Key Features (LSTM Traffic Predictor)
Prime Target
Status: Not yet implemented
- Achieve MAE < 10% on raw traffic values
Time-Series Forecasting
- LSTM Neural Network: Long Short-Term Memory architecture for temporal pattern recognition
- Learns long-term traffic patterns (morning/evening peaks)
- Remembers relevant historical events, forgets irrelevant ones
- Handles non-linear traffic spikes and sudden congestion
- Architecture: 64-unit LSTM layer with dropout regularization (0.2)
Prediction Capabilities
- 1-Hour Ahead Forecasting: Predicts edge traffic density for the next hour
- Sequence Length: Uses 3 hourly measurements (timesteps) for predictions
- Achieved Performance: MAE 0.2084 on normalized density values (test set)
- Input: 3 consecutive hourly measurements × 5 edge features
- Output: Predicted density for each of the 5 edges at next hour
Data Processing
- SUMO Integration: Loads historical traffic data from
edgeData.xml - Feature Selection: Focuses on top 5 most congested edges by average density
- Normalization: MinMaxScaler (0-1 range) for numerical stability
- Sliding Windows: Creates sequences of 3 timesteps with 1-step forecast target
- Data Split: 80/20 train-test split (temporal, no shuffle)
Supported Edges
Trained on the 5 most congested edges from SUMO simulation:
- -269002813
- -55825089
- 617128762
- -617128762
- -312266114#2
API Endpoints
-
GET /health— Service health check -
GET /model-info— Model architecture and performance metadata -
POST /predict— Predict edge density for next hour
Request Format: Array of (3 × 5) raw density measurements Response Format: Array of 5 predicted density values
Model Performance
- Test Loss (MSE): 0.0698
- Test MAE: 0.2084 (normalized scale)
- Training: 50 epochs with batch size 2
- Optimizer: Adam (learning rate: 0.001)
- Validation: Evaluated on 20% test set
Integration
- RL Inference Service: Can consume predictions for context-aware signal timing decisions
- API Gateway: Can proxy
/predictrequests with unified authentication - SUMO Source: Uses simulation outputs (
edgeData.xml) as training data
Future Enhancements
- 15-minute ahead forecasting (requires model retraining)
- Handle missing sensor data (KNN imputation)
- Bidirectional LSTM for enhanced accuracy
- Attention mechanism for temporal weighting
- Real-time model retraining capability
- Weather data integration