Railway-Centric Spatio-Temporal GCN

• The paper introduces RSTGCN, a model that predicts aggregate station-level average delays instead of individual train delays.
• It utilizes train frequency-aware spatial attention and Chebyshev polynomial-based graph convolutions to effectively capture railway-specific delay propagation.
• Evaluated on the Indian Railway Network, the model outperforms baselines by 13–15% in MAE, showcasing robust performance in real-world forecasting.

Searching arXiv for the named RSTGCN paper and closely related spatio-temporal GNN work in railway/transport forecasting. Railway-centric Spatio-Temporal Graph Convolutional Network (RSTGCN) is a spatio-temporal graph neural architecture introduced for train delay prediction at station level, specifically to forecast the average arrival delays of all the incoming trains at railway stations for a particular time period (Chowdhury et al., 26 Sep 2025). In contrast to earlier approaches centered on forecasting the exact delays of individual trains, RSTGCN targets an aggregate operational variable aligned with higher-level traffic management. The model is explicitly railway-centric in its graph construction, feature design, and spatial attention mechanism, most notably through train frequency-aware spatial attention, and is evaluated on a nationwide dataset covering the entire Indian Railway Network (IRN), spanning 4,735 stations across 17 zones (Chowdhury et al., 26 Sep 2025).

1. Problem formulation and scope

RSTGCN is designed for station-level aggregate delay prediction rather than individual-train forecasting (Chowdhury et al., 26 Sep 2025). The prediction target is the average arrival delay of all incoming trains at a station over a defined time period, such as an hour. This places the model in a distinct operational regime: its outputs are intended to support scheduling and dispatching decisions at network scale rather than train-specific passenger-facing estimates.

The railway network is modeled as an undirected graph

$G = (S, E, A, M, K),$

where $S$ is the set of stations, $E$ the set of edges, $A$ the adjacency matrix, $M$ the distance matrix, and $K$ the train frequency matrix (Chowdhury et al., 26 Sep 2025). An edge exists if two stations are consecutive on any train's route. The distance component is encoded as an inverse-distance matrix between adjacent stations, while the traffic intensity between adjacent stations is represented by a train frequency matrix in which $k_{ij}$ is the number of trains operating between adjacent stations $i$ and $j$ (Chowdhury et al., 26 Sep 2025).

This graph formulation situates RSTGCN within the broader spatio-temporal GNN literature, where network-structured time series are represented as signals on nodes. STGCN established the general principle that traffic systems can be modeled with graph convolutions for space and temporal convolutions for time (Yu et al., 2017). RSTGCN specializes that principle to railway operations by embedding railway-specific quantities—particularly train frequency and headway—directly into the representation and attention mechanism (Chowdhury et al., 26 Sep 2025).

2. Railway graph representation and input features

RSTGCN uses station nodes and railway adjacency defined by actual train-route consecutiveness, not by an image grid or a purely Euclidean neighborhood (Chowdhury et al., 26 Sep 2025). This is consistent with the graph-based treatment of transport systems in prior work such as STGCN (Yu et al., 2017), but RSTGCN adds railway-specific edge semantics through train frequency and distance weighting.

The station features used in the model are:

• Hourly average arrival delay
• Hourly average departure delay
• Total hourly arrival delay
• Total hourly departure delay
• Hourly headway (Chowdhury et al., 26 Sep 2025)

These features are central to the model’s railway-centric character. The paper identifies the inclusion of headway and total delay features as a methodological innovation relevant to real operational bottlenecks (Chowdhury et al., 26 Sep 2025). A plausible implication is that the model is intended to encode not only local lateness levels but also traffic density and temporal spacing, which are fundamental to railway propagation phenomena.

The corresponding dataset is described as follows: 4,735 stations, 9,336 edges (direct connections), 3,892 long-distance passenger trains, with hourly aggregated statistics per station for Sept 2024 over 720 hours (September 2024) (Chowdhury et al., 26 Sep 2025). The partitioning protocol uses the first 7 days as historical input (not for training/testing), 20% most recent as test set, rest for training/validation (Chowdhury et al., 26 Sep 2025). Because the paper also states that the IRN is the largest and most diverse railway network studied to date, RSTGCN is framed as a large-scale national-network model rather than a corridor- or city-level predictor (Chowdhury et al., 26 Sep 2025).

