Transfer study for efficient and accurate modeling of natural gas desulfurization process

Abstract

Background

Accurate modeling of the natural gas desulfurization process enables enterprises to maintain stable production, optimize efficiency, improve product gas quality, and ensure compliance with environmental regulations. Considering the limitations of the availability of industrial data, machine learning models, mechanism models, and hybrid models integrating both may become inefficient or inaccurate.

Methods

To bridge this gap, a transfer learning-based modeling method for the natural gas desulfurization process was proposed. Firstly, a deep neural network model was developed to predict the hydrogen sulfide content in the product gas, based on mechanism-based calculations. Subsequently, a small dataset from the target scenario was utilized to fine-tune model parameters for accurate predictions under actual production conditions.

Significant Findings

The result demonstrates that the established model provides more stable and accurate predictions compared to traditional machine learning models, achieving over a 20 % reduction in prediction error while also enhancing modeling efficiency. Finally, the interpretability analysis of the proposed model reveals that the prediction capability of the model in actual production scenarios was rationally and effectively improved at a low computational cost through transfer learning. This work offers a novel paradigm for developing modeling methods tailored to the practical production processes of natural gas desulfurization.