Source-linked AI summary
Deep-learning Architecture for Short-term Passenger Flow Forecasting in Urban Rail Transit
Jinlei Zhang, Feng Chen, Zhiyong Cui, Yinan Guo, Yadi Zhu
TL;DR
Urban rail operators need accurate, real-time, network-scale short-term passenger flow forecasts, while prior models often omitted external factors such as weather and air quality. The paper proposes ResLSTM, combining ResNet, GCN, and attention LSTM with four input branches. ResLSTM outperforms baseline models, quantifies external-factor effects, and achieves higher prediction precision at larger time granularities.
Problem
Prior urban rail forecasting models often neglected weather, events, or air quality, despite their relevance to passenger travel decisions.
Method
ResLSTM combines ResNet, GCN, and attention LSTM in four branches for inflow, outflow, graph topology, weather, and air quality.
Results
ResLSTM outperforms baseline models; at TG = 30 min, incorporating weather and air quality reduces RMSE from 60.13 to 56.96, MAE from 34.14 to 32.58, and WMAPE from 6.43 to 6.13%.
Takeaways & Limitations
The model provides favorable network-scale forecasting for nearly 300 subway stations and quantifies the influence of weather and air quality on prediction precision.
Takeaways & Limitations
The study reports several limitations but the supplied passage does not specify them beyond introducing the limitations discussion.
Abstract
from arXiv · showhide
Short-term passenger flow forecasting is an essential component in urban rail transit operation. Emerging deep learning models provide good insight into improving prediction precision. Therefore, we propose a deep learning architecture combining the residual network (ResNet), graph convolutional network (GCN), and long short-term memory (LSTM) (called "ResLSTM") to forecast short-term passenger flow in urban rail transit on a network scale. First, improved methodologies of the ResNet, GCN, and attention LSTM models are presented. Then, the model architecture is proposed, wherein ResNet is used to capture deep abstract spatial correlations between subway stations, GCN is applied to extract network topology information, and attention LSTM is used to extract temporal correlations. The model architecture includes four branches for inflow, outflow, graph-network topology, as well as weather conditions and air quality. To the best of our knowledge, this is the first time that air-quality indicators have been taken into account, and their influences on prediction precision quantified. Finally, ResLSTM is applied to the Beijing subway using three time granularities (10, 15, and 30 min) to conduct short-term passenger flow forecasting. A comparison of the prediction performance of ResLSTM with those of many state-of-the-art models illustrates the advantages and robustness of ResLSTM. Moreover, a comparison of the prediction precisions obtained for time granularities of 10, 15, and 30 min indicates that prediction precision increases with increasing time granularity. This study can provide subway operators with insight into short-term passenger flow forecasting by leveraging deep learning models.
I. INTRODUCTION
The paper motivates network-scale short-term passenger flow forecasting for urban rail transit and proposes ResLSTM to integrate spatial, temporal, topological, weather, and air-quality information.
- Short-term passenger flow forecasting supports passenger scheduling and operator decisions about congestion control and train timetables.
- Conventional statistical models often fail to meet real-time requirements and achieve high precision for traffic forecasting.
- Prior deep-learning studies omitted station spatial correlations, modeled only outflow, or applied environmental models to several stations rather than a whole network.
- Existing models often neglected weather and air quality, although passengers may adjust trips under bad weather or heavy pollution.
- ResLSTM combines ResNet, GCN, and attention LSTM for network-scale forecasting while incorporating spatiotemporal features, topology, weather, and air quality.
- The architecture is reported to support real-time network-scale predictions, remain robust when a branch is removed, and quantify weather and air-quality effects.
A. ResNet
The architecture uses residual blocks to support deeper feature extraction and GCNs to encode urban rail network topology. Improved residual connections address gradient problems, while graph transformations preserve topological information for later processing.
- A. ResNet: ResNet represents residual-block output as F(X_l) + X_l, where X_l and X_l+1 are the block input and output.
- A. ResNet: The improved residual block passes gradients through shortcut connections to mitigate vanishing or exploding gradients in deeper networks.
- B. GCN: GCN is applied because grid-based CNN representations neglect network-topology effects in urban rail traffic prediction.
- B. GCN: The graph uses vertices and edges to represent subway stations and their adjacent-node relationships, transforming inputs into features containing topological information.
- B. GCN: ResNet GCN is introduced to mitigate deeper-GCN drawbacks including backpropagation complexity, gradient vanishing, and oversmoothing.