3. Architecture and core computations

RSTGCN comprises three parallel modules that encode recent, daily, and weekly historical dependencies (Chowdhury et al., 26 Sep 2025). Each module contains the same sequence of components:

• Temporal attention layer
• Spatial attention layer incorporating train frequency
• Graph convolutional layer (GCN)
• 2D-CNN layer (Chowdhury et al., 26 Sep 2025)

The outputs of these three modules are fused through learnable weights, and a ReLU activation is applied at the end to ensure non-negative predictions (Chowdhury et al., 26 Sep 2025). The use of recent, daily, and weekly branches places RSTGCN in continuity with prior transport forecasting models that explicitly separate temporal regimes. For example, Conv-GCN processes recent, daily, and weekly inflow/outflow patterns separately before deeper integration (Zhang et al., 2020). In RSTGCN, however, these temporal regimes are embedded inside a station-delay forecasting architecture specialized to railway operations (Chowdhury et al., 26 Sep 2025).

The temporal attention module is defined by

$$Z = V_t \cdot \sigma\left( \left( (X U_1) U_2 \right) \odot (U_3 X) + b_t \right),$$

where $S$0 are learnable parameters, $S$1 is the input feature tensor, and $S$2 is a non-linearity (Chowdhury et al., 26 Sep 2025). The role of this module is to capture dependencies across multiple temporal granularities, specifically recent hours, daily pattern, and weekly pattern for each station.

After temporal attention, the model applies a spatial attention layer, followed by graph convolution and 2D-CNN integration (Chowdhury et al., 26 Sep 2025). The graph convolution uses Chebyshev polynomials for spectral filtering, with the implementation detail that the GCN uses order-3 Chebyshev polynomials (Chowdhury et al., 26 Sep 2025). This situates the model technically close to the Chebyshev-based spectral graph filtering family also used in STGCN (Yu et al., 2017).

The module outputs are fused as

$S$3

where $S$4 are learnable weights for combining recent, daily, and weekly modules (Chowdhury et al., 26 Sep 2025). The final prediction is

$S$5

which the paper states is intended to ensure non-negative predictions, aligning with the observation that early arrivals are rare (Chowdhury et al., 26 Sep 2025).

4. Train frequency-aware spatial attention

The defining innovation of RSTGCN is its train frequency-aware spatial attention, which modulates delay influence by both geographic separation and the intensity of rail service between adjacent stations (Chowdhury et al., 26 Sep 2025). This mechanism is introduced to model how delay at one node affects its neighbors under railway-specific operational constraints.

The correlation matrix from temporally attended features is computed as

$S$6

and the train frequency- and distance-weighted adjacency is defined as

$S$7

where $S$8 is the distance between stations $S$9 and $E$0, and $E$1 is the number of trains, normalized by the network maximum $E$2 (Chowdhury et al., 26 Sep 2025). The final spatial attention matrix is

$E$3

The stated effect of this design is that delay influence becomes stronger for nearby stations with many shared trains, thereby capturing real patterns of delay propagation more accurately than distance-only constructions (Chowdhury et al., 26 Sep 2025). This directly distinguishes RSTGCN from earlier distance-aware formulations such as TSTGCN, which is listed as a baseline, and from generic graph models whose adjacency is fixed by topology alone (Chowdhury et al., 26 Sep 2025).

This innovation also places RSTGCN in dialogue with a wider methodological trend in transport GNNs toward adaptive or multi-relational spatial structure. CCRNN replaces a single heuristic adjacency with layer-wise self-learned adjacency matrices coupled across layers (Ye et al., 2020). STGCGRN combines a predefined distance graph with a self-adaptive adjacency matrix and multi-head hidden dependencies (Zhao et al., 2022). MGC-RNN uses multiple graphs to encode heterogeneous inter-station correlations (He et al., 2021). RSTGCN differs from these approaches by encoding a specific railway operational variable—train frequency—into the spatial attention mechanism itself rather than introducing fully learned adjacency matrices or multiple parallel relational graphs (Chowdhury et al., 26 Sep 2025).

5. Experimental setting, baselines, and reported performance

The evaluation uses standard regression metrics:

• Mean Absolute Error (MAE)
• Mean Absolute Percentage Error (MAPE)
• Root Mean Square Error (RMSE) (Chowdhury et al., 26 Sep 2025)

Metrics are reported across all stations and all test timestamps (Chowdhury et al., 26 Sep 2025). The baselines include:

• Historical Average (HA)
• Random Forest (RF)
• LSTM
• GRU
• STGCN
• ASTGCN
• TSTGCN (Chowdhury et al., 26 Sep 2025)

All neural models use hidden size 64, batch size 4, learning rate 0.001, with a max prediction window: 3 hours in primary experiments; also validated for up to 12 hours (Chowdhury et al., 26 Sep 2025).

