, ignoring nonlinear interactions between modes (e.g., demand shifts caused by fare differences).Ye et al. (2019) used a CNN-LSTM hybrid model to simultaneously predict taxi and shared bike demand, but their spatial feature decomposition was based on static grid partitioning, failing to reflect cross-modal flow transmission under real road network topologies.Xu et al. (2022) attempted to introduce multi-spatial correlation graphs to depict competition between bikes and taxis, combining graph attention networks to quantify inter-modal demand substitution elasticity. However, reliance on manually defined correlation matrices made it difficult to adapt to dynamic scenarios (e.g., sudden weather changes causing sharp drops in cycling demand).Liang et al. (2022) proposed a Multi-Relation Graph Neural Network (MR-GNN), using fixed mode correlation matrices to model subway-bus synergy. Yet, static matrices could not capture emergency coordination demands during sudden disruptions, leading to prediction delays.

To break through static correlation constraints, some studies have shifted to adaptive graph modeling.Lu et al. (2020) introduced an adaptive adjacency matrix in road speed prediction, dynamically adjusting the weights of different traffic modes (e.g., private cars and trucks) through a gating mechanism, but it was not extended to multimodal passenger flow prediction scenarios.Huang et al. (2022) designed a dynamic residualGCN, using an attention mechanism to adaptively generate inter-mode correlation matrices, and validated the effectiveness of dynamic modeling in intercity traffic flow prediction;Bai et al. (2020) proposed an adaptive graph convolutional recurrent network, simultaneously optimizing node features and graph structure parameters, but it focused on intra-mode correlations and did not address cross-modal interactions. Recently, the multimodal former (M2-former)[2] addressed the challenge of heterogeneity in multimodal transportation by proposing an encoder-decoder architecture: the encoder separately models the spatiotemporal features of subways, buses, and shared bikes through adaptive multi-graph convolution and captures dynamic interactions using a cross-mode attention mechanism; the decoder then transfers features from high-density traffic modes (e.g., subways) to low-density modes (e.g., customized buses) through knowledge distillation, significantly alleviating data sparsity issues. This model validated the ripple effect of subway flow restrictions on sharedbike demand during holidays in a Beijing multimodal dataset, but its applicability to small-scale hub scenarios (e.g., airport landside traffic) has yet to be verified.

Knowledge transfer and adaptation to sparse data have become another research focus.Li et al. (2021) utilized memory neural networks to transfer dense passenger flow features from subway stations to sparse intercity bus station predictions, but relied solely on time-series similarity without considering spatial topological constraints;Li et al. (2022) further proposed an unsupervised knowledge adaptation model, aligning feature distributions of different transportation modes through adversarial training, but theirLSTM-based backbone network struggled to capture complex spatial dependencies. Notably, the Locality-Perception-Enhanced Spatiotemporal Graph Transformer Network (LPE-STGTN)[1] dynamically captures cross-temporal shared patterns (e.g., commuting cycles) and transient interactions (e.g., emergencies) between regions through a spatiotemporal graph generator, combined with a lightweightAFT-local module to enhance computational efficiency, offering a new approach for multimodal correlation modeling. However, it remains limited to unimodal predictions for taxis and ride-hailing services.

Multimodal prediction models primarily focus on spatiotemporal analysis at urban scales, with limited development for small-scale areas such as transportation hubs.

Summary of Research Status

Based on the above background, existing research on short-term traffic flow prediction for multimodal transportation faces the following challenges: (1) the difficulty of learning interaction mechanisms among multiple transportation modes; (2) the challenge of extracting complex dynamic spatiotemporal features; (3) the organizational difficulty of heterogeneous data structures.

Research Content and Technical Approach

Conducting research on integrated prediction of multi-modal landside transportation demand at airports driven by multi-source data aims to develop deep learning models capable of dynamically capturing complex intermodal correlations and spatiotemporal features. Through empirical analysis of real-world data from airports like Beijing Daxing International Airport, this research can not only validate the feasibility and effectiveness of the proposed methods but also provide theoretical support and practical references for demand prediction in other airport hubs and comprehensive transportation hubs.

北京交通大学毕业设计（论文）正文

Fundamental Theoretical Research on Airport Transfer Passenger Flow Prediction

As a core node of comprehensive transportation hubs, accurate prediction of transfer passenger flow at airports is of great significance for improving operational efficiency and optimizing resource allocation. This chapter systematically explores the aggregation characteristics, transfer mechanisms, and influencing factors of passenger flow in airport landside transportation systems, and conducts comparative analysis based on traditional statistical and machine learning methods, aiming to build a predictive model framework adaptable to complex scenarios.

Basic Theories

Definition of Airport Landside Transportation Concept

Airports, as critical nodes connecting air and ground transportation, can be functionally divided into two main parts: airside and landside. The airside encompasses the core operational areas for aircraft, including terminal approach airspace for takeoffs and landings, as well as movement areas for taxiing and parking, ensuring efficient aircraft operations.

