Prediction table
Report MAE, RMSE, R² for LSTM vs hybrid on the Thiruvananthapuram research split.
Compiling the hour
Compiling the hour
Research
A city does not grade you on mean absolute error. It grades you on whether a bus clears East Fort before the school bell. These tables are the lab. The policy is the field.
Traffic prediction
| Model | MAE | RMSE | R² |
|---|---|---|---|
| LSTM | 0.5354 | 0.7880 | 0.9766 |
| Hybrid LSTM–GAT | 1.3447 | 2.0056 | 0.9309 |
The standalone LSTM is the better univariate student. The hybrid is the better cartographer. We refuse to hide the first fact to advertise the second.
Adaptive routing
10–25%
Reduction in experienced travel time
Dynamic
Routes rewrite as predicted speeds move
Load
Traffic redistributes instead of stampeding
Policy
Learned — not a one-shot shortest path
Methods
Topology comes from OpenStreetMap via OSMnx around Palayam: the capital’s drive graph, Kerala’s hills and coastal approach. Simulation covers the same 5-minute cadence of speed, flow, congestion, weather, and density.
Thiruvananthapuram is a stress test dressed as a capital. Secretariat pulse. NH 66 pins the west. Coastal roads lie about capacity. A model that only works on a grid with four neighbors has not met a city.
Windows are causal. Features are normalized with statistics that cannot see the test week. GAT attention is computed on the true adjacency, not a fully connected fantasy. YOLO classes are limited to the vehicle set the municipal question actually asks.
We do not claim a world-best MAE. We do not claim live Kerala CCTV on every pole. We do not claim PPO is a replacement for signal timing, transit, or politics. We claim a measured loop from forecast to path on a real industrial graph, and a product face that does not insult the reader with gradients.
The Flask application, edge CSV, preprocessing utilities, and hybrid model code live beside this site in traffic-app. This Next.js 16 surface is the public narrative and the interactive instrument. Both should be able to point at the same numbers without blinking.
Future work is written in the same voice as the present: live APIs, multi-camera fusion, signal policies, Jetson-class edge, fuel as a second reward, multimodal legs. None of that is a promise. It is a queue.
0.9766 on LSTM means the series is learnable. 0.9309 on hybrid means the spatial tax is real and still high. If your procurement sheet only has one cell, put the routing delta in it. If it has two, put both. If it has a paragraph, you are already on this page.
The public Predict / Map / Influence tools currently run a deterministic demo predictor so anyone can explore Thiruvananthapuram without a GPU. The tables above are the measured lab story for the hybrid stack. When you demo the site, say both sentences: the graph is real OSM; the interactive speeds are scenario demos; the published MAE/R² are from the trained models on the research track.
Technology (stack) → this page (scores) → Predict (touch the city) → Map (see a route) → Influence (see links) → Detect (optional camera). That order matches how the loop was designed.
How to cite these results
Report MAE, RMSE, R² for LSTM vs hybrid on the Thiruvananthapuram research split.
Report PPO travel-time reduction vs Dijkstra as a range (10–25%) with the setting named.
Describe demo speeds as scenario forecasts on OSM — not as the same evaluation set unless you re-ran weights.
Cite OSMnx-derived topology for the capital; do not imply every Kerala district is loaded.
YOLO counts on uploads; do not imply city-wide live CCTV coverage.
Always include: no claim of world-best MAE; no replacement for signal timing or transit policy.
Common questions
Spatial models spend capacity on neighbors. Univariate fit can look better while routes stay naive. We publish both.
Yes — traffic-app holds Flask, edges, and web UI. Research weights and full training may need the ML track scripts.
Minutes saved on a corridor people actually drive — not only a prettier residual plot.