In the refinery process, a vast amount of data is generated in daily production. How to make full use of these data to improve the simulation's accuracy is crucial to enhancing the refinery operating level. In this paper, a novel deep learning framework integrating the self-organizing map (SOM) and the convolutional neural network (CNN) is developed for modeling the industrial hydrocracking process. The SOM is used to map input variables into two-dimensional maps to extract process features. Then, these maps are fed into the CNN to predict the outputs of the hydrocracking process. The SOM adopted is free of training, which reduces the computational complexity, simplifies the application, and improves the prediction accuracy. Practical guidance on the application of the proposed framework is provided by comparing and analyzing different structures and parameters. Finally, an online modeling scheme is developed and applied in an actual hydrocracking process. Experimental results demonstrate that the proposed framework has great performance in modeling the hydrocracking process and provides a good reference for process optimization.

Hybrid modeling, aiming to integrate the advantages of both first-principles models and data-driven models, is an important technology for refinery process simulation and optimization. The commonly used hybrid models include series models, parallel models, and series−parallel models. Many studies have reported the operational optimization results based on these models. However, it is unknown whether the results obtained based on these models are consistent with the actual plant optimal operation. Moreover, the hybrid models that have been used in operational optimization are of traditional structures and cannot ensure competent performance. To fill these gaps, we investigate the modeling and operational optimization of a typical refinery unit hydrocracking unit. First, we establish a first-principles model, a data-driven model, and three hybrid models and analyze the strengths and weaknesses of these models. Next, we propose the concepts of "mechanism-dominated models" and "data-dominated models" to classify the existing models according to the prediction abilities instead of the structural features. Furthermore, a novel type of hybrid model termed adaptive weighted hybrid model (AWHM) is proposed to gain a better balance between extrapolation and interpolation capabilities. Then, the performance of the operational optimization based on these models is compared under four optimization scenarios. Experimental results demonstrate that the two variants of the proposed AWHM have the best optimization performance and their optimization results are the most consistent with the actual process.

No abstract

No abstract

No abstract