Check out our model performance.
Evaluations of ST-GAT Single Edge Models on our graphs
| Graph |
15 min |
|
|
30 min |
|
|
45 min |
|
|
| |
RMSE |
MAE |
MAPE (%) |
RMSE |
MAE |
MAPE (%) |
RMSE |
MAE |
MAPE (%) |
| 1 |
3.77 |
2.01 |
4.67 |
3.90 |
2.06 |
4.83 |
4.04 |
2.13 |
5.03 |
| 2 |
3.89 |
2.04 |
4.73 |
3.97 |
2.08 |
4.90 |
4.07 |
2.11 |
5.03 |
| 3 |
3.85 |
2.01 |
4.67 |
3.94 |
2.05 |
4.82 |
4.08 |
2.15 |
5.08 |
| 4 |
3.79 |
2.05 |
4.68 |
3.91 |
2.08 |
4.87 |
4.07 |
2.16 |
5.06 |
| 5 |
3.91 |
2.07 |
4.79 |
4.00 |
2.14 |
5.02 |
4.16 |
2.13 |
5.00 |
| 6 |
3.89 |
2.03 |
4.72 |
4.07 |
2.09 |
4.93 |
4.03 |
2.06 |
4.93 |
Evaluations of ST-GAT Edge Type Models on our graphs
| Graph |
15 min |
|
|
30 min |
|
|
45 min |
|
|
| |
RMSE |
MAE |
MAPE (%) |
RMSE |
MAE |
MAPE (%) |
RMSE |
MAE |
MAPE (%) |
| 1 |
3.79 |
1.98 |
4.62 |
3.97 |
2.03 |
4.87 |
3.98 |
2.05 |
4.91 |
| 2 |
3.78 |
2.01 |
4.69 |
3.94 |
2.04 |
4.89 |
4.06 |
2.13 |
5.06 |
| 3 |
3.87 |
2.03 |
4.66 |
4.11 |
2.13 |
5.02 |
4.11 |
2.13 |
5.02 |
| 4 |
3.82 |
2.02 |
4.71 |
3.94 |
2.03 |
4.86 |
4.03 |
2.12 |
4.99 |
| 5 |
3.93 |
2.02 |
4.75 |
3.97 |
2.07 |
4.86 |
4.04 |
2.12 |
5.01 |
| 6 |
3.85 |
2.06 |
4.71 |
4.01 |
2.09 |
4.87 |
4.10 |
2.10 |
5.00 |
Results
- Forecasting traffic speeds for the upcoming 15 minutes consistently outperformed predictions for longer timeframes, demonstrating the model’s strength in near-term accuracy.
- The inclusion of extra features led to a dip in performance, likely attributed to the over-generalization of time to hourly increments rather than to more precise intervals.
- Graphs using ST-GAT Edge Type generally performed better than its ST-GAT Single Edge counterpart which shows the importance of differentiating edge types.
- Graphs containing Type 1 edges performed the best which may be due to the close proximity of our sensors and nearest neighbor speeds providing the most reliable information for predicting a target sensor’s speed.
Future Work
- Make data collection process automated for more efficient retrieval of sensor data
- Collect sensor data that extends a longer time frame to increase sample size for model training
- Integrate other highways in San Diego County to capture more highway relations
- Explore advanced optimization techniques such as Grid Search Cross Validation to optimize hyperparameters
- Incorporate additional features that factor into traffic patterns such as weather and lane closures to enrich feature vectors
- Integrate real-time feedback mechanisms into prediction models to enable continuous learning and adaptation to changing traffic and feature conditions