Forecasting
Air Pollution Forecasting
AirCast is an **air pollution forecasting system** using LSTM (Long Short-Term Memory) neural network.
Year :
2025
Industry :
E-commerce
ShopEase
Project Duration :
3 weeks
AirCast — Project Detail
AirCast is an **air pollution forecasting system** using LSTM (Long Short-Term Memory) neural network. The system analyzes historical weather data (temperature, humidity, wind speed) to predict air pollution levels in the next hour. Everything is packaged in an interactive dashboard that can be used directly.
**Tech Stack**: Python 3.12 · TensorFlow/Keras · Streamlit · Plotly · Scikit-learn
Why Use LSTM?
Air pollution prediction = **time series** problem (sequential data based on time). Regular neural networks (Dense/MLP) cannot "remember" time sequences. RNNs can, but suffer from **vanishing gradient** — fail to learn long-term patterns.
LSTM solves this with 3 gate mechanisms:
| Gate | Function |
|------|----------|
| **Forget Gate** | Remove irrelevant old information |
| **Input Gate** | Keep important new information |
| **Output Gate** | Determine output based on cell state |
Dashboard (app.py) Interactive dashboard built with **Streamlit + Plotly**:
### Sidebar — All Configuration in One Place
- Data Samples (1000-10000)
- Lookback Window (12-72 hours)
- Training Epochs (5-100)
- LSTM Units (16-128)
- Dropout Rate (0.0-0.5)
- Early Stopping toggle + patience
- Train/Test split ratio
### Step 1: Data Exploration (4 tabs)
- **Time Series** — Plot all 4 variables together (interactive, zoomable)
- **Correlations** — Heatmap of feature correlations
- **Distributions** — Histogram for each variable
- **Raw Data** — Data table + descriptive statistics
### Step 2: Model Training
- Real-time progress bar (scaling → windowing → training → predicting)
- Expandable architecture details (layer table + parameters)
### Step 3: Results
- **6 Metric Cards** — MSE, MAE, RMSE, R², Epochs Run, Training Time
- **Learning Curve** — Graph of train loss vs val loss (with best epoch marker)
- **Forecast vs Actual** — Comparison of predictions with actual values
- **Residual Analysis** — Residual bar chart + error distribution histogram
### Step 4: Export
- Download trained model (.h5)
- Download predictions (.csv) — Actual vs Predicted vs Residual
- Download metrics (.csv) — All metrics + training info
### pollution_forecast.py — Main Functions
| Function | Input | Output |
|----------|-------|--------|
| `generate_synthetic_data(n)` | number of samples | DataFrame 4 columns |
| `create_sequences(data, seq)` | scaled array, window size | X, y arrays |
| `train_model(...)` | train/test data, hyperparams | model, history |
| `compute_metrics(y_true, y_pred)` | actual, predicted | dict {MSE, MAE, RMSE, R²} |
More Projects
Forecasting
Air Pollution Forecasting
AirCast is an **air pollution forecasting system** using LSTM (Long Short-Term Memory) neural network.
Year :
2025
Industry :
E-commerce
ShopEase
Project Duration :
3 weeks
AirCast — Project Detail
AirCast is an **air pollution forecasting system** using LSTM (Long Short-Term Memory) neural network. The system analyzes historical weather data (temperature, humidity, wind speed) to predict air pollution levels in the next hour. Everything is packaged in an interactive dashboard that can be used directly.
**Tech Stack**: Python 3.12 · TensorFlow/Keras · Streamlit · Plotly · Scikit-learn
Why Use LSTM?
Air pollution prediction = **time series** problem (sequential data based on time). Regular neural networks (Dense/MLP) cannot "remember" time sequences. RNNs can, but suffer from **vanishing gradient** — fail to learn long-term patterns.
