Integrating Machine Learning in Process Control with LSTMc: A Case Study in Batch Crystallization

Over the past three decades, the primary methods for managing complex chemical processes have predominantly relied on PI/PID controllers and model predictive controllers (MPCs). Despite their advancements, both approaches have limitations. PI/PID controllers require specific tuning for each scenario, while MPCs are computationally intensive. To address these challenges, we introduce the LSTM-controller (LSTMc), which is a model-independent, data-driven framework that leverages the robust time-series prediction capabilities of LSTM networks. The LSTMc predicts future control inputs by analyzing the evolution of the system's state and error dynamics, taking into account both the current and previous time steps. Our experiments, particularly in dextrose batch crystallization, have demonstrated the effectiveness of this approach. Impressively, the LSTMc achieves a set-point deviation of less than 2% in 10+ case studies we tested. These results position the LSTMc as a promising alternative for process control, as it adeptly adapts to changing process conditions and set-points, and offers efficient computation for optimal control inputs.