7-Day Air Quality Index Forecasting with Conformal Prediction Intervals
| Field | Details |
|---|---|
| Name | Dhanush D |
| GitHub | github.com/Drdhx |
AURA-TFT is a novel hybrid deep learning architecture for multi-horizon urban AQI forecasting, applied to Chennai's Alandur monitoring station (2023–2025). It integrates five specialized components with Split Conformal Prediction for statistically rigorous uncertainty quantification.
| Metric | 1-Step | 7-Day Avg |
|---|---|---|
| MAE | 1.08 AQI units | 2.68 |
| RMSE | 1.45 | 3.76 |
| MAPE | 2.67% | 6.4% |
| R² | 0.9983 | 0.9879 |
| CP Coverage | 93.8% | 90.67% |
Raw Features (66)
│
▼
NeuralVSN ──► RFVSN ──► Multi-Head Attention ──► GRN ──► GBM Decoder ──► AQI Forecast
(ELU+GLU) (RF top-25) (n_heads=4, d=64) (skip+LN) (500 trees)
│
Split Conformal Prediction
(α=0.10 → 90% PI guarantee)
| Component | Description |
|---|---|
| NeuralVSN | ELU activation + GLU gating + softmax weights, d_hidden=64 |
| RFVSN | Random Forest importances → softmax top-25, per-horizon adaptive |
| MHA | Scaled dot-product, n_heads=4, d_model=64, LayerNorm + residual |
| GRN | ELU + GLU gating, skip connection, LayerNorm |
| GBM Decoder | 500 trees, lr=0.03, depth=4, subsample=0.75 |
| Conformal PI | Split CP, α=0.10, calibrated on 15% held-out validation set |
- Station: Alandur, Chennai (13.0°N, 80.2°E) — CPCB monitoring
- Period: January 2023 – December 2025 (3 years)
- Size: 26,281 hourly records, 66 features, 0 missing values
- Daily: 1,065 records after feature engineering (78 features)
- AQI Range: 18.0 – 171.3 (μ=79.2 ± σ=36.2)
- Pollutants: PM2.5, PM10, NO₂, SO₂, CO, O₃
- Meteorology: Temperature, Humidity, Wind Speed/Direction, Rainfall, Visibility
git clone https://github.com/Drdhx/AURA-TFT.git
cd AURA-TFT
pip install scikit-learn scipy seaborn matplotlib pandas numpyjupyter notebook AURA_TFT_CAQI.ipynbOr run the full pipeline directly:
# The notebook is self-contained — just set the CSV path and run all cells
df = pd.read_csv('chennai_aqi_complete_2023_2025.csv')AURA-TFT/
├── AURA_TFT_CAQI.ipynb # Main notebook (full pipeline)
├── chennai_aqi_complete_2023_2025.csv # Dataset
├── AURA_TFT_Presentation.pptx # presentation (12 slides)
├── results/
│ ├── viz_A_eda.png # EDA dashboard
│ ├── viz_B_overfit.png # Overfitting diagnostic
│ ├── viz_C_forecast.png # Forecast vs Actual + PI
│ ├── viz_D_scatter_qq.png # Scatter plot + Q-Q
│ ├── viz_E_horizon.png # 7-horizon performance
│ ├── viz_F_multistep.png # Multi-step fan chart
│ ├── viz_G_vsn.png # VSN feature weights
│ └── viz_H_comparison.png # Baseline comparison
└── README.md
78 features engineered from 66 raw variables:
- Cyclical encodings: sin/cos of day-of-year, day-of-week, month
- AQI lags: {1,2,3,4,5,7,10,14,21} days
- Rolling stats: mean, std, min, max over {3,7,14,21,30} day windows
- Momentum: first/7th differences, acceleration, 7-day range
- Meteorological lags: Temperature, Humidity, Wind, Rainfall, PM2.5, PM10, NO₂
- Interactions: temp×wind, PM2.5×NO₂, PM2.5×PM10, humidity×temp
Split CP provides distribution-free, finite-sample guarantees:
Ĉ(X) = [ŷ − q̂, ŷ + q̂]
where q̂ = ⌈(n+1)(1−α)⌉/n quantile of |y_i − ŷ_i| on calibration set
Guarantee: P(y_{n+1} ∈ Ĉ(X_{n+1})) ≥ 1 − α = 0.90
Achieved: 93.8% coverage with ±2.92 AQI interval width.
If you use this work, please cite:
@inproceedings{dhanush2025auratft,
title = {AURA-TFT: Adaptive Urban AQI Recurrent Transformer for 7-Day Air Quality Forecasting with Conformal Prediction Intervals},
author = {Dhanush D},
year = {2025},
note = {Independent Research, Chennai, India},
url = {https://github.com/Drdhx}
}MIT License — see LICENSE for details.