Distillation Column Temperature Prediction

Machine learning pipeline for predicting the top temperature of a distillation column from weather-driven process data. The project combines historical weather loading, synthetic process-data generation, feature engineering with lag variables, and LightGBM regression.

What Is Inside

• Weather loader for Moscow historical weather data from Visual Crossing.
• Synthetic process-data generator based on simplified heat-transfer and fouling models.
• LightGBM training pipeline for T_top_C prediction.
• Correlation analysis and report figures.
• Pretrained LightGBM artifact and fitted scaler in models/.
• Aspen HYSYS simulation files in simulations/.

Repository Structure

.
├── data/                  # local datasets; generated/downloaded CSV files are not required in git
├── models/                # trained model artifacts
├── reports/figures/       # process scheme and report images
├── simulations/           # Aspen HYSYS source files
├── src/
│   ├── analysis/          # exploratory analysis scripts
│   ├── data/              # data collection and synthetic-data generation
│   └── modeling/          # model training pipeline
├── .env.example
├── requirements.txt
└── README.md

Setup

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

To download weather data, create .env or set the variable in your shell:

$env:VISUAL_CROSSING_API_KEY = "your_api_key"

Usage

Download weather data:

python src/data/fetch_weather.py --start 2023-03-25 --end 2025-08-31 --out data/weather_summer_2025.csv

Generate synthetic process data:

python src/data/generate_synthetic_data.py --weather data/weather_summer_2025.csv --out data/synthetic_dataset.csv

Train the LightGBM model and save figures:

python src/modeling/train_lightgbm.py --input data/synthetic_dataset.csv --model-dir models --figures-dir reports/figures

Build a correlation heatmap:

python src/analysis/correlation_heatmap.py --input data/synthetic_dataset.csv --out reports/figures/corr_heatmap.jpg

Model

The model predicts T_top_C, the top temperature of the column. The training script uses:

• z-score filtering for numeric outliers;
• lag features over a 24-step window;
• chronological train/test split;
• TimeSeriesSplit for cross-validation;
• grid search over LightGBM parameters;
• MAE, RMSE, and R2 as evaluation metrics.

The current trained artifacts are stored in:

• models/best_lightgbm_model.pkl
• models/scaler.pkl

Notes

Raw .docx drafts and generated training logs are intentionally excluded from the repository. API keys must be provided through environment variables, not committed to source control.