Air Quality Index Prediction using Random Forest

Air pollution now shapes daily routines in Indian cities — school advisories, construction pauses and health warnings all hinge on the Air Quality Index. The Central Pollution Control Board publishes the National Air Quality Index, computed from pollutant sub-indices through a fixed breakpoint table, and streams station-level measurements through its portal alongside the global OpenAQ archive. But raw concentrations do not tell a citizen what to do, and advisories usually arrive after the smog does. Predicting the next hours' AQI from current station readings is a tabular time-series regression problem, and Random Forest handles it well: pollutant chemistry and dispersion are strongly non-linear, and forests tolerate the occasional sensor gaps real feeds contain. This project builds that forecaster end to end — ingestion, lagged features, per-pollutant forests, exact CPCB AQI math, and a dashboard with health advisories.

1. The scheduler polls CPCB/OpenAQ station feeds hourly; each record passes range validation, short gaps are linearly interpolated, and longer gaps are marked missing.
2. The feature builder materialises lagged pollutant levels (1/3/6/24 h), meteorological readings and calendar features into the training table.
3. One RandomForestRegressor per pollutant trains on the historical feature table to predict next-hour concentrations; forests are used because dispersion chemistry is non-linear.
4. Predicted concentrations pass through the exact CPCB breakpoint table: each pollutant gets a sub-index by interpolation, and the AQI is the maximum of the sub-indices.
5. The AQI maps to its category (Good → Severe), colour band and health-advisory text; the dashboard renders city cards, the hourly forecast with intervals, and the 7-day outlook.
6. Evaluation uses a strict time-based holdout — the last 14 days are never seen in training — and the notebook reports per-pollutant MAE/RMSE plus how often the predicted AQI category matches the actual one.

Tech stack:

• Python 3.10, scikit-learn (RandomForestRegressor, metrics)
• pandas, NumPy (time-series wrangling, lag features)
• requests (scheduled CPCB/OpenAQ feed polling)
• Matplotlib, Seaborn (forecast-vs-actual, error analysis)
• Jupyter notebook (buyer-run training and evaluation)
• Flask demo app (live board, outlook, AQI calculator)

• [Hourly station ingestion] Polls public CPCB/OpenAQ feeds on a schedule, validates every record against per-pollutant ranges, and interpolates only short gaps (≤3 h); longer gaps are flagged, never fabricated.
• [Lagged + meteorological features] Current and 1/3/6/24-hour lagged pollutant levels plus temperature, humidity, wind speed and direction, hour-of-day and day-of-week feed every prediction.
• [One forest per pollutant] A dedicated RandomForestRegressor (200 trees) predicts the next-hour concentration of PM2.5, PM10, NO₂, SO₂, O₃ and CO, since each pollutant disperses differently.
• [Exact CPCB NAQI computation] Sub-indices by linear interpolation between official breakpoints, AQI as the maximum sub-index, category and colour from the 0–500 bands — the same math the board itself uses.
• [Live board + 7-day outlook] City cards with dominant pollutant, an hourly next-day forecast with tree-vote prediction intervals, and a 7-day trend view.
• [AQI calculator] Enter pollutant concentrations manually and watch the identical breakpoint formula produce the AQI, category and health advisory — ideal for viva demonstrations.
• [Time-based holdout evaluation] The notebook trains on past weeks and tests on the most recent 14 days, reporting MAE, RMSE per pollutant and category-hit-rate for the final AQI.

• Complete source code (ingestion, features, training, AQI math, dashboard)
• Jupyter training and evaluation notebook (buyer-run: train, time-holdout, metrics, plots)
• Serialized per-pollutant forest models from the reference training run
• Project report PDF (background, CPCB methodology, evaluation, limitations)
• PPT presentation for final review
• Viva Q&A preparation document (lag features, breakpoint math, time-based validation)
• Setup guide (environment, data access, scheduler, running the demo)

• Forecast skill decays with horizon: next-hour predictions are strong, but the multi-day outlook is indicative — the report shows this decay curve honestly.
• Episodes absent from training data (dust storms, intense stubble-burning weeks) are out-of-distribution events; the system flags them rather than pretending confidence.
• Station coverage is uneven — forecasts exist only where public stations report; the system does not invent data for unmonitored areas.
• The AQI formula implementation is exact, but it is only as current as the CPCB breakpoint table it encodes; any future CPCB revision needs a table update.
• This is an advisory and academic prototype, not an official forecasting service; health advisories follow CPCB wording and do not replace medical advice.