Machine learning-based energy consumption forecasting and optimization in heavy manufacturing

This study presents an integrated framework combining multi-horizon load forecasting (15-minute, 1-hour, and 24-hour) with load-shifting optimization for the DAEWOO Steel plant, explicitly eliminating temporal data leakage. Leakage-safe features derived strictly from historical data available at prediction time are used to benchmark Random Forest, XGBoost, LSTM, and a two-stage hybrid architecture. Calibrated P10-P90 prediction intervals via Conformalized Quantile Regression (CQR) quantify operational uncertainty, while 24-hour day-ahead forecasts are embedded into a Linear Programming model to optimize flexible load shifting for peak reduction and cost savings. The results indicate that errors scale with the forecast horizon (MAE from 2.690 kWh at 15 minutes to approximately8,000 kWh at 24 hours). The two-stage hybrid model excels at short horizons via regime classification, whereas the P50 quantile regression performs best at 24 hours. The findings confirm that input data availability and multi-horizon evaluation are critical for practical industrial energy management.