Figure 2-1 Airport Functional Zones[22]

The landside area mainly consists of terminal buildings, ancillary facilities, and ground transportation networks connecting the airport. As the core functional area of the landside, terminal buildings handle processes such as check-in, security screening, and baggage handling, while ground transportation facilities organically connect the airport with urban transportation networks through diversified means such as roads, rails, and buses, as shown in Figure 2-1.

The focus of this study is the landside transportation transfer system at airports, which refers to the facilities and behavioral processes required for passengers to switch between different transportation modes within the airport hub.

Transfer process

Air passengers can be categorized into three types: transfer passengers, arriving (inbound flight) passengers, and departing (outbound flight) passengers..

Departing passengers arrive at the terminal curb or rail station via ground transportation (e.g., buses, taxis, rail transit, etc.), then enter the terminal to complete check-in and security procedures before boarding. Among them, rail transit passengers need to walk from the station to the terminal; if they choose other modes of transportation to the airport, all passengers disembark at the terminal curb before entering the terminal.

The evacuation process for arriving passengers is more complex: departing passengers who choose taxis transfer between the terminal and the terminal curb; those who opt for rail transit or private vehicles transfer between the rail station, public parking lots, and the terminal, involving large passenger volumes and complex walking routes; passengers taking buses or shuttles depend on the airport's specific conditions.

The transfer routes for connecting passengers are relatively independent, typically separating from arriving passengers after disembarking and quickly connecting to subsequent flight boarding gates via the terminal's automated people mover system or pedestrian walkways.

This study primarily focuses on the evacuation process of arriving passengers,After disembarking, domestic and international passengers follow different routes. Domestic passengers can directly collect their baggage and proceed to the arrival hall to choose transportation options for leaving the airport. International passengers, upon arrival, must undergo quarantine inspection, immigration clearance, baggage collection, customs procedures, and baggage quarantine before reaching the arrival hall to select transportation options. Connecting passengers, however, are separated from arriving passengers immediately after disembarking and proceed to boarding gates or directly board their next flight.

Figure2-2Transfer process for arriving passengers in the terminal[22]

Flight arrivals andairport passenger flowaggregationcharacteristics

The aggregation characteristics of airport passenger flow are closely related to the spatiotemporal distribution of flight arrivals. Flights typically follow a "peak-valley" arrival pattern, where concentrated arrival periods form significant passenger flow peaks (e.g., morning and evening rush hours), while other periods show relatively stable flows. This fluctuation is determined by airlines' route scheduling strategies, differentiated timetables for international and domestic flights, and transfer connection demands at hub airports. Temporally, domestic flights often concentrate during6:00-9:00and18:00-21:00—preferred travel time slots—while international flights may exhibit multi-peak distributions due to time differences and route network designs. Additionally, actual flight arrival times are affected by weather,operational conditions,and air traffic control, potentially causing sudden surges of passengers in terminals and intensifying instantaneous crowding pressures. Spatially, international and domestic passengers are diverted to different areas due to process differences, creating localized hotspots. For example, international arrival halls may experience baggage claim congestion during peak arrival times. The intermittent nature of flight arrivals leads to "pulse-like" passenger flow patterns, requiring dynamic monitoring and predictive models to capture spatiotemporal evolution patterns and provide scientific basis for resource allocation in passenger dispersal.

Transfer transportation organization characteristics

Different transfer modes exhibit significant differences in transportation organization, directly impacting passenger evacuation efficiency and overall airport service levels. Rail transit, such as subways and airport express lines, features high capacity and fixed schedules, making it suitable for large-scale passenger evacuation but requiring coordination with flight schedules to match demand. Their stations are typically located at a distance from terminal buildings, necessitating optimized pedestrian experiences through facilities like corridors and moving walkways. Additionally, to address tidal demand patterns, service frequency should be increased during peak hours and appropriately reduced during off-peak periods to lower costs. Taxis and ride-hailing services, with their flexibility and immediacy, are the preferred choice for business travelers with high time sensitivity. However, their efficiency is constrained by curb capacity and road congestion, requiring intelligent dispatch systems to balance supply and demand and avoid prolonged passenger waits or curb congestion. Buses and airport shuttles offer extensive route coverage, connecting urban areas and satellite regions, but their lower frequency necessitates dynamic alignment with flight peak times and optimized boarding/alighting zone designs to alleviate vehicle queuing and passenger congestion during peak periods. Private cars and taxis (self-provided vehicles) have evacuation efficiency closely tied to parking lot turnover rates. Dynamic parking guidance systems can reduce search time, but the process of passengers walking or taking shuttle buses from parking lots to terminals is time-consuming, potentially affecting overall experience.

