Energy Forecasting for European Steel Manufacturer

Background and industry context

Energy-intensive industries face volatile electricity markets where accurate consumption forecasting can mean the difference between millions in savings or losses. Steel manufacturing, with its massive furnaces consuming megawatts of power, requires precise day-ahead and week-ahead predictions to optimize electricity procurement and trading positions.

Intro problem

A leading European steel manufacturer needed to predict energy consumption across their facilities to optimize electricity trading strategies. Their existing approach relied on production schedules that were rarely followed in practice, leading to significant discrepancies between planned and actual consumption. This uncertainty forced conservative trading positions, leaving money on the table in a market where even 5% efficiency gains translate to millions annually.

Result

In just a week, I delivered a proof-of-concept that surpassed expectations:
97% accuracy for day-ahead consumption forecasting and ~97% accuracy (3% MAE) for week-ahead predictions (using only historical data). The solution proved very effective, and the client initiated internal rollout to additional manufacturing locations, potentially unlocking millions of euro in annual trading efficiency gains.

Solution Details

The breakthrough came from recognizing that while production schedules existed, they were aspirational rather than operational. My analysis revealed predictable patterns in how schedules deviated from reality: holidays were never worked despite being scheduled, Friday afternoons saw early shutdowns, and maintenance windows followed unwritten but consistent rules.

I developed a hybrid approach combining rule-based logic with machine learning to automatically adjust unrealistic schedules before feeding them into the forecasting model. This single feature improved prediction accuracy by 20%, transforming mediocre results into exceptional performance.

The solution compared multiple approaches (ARIMA, Prophet, LightGBM) before settling on a neural network architecture that best captured the complex temporal patterns while remaining robust to anomalies.

My involvement

• Led a one-day workshop with key stakeholders to define success metrics, project scope, and requirements
• One week independently developing the PoC
• Performed deep-dive data cleaning and analysis uncovering the critical insight about schedule reliability
• Delivered a production-ready model that the client’s team could immediately scale

Technologies used

• Darts – state-of-the-art time series forecasting library for model comparison
• Python & Pandas – data preprocessing and handling years of hourly consumption data
• Time series analysis – handling seasonality, trends, and COVID-period anomalies
• Feature engineering – developing the schedule adjustment algorithm combining rules and ML
• Docker - for secure model deployment in the client’s network