Industrial Time-Series Forecasting using Residual Learning

Problem

Forecast cumulative raw material deliveries for Hydro ASA’s aluminum smelting operations.

Approach

• Residual learning framework (baseline + ML correction)
• LightGBM quantile regression (P20 asymmetric loss)
• 14 engineered domain-specific features
• Leakage-free time-series validation

Results

• 35.8% improvement over statistical baseline
• Kaggle score: 6227
• 1.2M+ predictions generated
• Execution time: <10 minutes

Tech Stack

Python, LightGBM, SHAP, Pandas, NumPy

How to Run

python ultrafast_forecast.py