Arriving passengers' choice of evacuation modes

The choice of transfer modes by arriving passengers is influenced by multiple interacting factors, with complexity reflected in dimensions such as transportation mode attributes, individual passenger characteristics, external environment, and sociocultural aspects.

The accessibility of transportation modes is the primary consideration, such as rail transit station coverage,private vehicles enabling "door-to-door" servicedirectly determining passenger preferences;theeconomic cost of transportation modes includes direct expenses(such asticket fares, parking fees), as well ashidden costs(includingtime value, physical exertion);while time reliability,such as the risk of rail service delays,is particularly critical for business travelers' decisions.

Regarding individual passenger characteristics, travel purpose significantly influences choice preferences: business travelers tend to favor fast, comfortable car services, while leisure travelers prioritize cost and connectivity to attractions; passengers with large luggage may avoid modes requiring multiple transfers;first-time arrivals rely on the clarity of signage systems and prefer straightforwarddirectoptions.

In the external environment, adverse weather conditions may reduce the willingness to walk, prompting passengers to choose more direct transportation options like taxis, while temporary traffic controls or large-scale events may also force travelers to adjust their original plans.

Socio-cultural factors should not be overlooked either, such as environmental awareness prompting some passengers to choose low-carbon options like shared buses, whilegroup behaviors like family tripsmay lead to collective decisions overriding individual preferences.

These factors do not exist in isolation; for example, adverse weather may amplify the trade-off between economic and time costs, while travel purpose further modulates sensitivity. Therefore,using discrete choicemodelsor traditional regression models cannotdynamically capture the interaction mechanisms among these factors,improvingpredictionaccuracy.

Methodology Comparison

The selection of airport transfer passenger flow prediction methods requires comprehensive consideration of data characteristics, prediction objectives, and scenario complexity. Current mainstream methods include traditional statistical models, deep learning approaches, and emerging technologies, with core differences reflected in model assumptions, feature extraction capabilities, and dynamic adaptability. The following analysis explores their principles, applicable scenarios, and limitations, supported by research cases to evaluate their strengths and weaknesses.

Traditionalstatistical methods

Traditional statistical methods are represented by linear regression and time series models, such as ARIMA, SARIMA, which predict passenger flow by establishing explicit mathematical relationships. Linear regression fits the linear relationship between passenger flow and variables like flight schedules and weather by minimizing the sum of squared residuals. The model structure is simple, and the parameters are highly interpretable, making it suitable for scenarios with clear variable relationships and limited data volume. For example, the ARIMA model can capture data trends and periodicity, but it requires the time series to meet the stationarity assumption. However, traditional methods have limited expressive power for nonlinear relationships (such as the dynamic coupling between flight delays and traffic congestion) and high-dimensional interaction features. When passenger flow is affected by multiple overlapping factors (such as sudden weather changes or holiday effects), the model's accuracy tends to decline due to the failure of linear assumptions. Additionally, traditional statistical methods rely on the stability of historical data and struggle to adapt to unexpected events in airport operations (such as temporary traffic control), resulting in weaker robustness.

Machine learning methods

Shallow machine learning methods automatically learn data patterns through algorithms, breaking the linear limitations of traditional models. Decision trees and random forests handle nonlinear relationships through feature splitting and ensemble strategies. For example, random forests can quantify the contribution of flight density and transfer facility capacity to passenger flow, making them suitable for feature importance analysis. Support Vector Machines (SVM) map high-dimensional spaces using kernel functions and perform robustly in small-sample scenarios, having been used to classify and predict passenger transfer mode choices. Gradient Boosting Machines (GBM) iteratively optimize residuals to improve prediction accuracy and excel in structured data. However, such methods have limited capability in extracting spatiotemporal features. For instance, random forests struggle to capture long-term dependencies in passenger flow time series, such as lag effects caused by flight peaks; while SVM experiences significantly increased computational complexity when processing large-scale spatiotemporal data. Moreover, shallow models rely on manual feature engineering and have limited ability to integrate multi-source heterogeneous data.

Deep learning methods automatically extract high-order features of data through multi-layer neural networks, demonstrating outstanding performance in complex scenarios. Long Short-Term Memory networks (LSTM) are specifically designed for time-series data, effectively modeling the lagged correlation between flight arrival peaks and passenger flow surges. Convolutional Neural Networks (CNN) excel at capturing spatial features, such as the distribution of passenger density in different areas of the terminal. Graph Neural Networks (GNN) model passenger flow interactions between multiple transportation hubs through node and edge topological relationships, making them suitable for collaborative forecasting in large airport clusters. Another advantage of deep learning lies in its ability to integrate multi-source data, such as embedding real-time flight status into model inputs to dynamically adjust predictions in response to extreme weather impacts. However, deep learningdemands high computational resources, requires large-scale annotated data for training, and has poorer interpretabilitycompared to traditional statistical methods.

