As shown in Table 1, analysis reveals a stable “three-level periodicity” structure in the temporal distribution of passenger flow at GSRS [1]:

Daily Cycle: Exhibits a distinct pattern of “dual peaks on weekdays, single peak on weekends”. The weekday morning peak (8:00-10:00) is dominated by commuters on the Guangzhou-Shenzhen and Guangzhou-Zhuhai corridors, while the evening peak (18:00-20:00) combines arriving and departing flows. Weekend peaks are more evenly distributed between 13:00 and 18:00.

Weekly Cycle: Passenger volume climbs from Monday, peaks on Thursday-Friday for weekdays, and reaches the weekly maximum on Saturday, approximately 25% higher than the average weekday volume.

Annual Cycle: Characterized by four major peak periods: the “Spring Festival Extreme Peak”, “Summer Transport Sub-peak”, “Short Holiday Pulse Peaks”, and the “Canton Fair Peak” [1].

Table 1. Passenger flow impact indices for key holidays at GSRS (2023).

Note: Impact Index = Peak Daily Volume/Monthly Average Volume.

Spatial distribution shows strong “directional agglomeration” and “transfer dependency”, as shown in Figure 1.

OD Agglomeration: The Shenzhen/Hong Kong SAR direction accounts for the highest share (31.3%), forming, together with Zhuhai/Macao (19.0%) and Changsha/Wuhan (18.2%), the top three dominant flows, constituting nearly 70% of the

Figure 1.Spatiotemporal heatmap of passenger flow (Left) and major OD flow diagram (Right).

total and forming a stable “GBA Core Corridor” [2].

Transfer Structure: Metro is the overwhelmingly dominant transfer mode, with a share of 65.2%. Notably, the combined share of taxis and ride-hailing services can surge to over 35% during late-night hours (after 22:00) and under adverse weather conditions, demonstrating significant “spatiotemporal elasticity”.

Passenger categorization was performed using a combination of K-means clustering (applied to features such as booking lead time, travel frequency, ticket class, and temporal patterns) and rule-based classification (e.g., same-day return trips on weekdays classified as business travel).

Through analysis of ticket class, booking lead time, and POI correlation, passengers are profiled in detail [2]:

Business Travelers (42.3%): High-frequency travelers between Guangzhou-Shenzhen/Zhuhai, concentrated on weekdays, highly sensitive to departure/arrival times, and with the lowest tolerance for transfer time.

Tourists (28.7%): Concentrated on weekends, holidays, and summer travel periods, often carrying luggage, showing greater concern for wayfinding signage and rest facilities within the hub.

Commuters (8.9%): Exhibit stable “tidal” characteristics and are a primary source of pressure on metro systems during peak hours.

Addressing the highly nonlinear, non-stationary nature of HSR hub passenger flow series and their complex influence by multiple external factors [3], this study proposes a combined forecasting framework integrating hybrid deep learning and ensemble learning. The model architecture was designed to sequentially address different aspects of the forecasting challenge: VMD handles non-stationarity and multi-scale patterns; CNN extracts local spatial-temporal features; BiLSTM captures long-term bidirectional dependencies; Attention focuses on relevant historical periods; and XGBoost integrates deep features with external variables for robust ensemble learning. The core innovation lies in its three-stage architecture of Decomposition-Reconstruction-Fusion.

1) Signal Decomposition Layer: Employs Variational Mode Decomposition (VMD) to adaptively decompose the original passenger flow series into a set of quasi-stationary sub-series [7].

2) Deep Forecasting Layer: For each decomposed sub-series, a CNN-BiLSTM-Attention neural network is designed for deep feature extraction and forecasting [6][8].

3) Ensemble Output Layer: The deep forecasting outputs are combined with external features and fed into an XGBoost model for nonlinear ensemble and residual correction [4].

The mathematical representation of the model is:

where I M F K is the k-th modal component forecasted by the deep sub-network, E t is the vector of external features at time t.

VMD decomposes the original signal f(t) into K band-limited Intrinsic Mode Functions (IMFs) u k ( t ) by solving a constrained variational problem [8]:

The decomposition number K was determined through a combination of spectral analysis and trial evaluation on a validation set. K = 6 was selected as it clearly separated the series into interpretable components: ultra-high-frequency noise, daily periodicity, weekly periodicity, holiday/monthly periodicity, seasonal trend, and long-term trend.

Each IMF component is input into an independent, identically structured deep neural network consisting of three layers:

1) 1D Convolutional Layer (1D-CNN)

Extracts local fluctuation patterns:

2) Bidirectional Long Short-Term Memory Layer (BiLSTM)

Comprehensively learns the dynamic evolution of each IMF within its full context:

3) Attention Mechanism Layer:

Dynamically weights historical time steps, allowing the model to focus on key historical periods relevant to the current forecast [8]: C = ∑ t α t H t . The context vector C serves as the final deep feature representation for forecasting that IMF component.

This stage aims to fuse the intrinsic temporal patterns extracted by the deep network with rich external influencing factors.

1) External Feature Engineering: A feature pool containing over 30 features across 5 categories is constructed (temporal, historical, economic/event, weather, competing transport). Examples include: binary indicators for public holidays and Canton Fair periods; temperature, precipitation, and visibility as weather va- riables; and average ticket prices for competing transport modes (e.g., flights, coaches).

2) XGBoost Nonlinear Ensemble: The forecasted values of all IMF components are concatenated with the external feature vector as input to XGBoost. XGBoost learns the complex nonlinear mapping between these features and the final passenger flow target, performing residual correction [4]. Its objective function is:

The dataset was split into training (Jan 2021-Jun 2023), validation (Jul 2023-Sep 2023), and test (Oct 2023-Dec 2023) sets. Missing values were forward-filled, and outliers beyond three standard deviations were winsorized, as shown in Table 2. Evaluation on an independent test set shows the proposed model significantly outperforms benchmark models across multiple metrics [5], as shown in Table 3.

Table 2. Model hyperparameters and tuning strategy.

Table 3. Comprehensive comparison of model forecasting performance (test set results).

Accuracy Improvement: Compared to the second-best model (VMD-LSTM), our model reduces MAPE by a further 26.6%.

Peak Forecasting Capability: For the National Day holiday peak contained in the test set, our model's forecast error for the single-day maximum passenger volume is only 6.4%, demonstrating excellent fitting ability for extreme values.