Sustainability goals and compliance with ISO 50001 standards in industrial facilities require the development of high-accuracy energy forecasting models. However, operational dynamics on-site make it difficult to integrate external variables (e.g. production volume, weather conditions) into forecasting models in real time and without data loss. This study presents a practical, univariate forecasting framework that relies solely on historical electricity consumption data, without depending on complex external data. Daily electricity consumption data for 28 manufacturing and assembly buildings at a large-scale facility over a 6-year period (2020–2025) were dynamically aggregated based on their commissioning dates. During the data preprocessing stage, cumulative data gaps caused by communication interruptions were filled using the uniform distribution method, and outliers were corrected using IQR-based clipping. The predictive performance of ARIMA, Random Forest, XGBoost, and Long Short-Term Memory (LSTM) models was comparatively analyzed on the prepared dataset. The findings demonstrated that tree-based algorithms are superior in modeling sudden operational disruptions (such as shifts and holidays) in the industrial consumption profile. While ARIMA underperformed on nonlinear dynamics (83.04% test accuracy), LSTM—despite its stable performance (93.13%)—lagged behind other machine learning models on the tabulated dataset. Random Forest, supported by temporal feature engineering, emerged as the most successful predictor in the study, achieving a test accuracy rate of 94.10% and the lowest error metrics. Consequently, it has been confirmed that this proposed data-driven machine learning framework can be utilized as a highly reliable decision-support tool in energy budgeting, baseline creation, and planning processes for manufacturing facilities.

Keywords

Energy Forecasting, Time Series Analysis, Univariate Modeling, Machine Learning, Deep Learning