Airport transfer passenger flow prediction must balance spatiotemporal dynamics, multi-factor coupling, and real-time requirements. Traditional methods remain valuable when data is sparse or relationships are explicit but struggle with nonlinearity and high dimensionality. While shallow machine learning partially overcomes linear limitations, its ability to represent spatiotemporal features is limited. In contrast, deep learning methods can automatically extract implicit relationships between passenger flow and factors such as flights, transportation, and the environment, making them particularly suitable for handling unstructured data.

Chapter Summary

This chapter first defines the concept of landside transportation at airports, elaborates on the specific transfer processes, and discusses the "peak and trough" characteristics of flight arrivals, the organizational efficiency of transfer transportation modes, and individual passenger preferences, which togetherconstitute the challenges of passenger flow forecasting.Secondly, in the method comparison section, it systematically contrasts the applicability of traditional regression methods, shallow machine learning, and deep learning.Deep learning excels in multi-source data integration and real-time prediction adjustments, particularly suited for high-noise, unstructured data scenarios in airport transfer passenger flow. In summary, given the spatiotemporal dynamics, multi-factor coupling, and real-time requirements of airport passenger flow forecasting, deep learning emerges as the optimal choice to address the complex demands of airport transfer systems.

Case Analysis of Arrival Passenger Transfer at Airports

Current Situation Analysis of Beijing Daxing International Airport

DataFoundation

Multi-source data

The data used in the study primarily includes2023 April 1 to 2023 July 5 Beijing Daxing International Airport's arrival flight data, rail transit, taxis, airport shuttles, private cars (including ride-hailing services) and other landside multi-modal transportation operation data

(1) Arrival flight data

The data used in the studyArriving flightsData mainly comes fromthe airport management system,以xlsxstored in format,arrival and departure flight data are counted separately, with passenger numbers recorded byboarding gate check-in information,relatively high reliability.As shown3-1shownarrivingflightDataMainly includesFlight date、Flight number、Airline、Aircraft Type、Aircraft Number、Parking Stand、Origin station、Previous station、Scheduled landingTime,Actual landingTime,Passenger count等Information。All time information is accurate to the minute.自2023年4月1to6月30,totaling37170piece of data, the data integrity is good, no anomalies or missing values。

Figure3-1 Arrival flight data example

(2)Departuremultimodaltransportationdata

This dataprimarily comes fromcameras or sensors at the entrances of different transportation modes,includingtime,direction (1 Arrival 2 Departure)and hourlycumulativepassenger counts, all stored inCSVformat.Taking the subway as an example,as shown inFigure3-2.Ifthere is no passenger flow during a certainperiod, that period will not appear in the records.2023Year4Month1Day to2023Year7Month5Day, including subway2385records, taxi4749records,interprovincialbus total3607records, city bus4605records, private car4769records.The data completeness varies significantly across different transportation modes:taxi,private car, and subway data have good completeness,with no anomalies/missing values;city bus is missing approximately0.2%of directional data, while interprovincialbusis missing about13.5%of directional data, requiring data preprocessing.

The figure also shows4月1日0there are three data points, which have been verified,The smallest line (26) represents the passenger flow data for that time period, the largest line (209622) is the sum of all data for the month, while the remaining line between them represents the total data for the day.Other dates at midnight also have data representing the daily total, and the first day of each month also includes data for the monthly total.This part will be treated as an outlier during data preprocessing.

Figure3-2Example of Departure Traffic Data

Data Preprocessing

(1) Arrival Flight Data

WritePythoncode toautomatically parse the originalExcelfile and standardize column names,converttimedata intodatetimetype,and remove all values less than zero from the passenger count field.Then,use thefloor('h')function to align arrival times to full hour marks, thereby unifying the time granularity to hourly.After timestamp alignment,aggregate the number oflanding gearsandtotal passenger volume per corresponding time period by hour, filling missing time slots with zeros.The processed data is then stored inCSVformat.

In addition to routine passenger flow data cleaning, the study also conducted supplementary processing on flight delay information. By calculating the difference between actual and scheduled landing times, the "delay time" metric was constructed, and negative values for early arrivals were uniformly truncated to zero.According toCivil Aviation对delayrelateddefinition，set the actual landing time difference from the planned landing time to be greater than15the situation in minutes isdelay.Then proceed with 30 Minute window resampling, based on threshold conditions (delay proportion exceeding 50%, average delay surpassing 15 minutes with at least5flights) to identify potential periods of large-scale delays.If there areAdjacent or overlapping time periods, thenmerge and calculate the duration。

Figure3-3Delay Period Report

(2) Departure multimodal transportation data

First, separate all data by arrival/departure direction. For data missing direction information, first determine whether both arrival and departure directions have data for that time period. If one direction lacks data for that period, fill it accordingly. If both have data, verify using the total daily data mentioned in section 3.2.1. If the sum of data from other time periods matches the recorded total, delete the data for that period.

