Comparative evaluation of several models for forecasting hourly electricity use in a steel plant.

The iron and steel industry belongs to the most electricity-intensive branches of manufacturing, and power expenses often account for a large portion of overall production costs. For this reason, reliable short-term forecasts of plant-level electricity consumption are essential for optimizing production planning, avoiding excessive demand charges, and supporting low-carbon operation. This study compares three representative approaches for hourly power-load prediction in a steel enterprise: eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM). Using one year of operating data from an integrated steel plant, we build a dataset of 8,760 hourly records and design a unified preprocessing procedure, including three-sigma outlier detection, time-indexed linear interpolation, and chronological division into training, validation, and test subsets. All three models are trained within a 24-hour sliding-window framework and assessed with RMSE, MAE, and MAPE. The results indicate that each model is capable of depicting the cyclical variation of the load, whereas BiLSTM provides the most accurate predictions, achieving the lowest errors in both absolute and relative terms. XGBoost demonstrates competitive performance and robust trend following. In contrast, the unidirectional LSTM exhibits larger relative errors, particularly during low-load periods. The findings underline the benefits of bidirectional recurrent structures for short-term electricity-consumption forecasting in steel plants and offer guidance for model selection in industrial energy-management practice.