Digital Twin-Based Optimization of Energy Efficiency in Industrial Production Systems: A Simulation and Machine Learning Framework

Abstract

Factories draw heavily on energy supplies, leaving manufacturers with the difficult task of cutting power use without hurting output. We built a simulation-based digital twin of an industrial production system to forecast power usage and production efficiency, as well as test different operational scenarios. A full year of synthetic hourly operating figures was generated, capturing equipment load, machine utilization, run time, ambient temperature, output rates, and energy use. To forecast efficiency and energy demand, we trained Random Forest, XGBoost, and Artificial Neural Network models, then judged their accuracy using Mean Absolute Error, Root Mean Square Error, and the coefficient of determination. Across both targets, XGBoost yielded the lowest error rates alongside the highest coefficient of determination. We then integrated this top model into the digital twin, optimizing operating scenarios against a set floor for production efficiency. Pairing batch consolidation with balanced equipment loads trimmed simulated power draw by 17.1% and cut energy per finished item by 19.1% with a minor boost to efficiency; by contrast, an aggressive load-slashing plan saved more power but missed the efficiency floor. These results demonstrate that combining digital twins with machine learning helps operators make better, energy-conscious choices on the shop floor. We outline what these findings mean for plant managers, energy engineers, and Industry 4.0 specialists, while pointing to future research using data collected from operating plants.