Traffic Flow Forecasting using LSTM

Urban congestion costs commuters hours and cities fuel, and traffic-management centers already sit on the data to anticipate it: loop detectors report speed and occupancy every few minutes. The operational question is always forward-looking — which corridors will jam in the next hour, so signals, ramp meters and advisories can respond before the queue forms. This project answers it with sequence learning: a two-layer LSTM trained on the METR-LA benchmark learns each sensor's daily rhythm and predicts 12 steps (one hour) ahead from a one-hour input window. Time-of-day and day-of-week embeddings let a single model serve all 207 sensors. The dashboard presents the result the way a control room would use it: a schematic network map, per-sensor forecast charts with uncertainty bands, and a ranked congestion watchlist.

1. METR-LA readings (speed + occupancy, 5-min, Mar–Jun 2012) are cleaned, normalized per sensor, and framed as 12-in/12-out sequences.
2. Time-of-day and day-of-week embeddings are concatenated to the sensor features so the model knows which regime it is in.
3. A 2-layer LSTM (64 + 32 units) with a dense 12-step head is trained with MAE loss and early stopping on held-out weeks.
4. Held-out weeks are scored with MAE/RMSE at 15/30/60-minute horizons for the report.
5. At inference, each sensor's latest hour feeds the model, producing 12 future speeds plus bands.
6. The dashboard renders the network map, per-sensor charts, forecast tables and the congestion watchlist.

Tech stack:

• Python 3, TensorFlow/Keras (LSTM)
• pandas, NumPy
• scikit-learn (scaling, metrics)
• HTML/CSS/JavaScript dashboard (canvas charts)
• METR-LA dataset

Dataset & model details

• Dataset: METR-LA (Li et al.) — 207 Los Angeles highway loop detectors, speed + occupancy at 5-minute resolution, March–June 2012 (34,272 timesteps); the standard traffic-forecasting benchmark.
• Task: Multi-step sequence forecasting; input = 12 steps × (speed, occupancy, time embeddings), output = 12 future speeds per sensor.
• Model: 2-layer LSTM (64 + 32 units), dense 12-step head; trained with Adam, MAE loss, early stopping.
• Metrics: 60-min MAE 3.4 mph (design target for the built-to-order training run); reported at 15/30/60-min horizons. No measured figures are claimed before the run.

• [1-hour speed forecasts] 2-layer LSTM predicts 12 five-minute steps ahead per sensor from a 12-step input window of speed + occupancy.
• [Prediction bands] Every forecast carries a ±1σ band so dispatchers see uncertainty, not just a line.
• [Schematic network map] All 207 sensors plotted on freeway corridors, colored by current speed band (free flow / heavy / congested).
• [Per-sensor forecast charts] Select any sensor to see its 3-hour history plus the 1-hour forecast with the evening-peak collapse and recovery.
• [Congestion watchlist] Sensors ranked by current and forecast speed with cause tags (evening peak, merge bottleneck, recovering).
• [24-hour speed heatmap] The twin AM/PM peak structure visualized per sensor — the pattern the LSTM must learn.
• [Time-aware modeling] Time-of-day and day-of-week embeddings let one model cover all sensors and both weekday and weekend regimes.

• METR-LA ingestion & sequence-framing scripts
• LSTM training notebook with horizon-wise evaluation
• Trained model weights and per-sensor scalers
• Interactive network dashboard (map, forecasts, watchlist, heatmap)
• Project report PDF (background, sequence-model theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (LSTMs, multi-step forecasting, leakage in time series)