Machine learning meets process control: Unveiling the potential of LSTMc

Abstract

In the past three decades, proportional-integral/PI-differential (PI/PID) controllers and model predictive controller (MPCs) have predominantly governed complex chemical process control. Despite their advancements, these approaches have limitations, with PI/PID controllers requiring scenario-specific tuning and MPC being computationally demanding. To tackle these issues, we introduce the long-short-term-memory (LSTM)-controller (LSTMc), a model-free, data-driven framework leveraging LSTM networks' robust time-series prediction capabilities. The LSTMc predicts subsequent manipulated inputs by evaluating state evolution and error dynamics from both the current and previous time-steps, which proved effective in our dextrose batch crystallization case study. Remarkably, the (Formula presented.) achieves less than 2% set-point deviation, three times better than MPCs, and retains robustness even with 10%–15% sensor noise. With these results, LSTMc emerges as a promising alternative for process control, adeptly adjusting to changing process conditions and set-points, providing efficient computation for an optimal input profile, and effectively filtering out common industrial process noise.