Machine learning-based predictive control using noisy data: evaluating performance and robustness via a large-scale process simulator

Abstract

Machine learning modeling of chemical processes using noisy data is a practically challenging task due to the occurrence of overfitting during learning. In this work, we propose a dropout method and a co-teaching learning algorithm that develop long short-term memory (LSTM) neural networks to capture the ground truth (i.e., underlying process dynamics) from noisy data. To evaluate the performance and robustness of the proposed modeling approaches, we consider an industrial chemical reactor example and use a large-scale process simulator, Aspen Plus Dynamics that does not employ assumptions on reactor properties typically made in the derivation of first-principles models, to generate process operational data that are corrupted by sensor noise which is determined using industrial data. The dropout method is first utilized to reduce the overfitting of LSTM models to noisy data. Then, another approach termed co-teaching method is used to train LSTM models with additional noise-free data generated from simulations of the reactor first-principles model that employs several standard modeling assumptions not made in the Aspen model. Through open-loop and closed-loop simulations, we demonstrate the improvement of model prediction accuracy and of the open- and closed-loop performances under model predictive controllers using dropout and co-teaching LSTM neural network models compared to the LSTM model developed from the standard training process from the noisy data.