After filling in the directions, convert all "time" fields to datetime format. To compensate for randomly missing hourly data in the original observations, first construct a continuous hourly sequence covering the entire period based on the earliest and latest timestamps in the data sample. Then, by comparing the expected record count with the actual count, determine whether certain hourly periods require filling. The method used for flight data processing was not adopted here because some transportation modes have non-operating hours (e.g., midnight), during which no data would appear. However, since the daily total data includes these periods, simply skipping them would incorrectly treat the daily total as hourly data, affecting result accuracy.

Further verify data completeness by conducting a dedicated quantity check for the dual-record rule at midnight 0 o'clock. If the observed count does not match, generate a detailed exception report specifying the date, anomaly type, and discrepancy value.

After completing missing value repair and validation, retain the minimum value of the "number of people" field by timestamp, and finally store the processed results in a new file prefixed with processed_.

Through the aboveMultiplePreprocessingWith validation operations,effectively improving the quality and usability of airport transfer passenger flow data to meet the input requirements of subsequent time series analysis and demand prediction models.

Figure 2-4 Example of a multi-modal transportation data validation report

Arrival passenger flow distribution pattern

对2023年4月-6Hourly aggregation of monthly arrival flight data,As shown2-1in the figure, flight passenger flow exhibits obvious periodicity. Influenced by fixed daily flight schedules and people's travel habits, airport arrival passenger flow shows strong daily periodicity.。

图2-1 Flight passenger flow variation over time chart

Figure2-2Daily aggregated flight passenger flow variation over time chart

Figure 2-2 aggregates the inbound flight passenger flow by day, making it more evident that peaks occur around the "May Day" holiday and at the beginning of the summer vacation.

The number of passengers reaches its peak during these periods.

Figure2-3Flight delay period annotation

Figure2-3illustrateslarge-scale delay periods,with relevant datalaying the data foundationfor validating the model's effectiveness under delay conditions in subsequent sections.

Departingpassenger transferdistribution pattern

TransferMethods such as city buses, subways, and taxispassenger flowrespectivelyAs shown2-4、图2-5and the figure2-6As shown, it can be seen that the passenger flow of all three modes exhibits significant fluctuation patterns. In terms of total volume, the passenger flows transferring to the subway and taxis are similar, while the passenger flow for city buses is significantly lower than the former two.

图2-4 Passenger flow variation over time for transfer buses

Figure2-5Time-varying diagram of subway transfer passenger flow

Figure2-6Time-varying diagram of taxi transfer passenger flow

According to the capacity limits of the subway and parking pool set by Daxing Airport

, the current usage situationstill has considerable capacity space before reaching saturation.

Analysis of inter-modal transportation correlations

Calculate the hourly average for all transportation modes and plot a line chart as shown in the figure2-7The figure above shows flights, and the figure below shows three transfer transportation modes. Further analysis reveals the daily periodicity of flights, with passenger flow gradually increasing from 6 a.m. until reaching its peak around 10 a.m., followed by fluctuating passenger numbers until1 p.m. in the afternoonwhen passenger flowReached its peak,after which passenger flowfluctuateddecreased until 3 a.m. the next day, 4 a.m.-Occasionally, there is passenger flow at 5 a.m.。

Figure2-7Hourly distribution comparison of transportation modes

图2-7The figure below shows the hourly distribution comparison of three transfer modes. From midnight to 5 a.m., due to subway closure, taxis have the highest passenger flow, peaking at 5 a.m. After subway operations resume, passenger flow surges sharply, surpassing taxi flow around 6:50 a.m. and reaching its peak at 7 a.m. Subsequently, both modes show fluctuating declines. Taxis drop to their lowest point at 2 a.m. the next day, while subway passenger flow remains until service ends.

Figure2-8Lag Correlation

As shown in2-8, the immediate correlation between different transfer modes is displayed. Based on the correlationresults, the following conclusions can be drawn:Taxi-subwaycorrelation is0.649,indicating a strong positive correlation, suggesting thattheir passenger flows change synchronously and may share passenger sources.Taxi-buscorrelation is0.548, with a moderate positive correlation,possibly due to an overall increase in transportation demand (e.g., during peak hours).Subway-buscorrelation is0.567, with a moderate positive correlation, indicating that the two mayrouteshave certaincomplementarity, sharing someregionaltransportationdemands.

北京交通大学毕业设计（论文）正文

Methodology

Problem Definition

Set the hourly observation vector

$x_{} =[F_{}, T_{}, M_{}, B_{}]^{} ∈ R^{}$

Where $F_{}$ is the number of arriving flight passengers (persons), $T_{}$ , $M_{}$ , $B_{}$ are the transfer volumes of taxis, subways, and buses, respectively. Given the historical window length $w$ and prediction step length $H$ , the goal is to predict after time $t$ .

$_{} =[_{},_{},_{}]$

such that

$F:{x_{},…, x_{}}⟶_{}$

Multi-step multimodaltransfertransportationdemand prediction involves learning this mapping function $F$ .

