TL;DR
This paper introduces a novel deep learning model, CRS-ConvLSTM, that predicts short-term traffic states by focusing on "critical road sections" identified through spatiotemporal correlations. This approach effectively handles corrupted or missing data by leveraging the predictive power of these key sections, outperforming existing methods on a real-world Beijing traffic network.
What Problem Does This Paper Address?
Accurate short-term traffic prediction is crucial for efficient traffic management and public services. However, real-world traffic data often suffers from corruption and missing values due to factors like GPS signal degradation or equipment failures. Traditional methods struggle with this "structural missing data," leading to reduced prediction accuracy. This paper tackles this challenge by proposing a method that doesn't try to fill in all missing data but instead intelligently selects a subset of "critical road sections" that have the most influence on the overall network's traffic state. By focusing on these key sections, the model can predict the traffic evolution of the entire network, even when data from other sections is unreliable.
Methods
The core of this work is a hybrid deep learning model called CRS-ConvLSTM. It combines Convolutional Neural Networks (CNNs) for capturing spatial features and Long Short-Term Memory (LSTM) networks for temporal dependencies.
The novelty lies in the preceding step: identifying "critical road sections." This is achieved using a Spatiotemporal Correlation Algorithm (STCA). Here's a breakdown:
- Spatial Weights Matrix: The road network's topology is represented as a graph, and a spatial weights matrix is constructed to capture the relationships between adjacent road sections.
- Spatiotemporal Similarity: Pearson's correlation is used to measure the similarity between traffic speed time series of road sections, considering time lags. This is then combined with the spatial weights to calculate a "correlation distance" that quantifies the influence of one road section on its neighbors.
- TOPSIS for Ranking: To synthesize these correlations across different time lags and road sections, the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) is employed. This ranks road sections based on their overall impact.
- Critical Road Section Selection: A certain proportion of the top-ranked road sections are designated as "critical road sections."
- ConvLSTM Prediction: The traffic speed data from these identified critical road sections is then fed into the ConvLSTM model to predict the traffic states of the entire network. The ConvLSTM itself uses a CNN to extract spatial features from the critical road section data and an LSTM to model the temporal evolution.
Datasets
The study uses real-world traffic speed data collected from GPS devices in taxis in Beijing, China. The dataset covers 92 working days from June 1, 2015, to October 13, 2015. A specific sub-transportation network near the West Second Ring Road, comprising 278 road sections, was selected for the experiments. Two-thirds of the data (61 days) were used for training, and the remaining one-third (31 days) for testing.
Key Insights and Ideas
- Focusing on "Critical" Sections is Key: The paper demonstrates that not all road sections are equally important for predicting overall traffic. Identifying and leveraging "critical road sections" with high spatiotemporal influence can significantly improve prediction accuracy, especially in the presence of noisy or missing data.
- Spatiotemporal Correlation as a Measure of Influence: The STCA effectively quantifies how much one road section's traffic state influences others, considering both spatial proximity and temporal dynamics. This is a novel way to distill complex network interactions.
- TOPSIS for Multi-Criteria Ranking: Using TOPSIS to integrate correlation distances across various time lags provides a robust method for ranking road sections by their overall impact, overcoming the limitations of considering only single time lags.
- Handling Structural Missing Data: Instead of complex imputation, the model bypasses problematic data by focusing on the predictive power of critical sections, offering a more practical solution for real-world data challenges.
- Dynamic Critical Sections: The study acknowledges that critical road sections can change over time (e.g., during peak vs. off-peak hours), suggesting that a dynamic approach to identifying these sections could further enhance predictions.
Key Results
- The proposed CRS-ConvLSTM model outperforms several established deep learning algorithms (ANN, CNN, LSTM, SAE) in terms of RMSE and RMSEP.
- With an extracting rate of 0.70 (selecting 195 out of 278 road sections), the CRS-ConvLSTM achieved an RMSE of 6.955 and an RMSEP of 21.802%.
- The model's performance is shown to be sensitive to the number of critical road sections used. Generally, as the extracting rate decreases (fewer critical sections), the RMSE increases, indicating that a certain level of redundancy or inclusion of less critical sections can be beneficial.
- The study found that using 5-order neighbors (k=5) for spatial weights provided the best performance, suggesting that influence beyond a certain spatial range is minimal.
- In stochastic and extreme missing data scenarios, the CRS-ConvLSTM still showed better performance than using the entire network data, highlighting its robustness.
Limitations and Future Work
- Data Quality: The paper acknowledges that random missing values still affect prediction accuracy. Future work could explore tensor decomposition for more advanced imputation methods.
- Road Attributes: The current STCA primarily uses topological information. Future research could incorporate road hierarchy, length, and other inherent attributes for more nuanced spatial correlation extraction.
- External Factors: The authors suggest that incorporating external factors like weather, temperature, and air quality could further improve prediction accuracy.
- Dynamic Critical Sections: While the paper notes that critical sections can change over time (e.g., by time of day), the current STCA identifies them based on aggregated data. Future work could explore dynamically identifying critical sections for more adaptive predictions.
- Model Complexity: The paper notes that deep learning models require substantial historical data (at least 3 months) for training.
Summary updated