LSTM solves this with 3 gate mechanisms:
| Gate | Function |
|------|----------|
| **Forget Gate** | Remove irrelevant old information |
| **Input Gate** | Keep important new information |
| **Output Gate** | Determine output based on cell state |
Dashboard (app.py) Interactive dashboard built with **Streamlit + Plotly**:
### Sidebar — All Configuration in One Place
- Data Samples (1000-10000)
- Lookback Window (12-72 hours)
- Training Epochs (5-100)
- LSTM Units (16-128)
- Dropout Rate (0.0-0.5)
- Early Stopping toggle + patience
- Train/Test split ratio
### Step 1: Data Exploration (4 tabs)
- **Time Series** — Plot all 4 variables together (interactive, zoomable)
- **Correlations** — Heatmap of feature correlations
- **Distributions** — Histogram for each variable
- **Raw Data** — Data table + descriptive statistics
### Step 2: Model Training
- Real-time progress bar (scaling → windowing → training → predicting)
- Expandable architecture details (layer table + parameters)
### Step 3: Results
- **6 Metric Cards** — MSE, MAE, RMSE, R², Epochs Run, Training Time
- **Learning Curve** — Graph of train loss vs val loss (with best epoch marker)
- **Forecast vs Actual** — Comparison of predictions with actual values
- **Residual Analysis** — Residual bar chart + error distribution histogram
### Step 4: Export
- Download trained model (.h5)
- Download predictions (.csv) — Actual vs Predicted vs Residual
- Download metrics (.csv) — All metrics + training info
### pollution_forecast.py — Main Functions
| Function | Input | Output |
|----------|-------|--------|
| `generate_synthetic_data(n)` | number of samples | DataFrame 4 columns |
| `create_sequences(data, seq)` | scaled array, window size | X, y arrays |
| `train_model(...)` | train/test data, hyperparams | model, history |
| `compute_metrics(y_true, y_pred)` | actual, predicted | dict {MSE, MAE, RMSE, R²} |
More Projects
Forecasting
Air Pollution Forecasting
AirCast is an **air pollution forecasting system** using LSTM (Long Short-Term Memory) neural network.
Year :
2025
Industry :
E-commerce
ShopEase
Project Duration :
3 weeks
AirCast — Project Detail
AirCast is an **air pollution forecasting system** using LSTM (Long Short-Term Memory) neural network. The system analyzes historical weather data (temperature, humidity, wind speed) to predict air pollution levels in the next hour. Everything is packaged in an interactive dashboard that can be used directly.
**Tech Stack**: Python 3.12 · TensorFlow/Keras · Streamlit · Plotly · Scikit-learn
Why Use LSTM?
Air pollution prediction = **time series** problem (sequential data based on time). Regular neural networks (Dense/MLP) cannot "remember" time sequences. RNNs can, but suffer from **vanishing gradient** — fail to learn long-term patterns.
LSTM solves this with 3 gate mechanisms:
| Gate | Function |
|------|----------|
| **Forget Gate** | Remove irrelevant old information |
| **Input Gate** | Keep important new information |
| **Output Gate** | Determine output based on cell state |
Dashboard (app.py) Interactive dashboard built with **Streamlit + Plotly**:
### Sidebar — All Configuration in One Place
- Data Samples (1000-10000)
- Lookback Window (12-72 hours)
- Training Epochs (5-100)
- LSTM Units (16-128)
- Dropout Rate (0.0-0.5)
- Early Stopping toggle + patience
- Train/Test split ratio
### Step 1: Data Exploration (4 tabs)
- **Time Series** — Plot all 4 variables together (interactive, zoomable)
- **Correlations** — Heatmap of feature correlations
- **Distributions** — Histogram for each variable
- **Raw Data** — Data table + descriptive statistics
### Step 2: Model Training
- Real-time progress bar (scaling → windowing → training → predicting)
- Expandable architecture details (layer table + parameters)
### Step 3: Results
- **6 Metric Cards** — MSE, MAE, RMSE, R², Epochs Run, Training Time
- **Learning Curve** — Graph of train loss vs val loss (with best epoch marker)
- **Forecast vs Actual** — Comparison of predictions with actual values
- **Residual Analysis** — Residual bar chart + error distribution histogram
### Step 4: Export
- Download trained model (.h5)
- Download predictions (.csv) — Actual vs Predicted vs Residual
- Download metrics (.csv) — All metrics + training info
### pollution_forecast.py — Main Functions
| Function | Input | Output |
|----------|-------|--------|
| `generate_synthetic_data(n)` | number of samples | DataFrame 4 columns |
| `create_sequences(data, seq)` | scaled array, window size | X, y arrays |
| `train_model(...)` | train/test data, hyperparams | model, history |
| `compute_metrics(y_true, y_pred)` | actual, predicted | dict {MSE, MAE, RMSE, R²} |