Modelframework

This paperproposes a method based on theCNN-LSTMframework for predicting hourly taxi, subway, and bus demand using flight arrival data. This framework employs a sliding window for feature extraction, withCNNas the feature extraction layer andLSTMfor time-series modeling to achieve multi-output prediction.

To process heterogeneous information within the same network, this study treats $X_{}$ as a 2D tensor of size $r × c$ , where $r = w =24$ hrepresents the time dimension (one day), and $c =4$ is the feature dimension. This operation enables multi-source sequences to be"imagified", facilitating convolutional kernels to slide along the time axis for extracting cross-variable local patterns. The image tensor first undergoes min–max normalization performed independently per feature, which linearly maps each dimension's values to the[0,1]interval, avoiding gradient interference from unit differences. After normalization, the data is divided chronologically: the first80%for training and the remaining20%for testing, with a sliding window step size of1 h.

CNN-LSTMThe architecture of the model is shown in Figure 4-1.The model consists of two parts: a feature extraction encoder and a temporal prediction decoder. The encoder includes two layers of 1D convolution, each with 64 filters and a kernel width of 1–12, covering the entire feature dimension (i.e., kernel size $1× c$ )—this is equivalent to scanning all variables row by row at the same time, enabling the learning of cross-variable local correlations such as "flight surge—passenger flow response." Each convolutional layer is followed by max pooling with a stride of 1 and ReLU activation to reduce dimensionality, suppress noise, and deepen the network. The flattened convolutional output is repeated 24 times to form $(24, d)$ sequences, which are then fed into the downstream LSTM. This repetition ensures that the LSTM receives feature streams consistent with the original time length.

图4-1 LSTM-MHAModel Architecture

The decoder employs a single-layer LSTM (128 units) to carry long-term memory. Its gating mechanism updates the cell state $c_{}$ through the coordinated action of the forget gate $f_{}$ , input gate $i_{}$ , and output gate $o_{}$ , thereby preserving trend information across days or even weeks.

Convolutional Feature Extraction

The core of the convolutional module is the use of shared-weight filters sliding along the time axis to capture short-term high-frequency patterns. A stride of $s =1$ ensures that all adjacent hours are covered by the convolutional kernel, preventing information leakage. Dilated convolution further expands the receptive field by introducing a dilation rate of $d = 2^{}$ , allowing the network to simultaneously perceive daily and weekly periodic structures without increasing parameters.

Apply 1-D convolution in the time dimension:

$= σ (\sum_{}^{} W_{} x_{} + b), k = kernel size$

The shared-weight convolutional kernel captures "flight peaks → followed by passenger surges" and other short-term local patterns, providing an easily interpretable temporal embedding representation for subsequent LSTM.

LSTM temporal modeling

Based on the important local features extracted by CNN, the convolutional output sequence $z_{}$ is fed into single or multiple layers of LSTM.LSTM focuses on macro-level trends across windows. Its memory cells dynamically regulate old and new information through gating mechanisms: the forget gate suppresses transient noise, the input gate writes key events encoded by convolution, and the output gate determines the current moment's influence on the next prediction. This structure is suitable for handling the "burst-decay-stabilization" sequence evolution of airport passenger flow.The gating mechanism operates as follows:

$\begin{matrix} f_{} & = σ (W_{} [h_{}, z_{}]+ b_{}), \\ i_{} & = σ (W_{} [h_{}, z_{}]+ b_{}), \\ _{} & =tanh (W_{} [h_{}, z_{}]+ b_{}), \\ c_{} & = f_{} ⊙ c_{} + i_{} ⊙_{}, \\ o_{} & = σ (W_{} [h_{}, z_{}]+ b_{}), \\ h_{} & = o_{} ⊙tanh⁡(c_{}), \end{matrix}$

thereby establishinglong-rangetemporaldependenciesat the local feature level.

北京交通大学毕业设计（论文）正文

Model validationexperimentsand results

Experimentaldataset

Adopt

3.2the processed dataset of Daxing Airport.First, loadthe data of four transportation modes and merge them by time,The time span for which all four transportation modes have data records is2023年4月1日0to6月30日23points, and data beyond this time rangeis filtered outSubsequently, merge the four types of data into a single file and perform time range validation.

Figure5-1merged data validation results

According to the formula,normalize the data for flights, buses, subways, and taxis,eliminating units of measurement, while saving the scalerasscaler.saveto enable denormalization for obtaining actual values after prediction.

$\begin{matrix} = \frac{x_{} -}{-} #(1) \end{matrix}$

Standardize the four types of dataThe text description is as shown in the figure5-2，Visualization results are as shown in the figure5-3。

Figure5-2Standardization Results

图5-3 Standardized transportation demandVisualizationComparison

Next, create a time window with a time step length oflook_back = 24, and set the prediction target's relative offset from the input end positionas the variabletarget_offsetto1,indicatingthe use of24hours of historical observation datato predict the next hour.Convert timestamps toISOstring formatfor loading asx-axis coordinates in subsequent visualizations.

