Abstract:

Methods for modeling the steam cracking process were reviewed, and the problem of data scarcity faced in industrial realities was described. Facing the massive small dataset modeling requirements in petrochemical industry, an ensemble transfer learning framework was proposed by making full use of the historical production data. First, basic deep learning models were established on a specific working condition with sufficient data. Then, transfer learning techniques were applied to the new working conditions with a small dataset. The process knowledge from the source domain was transferred to the target domain using parameter-based methods. Finally, ensemble learning was introduced to integrate the obtained transfer learning models, resulting in enhanced performance. The performance of entire modeling framework was found to be industrially acceptable on several practical cases, and further layer transferability analysis and SHapley Additive exPlanation (SHAP) feature importance analysis were implemented to provide a better understanding of the model. The results illustrated that the model trained by this method had good accuracy, stability, computational efficiency and interpretability to meet industrial requirements.

Key words: model, transfer learning, ensemble learning, algorithm, model interpretability, petroleum, prediction, neural networks

摘要：

回顾了蒸汽裂解过程建模的方法，阐述了工业实际情况中面临的数据匮乏问题。面对石油化工行业大量的小数据集建模需求，充分利用历史生产数据，提出了一种集成迁移学习框架。首先，利用充足的数据在特定工况下建立了基本的深度学习模型。然后，利用小数据集将迁移学习技术应用于新的工况，源域的专家知识通过基于参数的方法转移到目标领域。最后，引入集成学习来整合获得的迁移学习模型，从而提高性能。在几个实际案例上进行实践，研究了该模型框架的性能。为了更好地理解模型，还进一步实施了层可迁移性分析和SHapley Additive exPlanation（SHAP）特征重要性分析。结果说明该方法训练出的模型具有良好的准确性、稳定性、计算效率和可解释性，可以满足工业需求。

关键词: 模型, 迁移学习, 集成学习, 算法, 模型可解释性, 石油, 预测, 神经网络