Real-Time Optimization of Vertical Roller Mills Using XGBoost Prediction and Q-Learning Control

Vertical roller mills are essential for energy-intensive grinding in cement, minerals, and metallurgy industries, consuming up to 50% of plant electricity and frequently experiencing operational instabilities (including excessive vibration and main motor current fluctuations) that drive unplanned downtime, increased wear, and reduced throughput. Despite their importance, real-time autonomous optimization remains challenging due to the nonlinear interactions among grinding pressure, feed rate, separator speed, and aerodynamic factors, which limit traditional control strategies under varying loads. This paper presents a real-time operational optimization system for large-scale vertical roller mills using big industrial data and artificial intelligence (AI). From a 5400 kW Loesche LM56.4 mill, 2,764,800 samples were collected at 1 Hz over 32 days of continuous production. A systematic pipeline was developed: quartile-based outlier-robust cleaning; domain-informed feature engineering including Total Current; Random Forest (RF) permutation importance selection of the top 15 parameters; and Extreme Gradient Boosting (XGBoost) regression models with hyperparameters tuned by Tree-structured Parzen Estimator (TPE) Bayesian optimization. The resulting models achieved strong predictive performance, Mean Absolute Percentage Error (MAPE) of 1.3% (95% CI: 1.1%–1.5%) for main motor current (R² = 0.9997) and 5.8% (95% CI: 5.3%–6.3%) for shell vibration (R² = 0.9717), representing reductions of 89% and 59%, respectively, relative to the Long Short-Term Memory (LSTM) baseline. These surrogates were embedded into a tabular Q-learning Reinforcement Learning (RL) agent that autonomously adjusts feed rate, grinding pressure, separator speed, and exhaust damper position via a discrete action space and multi-objective reward function, communicating with the Distributed Control System (DCS) via Open Platform Communications Unified Architecture (OPC-UA). Closed-loop evaluation yielded simultaneous reductions of 6.0% in peak current (181.92 → 170.04 A) and 9.4% in peak vibration (5.51 → 4.99 mm/s) while maintaining throughput. A PyQt5-based graphical interface enabling real-time monitoring, predictive alerts, and automatic DCS write-back was deployed and operated stably for two weeks.