Next, perform time series division on the data,前80%as the training set, followed by20%as the test set.The training set consists ofX_train, y_traincomposed of two sets of variables, the former has a dimension of4, representing4types of transportation modes'observed values; the latter dimension is3, representing predictions for3different transfer modes. The division results are shown in Figure5-3.

Figure5-4Dataset Partitioning Results

Finally, the structure of the training set and test set was validated, as shown in the figure5-5shown, which is consistent with expectations.

Figure5-5Dataset Structure Validation Results

Evaluation Metrics

This paper adoptsfourevaluation metrics to assess model performance, namelyCoefficient of Determination(Coefficient of Determination, $R^{}$ ), Mean Absolute Error(Mean Absolute Error, MAE), Root Mean Squared Error(Root Mean Squared Error, RMSE)andSymmetric Mean Absolute Percentage Error(Symmetric Mean Absolute Percentage Error, sMAPE).

$R^{}$ reflects the goodness of fit,RMSE,MAEreflect theoverall magnitude of errors,sMAPEreflects thesymmetry of relative errors.Each metric'scalculation formula is as follows:

1. Coefficient of determination（Coefficient of Determination, $R^{}$ ）

$R^{} =1- \frac{\sum_{}^{} (y_{} -_{})^{}}{\sum_{}^{} (y_{} -)^{}}$

2. Mean Absolute Error（Mean Absolute Error, MAE）

$MAE= \frac{1}{n} \sum_{}^{} | y_{} -_{} |$

3. Root Mean Squared Error（Root Mean Squared Error, RMSE）

$RMSE= \sqrt{\frac{1}{n} \sum_{}^{} (y_{} -_{})^{}}$

4. Symmetric Mean Absolute Percentage Error (Symmetric Mean Absolute Percentage Error, sMAPE)

$sMAPE= \frac{200}{n} \sum_{}^{} \frac{| y_{} -_{} |}{| y_{} |+|_{} |+ ε}$

where $n$ is the number of samples, $y_{}$ is the true value, $_{}$ is the predicted value, is the mean of the true values, $ε$ is a very small constant $10^{}$ (to prevent division by 0).

Baseline models and model parameters

To evaluate the performance of the proposed model, the following three baseline models were selected for comparison:

(1)Statistical method-based ARIMA model

(2)Based ona singlemachine learningmethod, theGRUmodel

(3)Based on an ensemblemachine learning method, theLSTM-Multi-HeadAttention(hereinafter referred to asLSTM-MHA)model

All models areimplemented based on the deep learningPyTorchframework. During training,

This paper'sCNN-LSTMmodel, with

ARIMA model uses Python's

Experimental results analysis

Baseline model comparison

From Table 5-1, it can be seen that this paper's model achieves the best performance in predicting all three indicators for buses.

Table 5-1 Model Performance in Bus Prediction

Model	Bus
Model		$R^{}$	RMSE	MAE	sMAPE
ARIMA	0.75	6.5	5.2	27.3%
GRU	0.74	6.5	4.9	33.0%
LSTM-MHA	0.75	6.5	4.9	32.2%
CNN-LSTM	0.75	6.4	4.8	31.4%

From Table 5-2, it can be seen that although the model did not achieve optimal performance in subway predictions, it secured two suboptimal results. For subway forecasting, the ARIMA model performed relatively well, likely due to the more regular operating hours of subways.

Table 5-2 Model Performance in Subway Predictions

Model	Subway
Model		$R^{}$	RMSE	MAE	sMAPE
ARIMA	0.95	53.1	48.4	41.7%
GRU	0.94	61.4	42.5	35.2%
LSTM-MHA	0.94	62.1	44.4	27.1%
CNN-LSTM	0.94	61.0	45.0	44.3%

From Table5-3, it can be seen that the model achieved one optimal and two sub-optimal results when predicting taxi flow. The best-performing part is theGRUmodel. During themodel performance evaluation phase, these two models will be further compared.

Table5-3Performance of models in taxi flow prediction

Model	Taxi
Model		$R^{}$	RMSE	MAE	sMAPE
ARIMA	0.61	108.8	95.8	39.4%
GRU	0.91	51.7	36.8	23.0%
LSTM-MHA	0.88	59.3	41.8	26.1%
CNN-LSTM	0.91	53.0	39.7	24.2%

From Table5-4,it can be seen that the modelachieved the best performance across all metrics overall.

Table5-4Overall Performance of the Model

Model	Total
Model		$R^{}$	RMSE	MAE	sMAPE
ARIMA	0.95	85.5	76.3	27.1%
GRU	0.96	81.8	60.5	19.3%
LSTM-MHA	0.95	90.3	68.2	16.1%
CNN-LSTM	0.96	79.9	60.2	18.0%

Ablation experiment

By constructing model variants and comparing scenarios with missing components, we validate the effectiveness of each component in the model, specifically removing LSTM, retaining only CNN, and removing CNN, retaining only LSTM in these two cases.

