Residuum hydrogenation model based on semi-supervised deep GRU and establishing method thereof

Abstract

The invention discloses a residuum hydrogenation model based on semi-supervised deep GRU and an establishing method thereof. The establishing method includes the steps: carrying out time sequence processing on original data to obtain a label-free sample training set A and a labeled sample training set A *; normalizing the sample data by using a min-max standardization method so as to eliminate theinfluence caused by different dimensions, wherein the normalized sample data sets A and A * are input data of the modeling process; constructing a deep GRU model by using the processed label-free sample data set A, and performing unsupervised pre-training on the deep GRU model layer by layer according to a greedy principle by using a GRU automatic encoder; and by means of parameters obtained through pre-training, initializing the deep GRU model of the same structure, carrying out secondary training by means of the labeled data set A *, and after supervised learning fine adjustment, obtaininga final residuum hydrogenation model based on the semi-supervised deep GRU. Compared with other existing methods, the establishing method provided by the invention can effectively improve the data utilization rate.

Description

technical field [0001] The invention relates to the technical field of residual oil hydrogenation, in particular to a residual oil hydrogenation model based on a semi-supervised deep GRU and a method for establishing it. Background technique [0002] The residual oil hydrogenation process can remove impurities such as sulfur, nitrogen, and metals, reduce the residual carbon content, and provide high-quality raw materials for heavy oil catalytic cracking units. An important means of quality oil. Hydrogen is an important raw material for the hydrofining process. The residual oil hydrogenation unit is a hydrogen consumption unit. The source of hydrogen includes new hydrogen from the hydrogen production unit and recycled hydrogen after purification. Due to the time delay in the production of hydrogen by the hydrogen production unit and the scheduling of hydrogen by the system, it is possible to establish a stable and accurate forecast of the new hydrogen flow required by the re...

AI Technical Summary

Problems solved by technology

For multivariate statistical analysis modeling methods such as principal component analysis, partial least squares, etc., it is impossible to perform good nonlinear fitting on complex chemical processes; while common machine learning methods such as support vector machines and BP neural networks cannot It handles potential features in time series data very well; as a commonly used method for processing time series data, recurrent neural network contains a large number of structural parameters, and is prone to gradient disappearance problems that are difficult to train

At the same time, the above-mentioned traditional methods can only model for labeled data samples, and there are problems such as insufficient data feature mining and low utilization rate.

The above limitations affect the running accuracy and robustness of the model

Images