Improving computational efficiency of machine learning modeling of nonlinear processes using sensitivity analysis and active learning

Abstract

In this work, we develop a model reduction method using sensitivity analysis and active learning to improve the computational efficiency of machine learning modeling of nonlinear processes. Specifically, sensitivity analysis is first used to identify important connections between model outputs and inputs. Subsequently, active learning is used to enrich the training set by iteratively identifying the training data that most efficiently improve model performance. Reduced-order recurrent neural networks (RNN) using the important input features obtained from sensitivity analysis are developed to approximate the nonlinear system, and are incorporated within model predictive control (MPC) to stabilize the nonlinear system at the steady-state. Finally, the effectiveness of the proposed machine learning modeling approach using sensitivity analysis and active learning and machine-learning-based predictive control scheme are demonstrated using a reactor-reactor-separator process example.

Keywords

• Model predictive control
• Machine learning
• Recurrent neural networks
• Feature selection
• Sensitivity analysis
• Active learning