The experimental results are shown in Table 5-5.

Model performance evaluation

We selected the optimal GRU model from the baseline models and conducted experiments with our proposed model on the dataset. The training time comparison results are shown in Table 5-6.

Table5-6Model Performance Comparison

Model	Training time/min	Training epoch
GRU	2.9	139
CNN-LSTM	0.9	108

Visualization of prediction results

From the figure5-6it can be seen that the modelpredicted values and true valuesfitrelatively goodThe demand for subway and taxis nearly overlaps in multiple segments; buses show the most significant fluctuations.The model's prediction during peak bus hoursstill has room for improvement.

图5-6 Visualization of the model prediction results in this paper

Figure5-7Prediction Residual Distribution

Ideally, residuals should exhibit a symmetrical bell-shaped curve, approximating a normal distribution.From the model's residual distribution plot5-7, it can be observed that most residuals are distributed around0, with nosignificantskewness or heavy-tailed phenomena, which also indirectly indicates the reasonableness of the model's predictions.

北京交通大学毕业设计（论文）正文

Conclusion

This paper aims to study prediction methods for transfer traffic flow of airport arrival passengers.

Initially, this paper discussesthe current research status domestically and internationally, revealing that existing prediction methods have shortcomings in capturing multi-modal interaction mechanisms and complex spatiotemporal features. Traditional statistical models and single deep learning models struggle to adapt to the dynamic demands of airport landside transportation.

Based on this, this paper proposes a deep learning framework that integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory networks (LSTM). It leveragesCNNto extract local spatiotemporal features from multi-source data and employsLSTMto model long-term dependencies, achieving collaborative prediction of demand for three transfer modes: taxis, subways, and buses.

Taking Beijing Daxing International Airport as a case study, this research systematically analyzes the spatiotemporal aggregation characteristics of arriving flight passenger flows, transfer process mechanisms, and the correlations among transportation modes. Through in-depth exploration of features such as flight delay periods and lagged correlations between transfer modes, the importance of multi-source data fusion in improving prediction accuracy was validated. During the model validation phase, experimental results demonstrated that the CNN-LSTM model achieved R² values of 0.75 and 0.91 for bus and taxi demand predictions, respectively, with overall performance significantly outperforming baseline models such as ARIMA, GRU. Additionally, training efficiency was improved to 0.9 minutes, providing technical support for real-time prediction. Furthermore, ablation experiments confirmed the complementary nature of CNN and LSTM components, demonstrating that their combination effectively balances local feature extraction and long-term temporal modeling requirements.

This study has achieved certain results in multi-source data-driven airport landside transportation demand prediction, but there remains room for improvement and directions worthy of further exploration:

(1) The current research does not fully capture the geographical location information of transportation data collection points (e.g., spatial coordinates of taxi waiting pools, subway stations, and bus stops), making it difficult to construct accurate geospatial matrices to characterize the topological relationships of transfer facilities. Future work could integrate Geographic Information Systems (GIS) or high-precision sensor positioning data to establish dynamic spatial weight matrices, leveraging Graph Neural Networks (GNN) to model spatial interaction effects among transportation modes. For instance, node embeddings could be used to represent the differences in attractiveness and accessibility of various transfer facilities, thereby enhancing the prediction model's ability to capture local passenger flow distribution characteristics.

(2) This study primarily relies on structured data provided by airport operational systems and has not yet incorporated unstructured data sources such as mobile signaling data, social media dynamics, or transit card swipe records. For instance, mobile signaling data can reveal the spatiotemporal continuity of passenger travel chains, aiding in identifying cross-modal transfer route preferences; social media sentiment data can indirectly reflect passenger satisfaction with transportation services or the real-time impact of突发事件. Future research should explore fusion mechanisms for multi-source heterogeneous data, such as addressing data privacy issues through federated learning or designing attention mechanisms to dynamically weight the contributions of different data sources, thereby enhancing the model's generalization capability for complex scenarios.

(3) Although the model performs well in常规 scenarios, its adaptability to abnormal conditions such as large-scale flight delays, extreme weather, or突发公共卫生 events has not been fully validated. Future work could involve constructing synthetic datasets (e.g., simulating the coupling effects of widespread flight cancellations and ground traffic congestion during typhoons) or introducing adversarial training techniques to improve the model's tolerance to noise and outliers. Additionally, a reinforcement learning framework could be integrated to design dynamic feedback mechanisms, enabling the model to quickly adjust prediction strategies based on real-time events (e.g.,临时交通管制), thereby enhancing airport emergency management capabilities.

北京交通大学毕业设计（论文）附录

References