C. Attention LSTM
The architecture uses attention LSTM to weight features from preceding network layers after branch processing. Its inflow branch organizes real-time, daily, and weekly histories into a three-channel input before residual processing and feature fusion.
- C. Attention LSTM: Attention LSTM obtains weights for features extracted from preceding network layers.The weighting mechanism is intended to account for complex influences beyond simple timestep recency.
- C. Attention LSTM: The model uses a fully connected network to score attention weights from the input or LSTM output.This replaces attention based solely on temporal proximity.
- B. Branch 2: Outflow: Branch 2 applies the same processing as Branch 1 to outflow data.The architecture therefore separates inflow and outflow processing.
- A. Branch 1: Inflow: The inflow branch processes its input through two residual blocks, flattening, and a 276-neuron fully connected layer before fusion.The residual blocks use 32 and 64 filters, respectively.
B. Branch 2: Outflow
Branch 2 processes outflow with the same architecture used for inflow in Branch 1.
- B. Branch 2: Outflow: Outflow processing in Branch 2 is identical to inflow processing in Branch 1.The Branch 2 input uses outflow data instead of inflow data.
C. Branch 3: Graph signal
The graph-signal branch uses ResNet GCN processing to capture subway network topology, while the external-factor branch incorporates weather and air-quality data. Weather and air-quality inputs are aligned to the forecasting granularity and passed through stacked LSTM processing.
- C. Branch 3: Graph signal: Branch 3 uses a ResNet GCN to capture the influence of subway network topology.The topology is treated as a real-time pattern because the network topology does not change.
- C. Branch 3: Graph signal: The graph transformation preserves input shape while adding rich network topological information for subsequent ResNet processing.The input represents stations across historical timesteps.
- D. Branch 4: Weather conditions and air quality: The external-factor branch uses weather indicators recorded every half-hour and air-quality indicators recorded every hour.Air-quality inputs include AQI, particulate matter, SO2, NO2, CO, and O3.
- D. Branch 4: Weather conditions and air quality: For 10-minute forecasting, weather observations are shared across their 30-minute recording interval and air-quality observations across their 60-minute interval.The preprocessing aligns lower-frequency environmental data with the forecasting timetable.
- D. Branch 4: Weather conditions and air quality: The environmental input contains 11 weather and air-quality indicators and is flattened before weighted processing with stacked LSTM layers.The stacked LSTM uses 128 neurons in its first layer and 276 in its second layer.
E. Feature fusion
The four branch outputs are combined through weighted feature fusion, followed by attention LSTM processing and a fully connected output layer. Model configuration uses separate training, validation, and test data with checkpointing and early stopping.
- E. Feature fusion: Feature fusion combines the four branch outputs through elementwise weighted addition.Each branch has a corresponding weight vector that is updated during backpropagation.
- E. Feature fusion: The fused features are processed by attention LSTM and then flattened into a 276-neuron fully connected layer for final prediction.Attention LSTM is applied after feature fusion.
- B. Model configuration: The experiments use Beijing subway AFC data collected across 17 lines and 276 stations over 25 workdays.The dataset contains 130 million records from February 29 to April 3, 2016.
- B. Model configuration: The first four weeks are used for training and the final week for testing, with a validation split rate of 0.2.The previous five network timesteps forecast the next one.
- B. Model configuration: Model Checkpoint and Early Stopping are used to retain the best model and avoid overfitting.Training and validation losses stabilize after 150 epochs following early oscillation.
C. Baseline models
The study benchmarks conventional, machine-learning, deep-learning, and ResLSTM variants using specified architectures, training settings, and evaluation metrics.
- Benchmark configurations: ARIMA uses station-specific models, whereas other benchmarks obtain overall results for 276 stations with a single model.The ARIMA comparison uses 276 models; other benchmark optimizers use Adam with learning rate 0.0001.
- Benchmark configurations: The benchmark set includes ARIMA, BPNN, RBF-SVR, Vanilla RNN, LSTM, GRU, CNN, ConvLSTM, and ResLSTM variants.Variants include ResLSTM-GCN, ResLSTM-No graph, ResLSTM-No W&A, ResLSTM-No A, and ResLSTM-TC.
- ResLSTM variants: ResLSTM-TC reorganizes inflow and outflow into two channels across three branches, with each branch containing one pattern and two channels.This configuration changes the representation of the first two branches while separating inflow and outflow.
- Training and evaluation: The models are trained end-to-end with MSE loss and Adam, using a learning rate of 0.001 for ResLSTM training.Performance is evaluated with RMSE, MAE, and WMAPE.
