Handling Noisy Data in Machine Learning Modeling and Predictive Control of Nonlinear Processes

Long short-term memory (LSTM) networks, as one type of recurrent neural networks, has been widely utilized to model nonlinear dynamic systems from time-series process operational data. This work focuses on LSTM modeling and predictive control of nonlinear processes using a noisy training data set, where the noise can stem from different sources, such as sensor variability and common plant variance. We first consider a dataset with Gaussian noise, and demonstrate that the standard LSTM network is able to capture the underlying (nominal) nonlinear process dynamic behavior. Then, we consider a noisy dataset from industrial operation (i.e., non-Gaussian noisy data), and demonstrate a poor training performance of the standard LSTM network despite its denoising capability for Gaussian noise. Therefore, to train an LSTM more efficiently with noisy data, we propose an LSTM network using Monte Carlo dropout method to reduce the overfitting to noisy data. The proposed dropout LSTM method is applied to a chemical process example with state measurements corrupted by industrial noise to demonstrate its improved prediction accuracy in both open- and closed-loop operation under a Lyapunov-based model predictive controller.

Automated summary

This research focuses on using LSTM to model and predict control nonlinear processes, comparing the performance of standard LSTM on datasets with Gaussian noise and noisy industrial datasets, and proposing a dropout LSTM method using Monte Carlo dropout to train LSTM more efficiently with noisy data.