A hybrid framework of first-principles model and machine learning for optimizing control parameters in chemical processes

Abstract

Artificial intelligence (AI) has recently gained prominence for addressing complex problems in chemical plants. Despite its enthusiastic attention, the industrial application of AI is limited due to a lack of both reliability and diversity in its operation data at the plant scale. To address this issue, a framework that integrates the machine learning (ML) model and first-principles approach is proposed herein. The performance of the proposed framework is demonstrated by its application to the control system of the liquefied natural gas fuel gas supply system. In this framework, commercial simulation software was used to implement a high-accuracy first-principles model using operation data. Thereafter, a wide range of data was generated that cannot be obtained in an industrial plant. The generated data was fed to the ML model that predicted the control performance with variations of the control parameters. The ML model, built with high-quality data, can predict the control performance with high accuracy. The optimal control parameters were quickly found using the ML model, thereby improving the control performance. This study presents a solution that can overcome the limitations of using an ML model alone by exploiting the advantages of both the first-principles and data-driven approaches at the plant scale.