The main reported result is that RSTGCN consistently outperforms all baselines on all metrics (MAE, MAPE, RMSE) across 17 railway zones and the full IRN (Chowdhury et al., 26 Sep 2025). For the whole IRN, averaged over the 1–3 hour horizon, the paper reports:

Model	MAE	MAPE	RMSE
RSTGCN	0.077	9.64%	0.370
TSTGCN	0.089	11.10%	0.375
ASTGCN	0.091	11.15%	0.413

The paper further summarizes the gains as 13–15% in MAE, 9–13% in MAPE, 1–10% in RMSE over closest baselines (Chowdhury et al., 26 Sep 2025). It also states that for long-term (up to 12-hour horizon), RSTGCN maintains lower error increase relative to baselines, which is presented as evidence of robust generalization (Chowdhury et al., 26 Sep 2025).

Ablation studies report that the inclusion of the novel features, especially headway and total delay, and the architectural innovations, especially train frequency-aware spatial attention and ReLU activation, each lead to clear improvements, and that both refinements are needed for best performance (Chowdhury et al., 26 Sep 2025).

6. Position within related spatio-temporal graph learning research

RSTGCN belongs to the broader class of spatio-temporal graph forecasting models, but its distinguishing contribution lies in adapting this class to railway-wide delay aggregation. STGCN provided an influential fully convolutional template using graph convolution for spatial structure and gated temporal convolution for sequential dependence, emphasizing faster training speed with fewer parameters relative to recurrent alternatives (Yu et al., 2017). RSTGCN adopts the graph-based forecasting paradigm but uses attention-based temporal modeling, a railway-specific spatial weighting, and 2D-CNN integration (Chowdhury et al., 26 Sep 2025).

Relative to transport demand models, RSTGCN shares with Conv-GCN the use of recent, daily, and weekly historical structure (Zhang et al., 2020). It also shares with MGC-RNN and CCRNN the general aim of going beyond a purely heuristic adjacency representation of transport systems (He et al., 2021, Ye et al., 2020). However, those models target transportation demand or passenger flow, whereas RSTGCN focuses on delay prediction and introduces railway-operational variables that are absent from generic urban mobility models (Chowdhury et al., 26 Sep 2025).

A useful contrast arises with STG-GAN, which addresses short-term passenger flow prediction in urban rail transit through an adversarial framework composed of a generator using gated temporal conventional networks (TCN) and weight sharing graph convolution networks (GCN) and a discriminator with separate spatial and temporal components (Zhang et al., 2022). STG-GAN is centered on passenger flow realism, efficiency, and memory occupancy, whereas RSTGCN is centered on average station delay prediction and train frequency-aware propagation (Zhang et al., 2022, Chowdhury et al., 26 Sep 2025).

Another contrast appears in event-level railway delay forecasting. The later GATv2-based work on cascading train delays represents the system as a spatio-temporal graph of operational events (arrivals and departures) with Running, Dwell, and Headway edges, and emphasizes explainability and autoregressive multi-step evaluation (Nguyen et al., 10 Oct 2025). By comparison, RSTGCN is station-level rather than event-level, and one-stage regression-oriented rather than a two-stage hurdle model (Chowdhury et al., 26 Sep 2025, Nguyen et al., 10 Oct 2025). This suggests a methodological divide between aggregate station management models and fine-grained operational event models.

7. Limitations, interpretation, and research directions

The paper states several limitations explicitly. RSTGCN focuses on aggregate, station-level average delays, not on individual train delays or downstream customer experience (Chowdhury et al., 26 Sep 2025). It excludes factors like weather and real-time disruptions due to lack of uniform data, is primarily trained and tested on one month's data, and does not yet address dynamic (real-time) adaptation or integration with external data sources (Chowdhury et al., 26 Sep 2025).

These limitations matter because they delimit what the model can legitimately be interpreted as predicting. RSTGCN is not presented as a full causal model of delay propagation, nor as a train-by-train operational simulator. Rather, it is a large-scale predictive model for station-level average arrival delay under the feature regime available in the released dataset (Chowdhury et al., 26 Sep 2025).

The future directions listed in the paper are to incorporate external factors such as weather, maintenance events, or operational disruptions; extend to real-time deployment for adaptive schedule management; explore individual train-level predictions and customer-centric metrics; improve scalability and robustness in sparser or dynamically changing sections of the network; and use the open IRN dataset as a foundation for further railway analytics research (Chowdhury et al., 26 Sep 2025).

A plausible implication is that RSTGCN may serve as a benchmark architecture for railway-scale aggregate delay forecasting in the same way STGCN became a benchmark architecture for traffic forecasting on graphs (Yu et al., 2017). Its significance lies less in introducing graph forecasting itself than in specifying how graph forecasting should be adapted when the domain is a nationwide railway network and the target is average station delay rather than passenger flow or sensor speed (Chowdhury et al., 26 Sep 2025).