A Comparative Study of Multi-Step Forecasting Algorithms for Real-Time Load Profiling in Smelting Plant Electricity Demand

The transition to deregulated energy markets has transformed electricity from a fixed cost into a strategic variable for power-intensive industries. Smelting plants, characterized by volatile, multi-megawatt load signatures, face acute exposure to price volatility and grid penalties, creating an urgent need for high-fidelity, multi-step load forecasting. This study addresses this need through a rigorous, empirical benchmark of six forecasting models, ARIMA, VAR, XGBoost, Random Forest, LSTM, and TCN, using high-resolution data from the Smart Energy Management Platform at Lanzhou Heqiao Min Metals Resources Co., Ltd.We move beyond aggregate accuracy to perform a horizon-degradation analysis, revealing a critical trade-off: deep learning models (LSTM) excel at minimizing typical error (MAE) over 60-minute horizons, but exhibit higher vulnerability to large errors (RMSE). In contrast, gradient boosting (XGBoost) demonstrates superior robustness, bounding worst-case scenarios. Furthermore, model interpretation identifies load imbalance factor and power quality metrics as the dominant predictive features, providing novel operational intelligence. Our findings establish that model selection is not a universal verdict but a strategic decision based on operational horizon and risk tolerance. This research provides a definitive framework for deploying forecasting as a core component of energy strategy, demand response, and grid resilience in the smelting industry.