AI Model & Forecasting
Ensemble Model
Three models combined with dynamic Inverse-MAE weighting:
| Model | Features | Min Weight |
|---|---|---|
| Random Forest | Lag + rolling + weather + cyclical | 30% (floor) |
| ARIMA | Auto-order via AIC grid p∈[0,3] q∈[0,2] | Dynamic |
| XGBoost Lite | Lag 1/2/3/6/12h — trained at runtime | Dynamic |
Weighting Formula
w_i = (1 / MAE_i) / Σ(1 / MAE_j) with w_RF ≥ 0.30
PM2.5_final = w_RF × pred_RF + w_ARIMA × pred_ARIMA + w_XGB × pred_XGB
Weights recompute every 60-minute cycle based on one-step-ahead MAE against the last 6 hours of actual observations.
Delta clamping: MAX_HOURLY_DELTA = 12.0 µg/m³ prevents implausible step-wise jumps in multi-step iterative forecasts.
Feature Engineering
| Group | Features |
|---|---|
| Cyclical time | sin/cos(hour), sin/cos(day_of_week), sin/cos(month) |
| Lag features | pm25_lag_1h, _2h, _3h, _6h, _12h, _24h |
| Rolling stats | pm25_roll_mean_3h, _6h, _24h |
| Diff features | pm25_diff_1h, _3h |
| Weather | temperature, humidity, wind_speed, wind_direction, pressure, CO, NO₂ |
Total: 157 features
Training Results
Training data: 01/01/2024 → 02/04/2026 (19,752 hourly points, TPHCM)
Cross-Validation (TimeSeriesSplit — 5 folds)
| Model | CV MAE | CV RMSE | CV R² |
|---|---|---|---|
| ExtraTrees | 1.167 | 2.210 | 0.9693 |
| XGBoost | 1.503 | 2.771 | 0.9533 |
| Random Forest | 1.777 | 2.957 | 0.9397 |
| Baseline (last hour) | 2.895 | 4.552 | 0.8809 |
Test Set (3,951 points — unseen)
| Model | MAE (µg/m³) | RMSE (µg/m³) | R² |
|---|---|---|---|
| ExtraTrees | 0.292 | 0.564 | 0.9982 |
| XGBoost | 0.265 | 0.509 | 0.9985 |
| Random Forest | 0.283 | 0.621 | 0.9978 |
All AI models outperform the naive baseline by 6–7× on RMSE.
Adding / Updating a Model
- Train in Google Colab (see
src/airsafenet_ai/notebooks/) - Export
model.pkl+metadata.json - Place in
src/airsafenet_ai/models/ - Restart the AI server or call
POST /admin/compute - Verify via
GET /healthandGET /model/info
See MODEL_VERSIONING.md for naming conventions.
Anomaly Detection
| Parameter | Value |
|---|---|
| Spike threshold | +20 µg/m³ within 6-hour lookback |
| Cooldown | 2 hours between alerts |
| XAI method | Feature importance × delta → top 3 factors |
| Storage | data/anomaly_log.json |
| Alert trigger | POST to .NET /api/notification/check-and-alert |