---
library_name: pytorch
pipeline_tag: tabular-regression
tags:
  - machine-learning
  - deep-learning
  - air-quality
  - aqi-prediction
  - xgboost
  - lstm
  - pytorch
  - time-series
  - regression
---
# AQI Eco-Nav Models

Machine learning and deep learning models for **Air Quality Index (AQI) prediction** using air-quality and temporal features.

This repository contains two independently trained models:

- **XGBoost** — traditional machine learning model for tabular AQI prediction
- **BiLSTM** — deep learning model for time-series AQI prediction

Both models were trained using air-quality data from India.

## Models

### 1. XGBoost — Machine Learning

**Model:** XGBoost Regressor

**Task:** AQI regression

**Model file:**

```text
ml/xgb_aqi_model.pkl
```

#### Performance

| Metric | Score |
|---|---:|
| MAE | 20.28 |
| RMSE | 41.14 |
| R² | 0.9076 |

Training samples: **19,880**

Test samples: **4,970**

#### Preprocessing

```text
ml/
├── xgb_aqi_model.pkl
├── city_encoder.pkl
├── feature_medians.pkl
└── model_meta.json
```

---

### 2. BiLSTM — Deep Learning

**Model:** 2-layer Bidirectional LSTM

**Framework:** PyTorch

**Task:** Time-series AQI regression

**Architecture:**

```text
2×BiLSTM(hidden=128, bidirectional=True)
        ↓
FC(256)
        ↓
AQI Output
```

**Sequence length:** 7

**Model file:**

```text
dl/lstm_aqi_model.pt
```

#### Performance

| Metric | Score |
|---|---:|
| MAE | 17.21 |
| RMSE | 25.39 |
| R² | 0.8102 |

Training samples: **19,734**

Test samples: **2,468**

#### Preprocessing

```text
dl/
├── lstm_aqi_model.pt
├── dl_scaler.pkl
├── dl_city_encoder.pkl
├── dl_feature_medians.pkl
└── dl_model_meta.json
```

## Model Comparison

| Model | Type | MAE | RMSE | R² |
|---|---|---:|---:|---:|
| XGBoost | Machine Learning | 20.28 | 41.14 | **0.9076** |
| BiLSTM | Deep Learning | **17.21** | **25.39** | 0.8102 |

The **BiLSTM achieves lower MAE and RMSE**, while the **XGBoost model achieves a higher R²** on its respective test evaluation.

Because the models use different evaluation setups, the metrics should not be interpreted as a direct apples-to-apples benchmark.

## Input Features

Both models use the following 16 features:

```text
PM2.5
PM10
NO
NO2
NOx
NH3
CO
SO2
O3
Benzene
Toluene
Xylene
Month
DayOfYear
DayOfWeek
City_Enc
```

## Supported Cities

The models were trained using data from these cities:

```text
Ahmedabad
Aizawl
Amaravati
Amritsar
Bengaluru
Bhopal
Brajrajnagar
Chandigarh
Chennai
Coimbatore
Delhi
Ernakulam
Gurugram
Guwahati
Hyderabad
Jaipur
Jorapokhar
Kochi
Kolkata
Lucknow
Mumbai
Patna
Shillong
Talcher
Thiruvananthapuram
Visakhapatnam
```

## Dataset

The models were trained using air-quality data from India.

Dataset source:

https://www.kaggle.com/datasets/rohanrao/air-quality-data-in-india

## Model Files

```text
aqi-eco-nav-models/
│
├── ml/
│   ├── xgb_aqi_model.pkl
│   ├── city_encoder.pkl
│   ├── feature_medians.pkl
│   └── model_meta.json
│
└── dl/
    ├── lstm_aqi_model.pt
    ├── dl_scaler.pkl
    ├── dl_city_encoder.pkl
    ├── dl_feature_medians.pkl
    └── dl_model_meta.json
```

## Technologies

### Machine Learning

- Python
- XGBoost
- Scikit-learn
- Pandas
- NumPy
- Joblib

### Deep Learning

- Python
- PyTorch
- Bidirectional LSTM
- Scikit-learn
- NumPy

## Intended Use

These models are designed for:

- AQI prediction
- Air-quality analysis
- Environmental monitoring
- Machine learning research
- Time-series forecasting experiments
- Educational projects

## Disclaimer

These models are intended for educational and research purposes. Predictions may vary when applied to real-world air-quality conditions outside the training data.