The accurate process modelling in natural gas industry is beneficial for enterprises to increase potential profit and realize sustainable production. The integration of industrial data and first-principles computation is effective in hybrid process modelling. However, how to enhance their accuracy, robustness, and interpretability remains a challenge. Herein, a general framework integrating artificial intelligence algorithms and process mechanisms for natural gas desulfurization processes is proposed for accurate hybrid modelling. Firstly, the modelling is driven by the biased sulfur content target data generated by the mechanism model. Secondly, the general data information estimation of multivariate interaction and the independence test framework based on proxy data resampling are proposed based on the high-order information relationship. Then, based on Markov Blanket and causal feature selection principle, the compact variable recognition algorithm and the bias prediction model are further established. Finally, the industrial case of natural gas desulfurization is applied for method validation. On one hand, the prediction effect of the developed hybrid model (R 0.9453, MAE 0.0982, MSE 0.0168, MAPE 0.03018) performs better than that of the mechanism model. On the other hand, it is indicated by the target deviation modelling that the proposed method can not only effectively reduce the excessive variables into 11 compact variables, which proves the effectiveness and accuracy of the proposed method, with the coefficient of determination R2 reaching 0.9152. Moreover, the rationality and interpretability of the proposed method is validated combining the process knowledge and some interpretability techniques. The contribution of this work is to provide an accurate, robust and interpretable method involving both process mechanism and realistic production data for the modelling for the natural gas industry.

中文翻译：

---

天然气脱硫过程的混合建模方法：通过紧凑变量识别的耦合机制和数据建模

天然气行业准确的过程建模有利于企业增加潜在利润并实现可持续生产。工业数据和第一原理计算的集成在混合过程建模中是有效的。然而，如何提高其准确性、稳健性和可解释性仍然是一个挑战。本文提出了一种集成天然气脱硫过程人工智能算法和过程机制的通用框架，用于精确的混合建模。首先，建模由机制模型生成的有偏差的硫含量目标数据驱动。其次，基于高阶信息关系，提出了多元交互作用的通用数据信息估计和基于代理数据重采样的独立性检验框架。然后，基于马尔可夫毯和因果特征选择原理，进一步建立了紧凑变量识别算法和偏差预测模型。最后应用天然气脱硫工业案例进行方法验证。一方面，所开发的混合模型（R 0.9453，MAE 0.0982，MSE 0.0168，MAPE 0.03018）的预测效果优于机制模型。另一方面，通过目标偏差建模表明，该方法不仅能够有效地将过多变量减少为11个紧凑变量，证明了该方法的有效性和准确性，判定系数R2达到0.9152。此外，结合过程知识和一些可解释性技术，验证了所提出方法的合理性和可解释性。这项工作的贡献是提供一种准确、稳健和可解释的方法，涉及天然气行业建模的过程机制和实际生产数据。