Objective  Accurate modeling of the natural gas desulfurization process is helpful for natural gas enterprises to achieve efficient and stable production, improve product gas quality, increase economic benefits, and ensure that exhaust emissions meet standards. Although the existing modeling methods based on artificial intelligence (AI) have achieved notable advancements, they typically focus on the prediction of a single production objective and cannot cope with the diverse application requirements, which limit their application in industrial devices.

Method  Using multi-task learning (MTL) technology, the intrinsic correlations among production variables in the production data of natural gas desulfurization processes were fully explored. By incorporating Savitzky-Golay filtering for noise reduction, and feature dimensionality reduction through Pearson correlation analysis and Random Forest algorithm, key production features of the desulfurization process were accurately identified. Based on this, using the MTL method, the multi-objective prediction tasks of total sulfur content and hydrogen sulfide content in the product gas were accomplished simultaneously.

Result Compared to traditional machine learning methods, the MTL model demonstrated significant advantages in prediction accuracy and prediction stability. The determination coefficient mean value ( \overlineR^2 ) of total sulfur mass concentration and hydrogen sulfide mass concentration in product gas were 0.980 and 0.972, respectively. The mean squared error mean value ( \overline\mathitS_\mathrmME ) were 0.127 (mg/m³)² and 0.008 (mg/m³)², respectively, and the prediction error was reduced by more than 50%. The assessment of input variable importance using Shapley additive explanations (SHAP) method revealed that the MTL model could accurately identify key operational variables. Additionally, the importance ranking of key variables and their impact patterns on prediction objects were consistent in the two prediction tasks, indicating that the model has advantages in accurately capturing the synergistic effects among multiple tasks.

Conclusion  The proposed modeling method not only enables accurate prediction of multiple production objectives, enhances the model's robustness and interpretability, but also reduces modeling computational costs, thereby increasing its practicality and adaptability in industrial applications, which can provide a reference for the development of modeling methods suitable for the natural gas desulfurization process.