E. Results and discussion
Deep-learning models generally outperform earlier baselines, while the complete ResLSTM achieves the strongest performance and remains robust when branches are removed.
- Network-wide prediction performance: Deep-learning models considerably outperform mathematical-statistics and machine-learning models in most cases, while RBF-SVR and ARIMA perform worst.The authors attribute ARIMA’s poor performance to its inability to capture comprehensive nonlinear passenger-flow features.
- Network-wide prediction performance: Convolution-based models outperform recurrence-based models for single-model network prediction, and ConvLSTM outperforms CNN by capturing more temporal information.LSTM and GRU also outperform Vanilla RNN.
- Network-wide prediction performance: ConvLSTM performance worsens as time granularity increases, possibly because larger granularity reduces the number of samples.The proposed explanation is explicitly presented as a possible account.
- ResLSTM robustness: The complete ResLSTM performs best among its variants, while deleting one branch has only a negligible effect on prediction precision.The compared deletions include graph, weather-and-air-quality, and air-quality branches.
- ResLSTM robustness: ResLSTM’s advantage over its variants becomes clearer from 10 to 30 minutes, while topology contributes to prediction precision.Using only the topology branch also produces satisfactory results.
- External factors: At 30 minutes, adding weather and air quality decreases RMSE from 60.13 to 56.96, MAE from 34.14 to 32.58, and WMAPE from 6.43 to 6.13%.These comparisons quantify the reported contribution of the external-factor branch.
- ResLSTM variants: Treating passenger flow separately in ResLSTM-TC retains prediction precision while reducing computation cost.The comparison is between ResLSTM-TC and the complete ResLSTM.
COMPARISON OF PREDICTION PERFORMANCES OBTAINED USING DIFFERENT TGS IN DIFFERENT MODELS
Station-level comparisons across three time granularities show that ResLSTM tracks regular and irregular passenger-flow patterns, with better fitting at larger granularities.
- Station-level comparisons: The evaluation examines Tiantongyuan, Xizhimen, and Beijing West Railway stations, representing a residential community, traffic hub, and railway-station setting.The three stations provide contrasting passenger-flow contexts.
- Tiantongyuan station: At Tiantongyuan, predicted values align with actual values across 10-, 15-, and 30-minute granularities during peak and non-peak periods.The station’s large residential community produces a pronounced and regular morning peak.
- Xizhimen station: At Xizhimen, ResLSTM maintains favorable performance across all three granularities, especially during peak periods despite double-peak inflow patterns.Xizhimen is characterized as a large traffic hub.
- Beijing West Railway station: At Beijing West Railway station, ResLSTM captures the variation trend despite low regularity, and fitting improves from 10 to 30 minutes.This station exhibits significant inflow variation.
- Station-level comparisons: The proposed model provides precise prediction at both network and individual-station scales.This is the section’s stated summary conclusion.
3) Prediction performance in different TGs
Aggregating predictions to larger time granularities improves evaluation performance, and the study concludes that ResLSTM is robust for practical short-term passenger-flow forecasting.
- Prediction performance in different TGs: Prediction precision gradually increases as time granularity increases from 10 to 30 minutes.Results at 10 and 15 minutes were aggregated into corresponding 30-minute results before evaluation.
- Prediction performance in different TGs: When comparing 10×3 minutes with 30 minutes, RMSE decreases from 61.55 to 56.96, MAE from 35.22 to 32.58, and WMAPE from 6.63 to 6.13%.The authors associate this improvement with greater passenger-flow similarity and regularity after aggregation.
- Conclusion: ResLSTM shows satisfactory short-term passenger-flow forecasting ability and strong robustness favorable for practical real-world applications.This is the study’s concluding assessment.
V. CONCLUSION
ResLSTM combines ResNet, GCN, and attention LSTM while incorporating weather and air-quality effects into network-scale passenger-flow forecasting. The study reports robust forecasting and finds that prediction precision increases with larger time granularity.
- ResLSTM combines ResNet, GCN, and attention LSTM to capture passenger-flow features and network topology while incorporating weather and air-quality influences.The architecture also quantifies how weather conditions and air quality affect prediction precision.
- Air-quality indicators, including PM2.5 and PM10, were considered in short-term passenger flow forecasting for the first time to the authors’ knowledge.
- ResLSTM showed strong robustness and favorable prediction precision for network-scale forecasting involving nearly 300 subway stations.
- Weather conditions and air quality had considerable influence on prediction precision, and their influence was quantified.
- Prediction precision increased with increasing time granularity because aggregated passenger flows had higher similarity and regularity.