WO2020147461A1 - 一种基于数据驱动的溶剂脱沥青装置建模及优化方法 - Google Patents

一种基于数据驱动的溶剂脱沥青装置建模及优化方法 Download PDF

Info

Publication number
WO2020147461A1
WO2020147461A1 PCT/CN2019/124325 CN2019124325W WO2020147461A1 WO 2020147461 A1 WO2020147461 A1 WO 2020147461A1 CN 2019124325 W CN2019124325 W CN 2019124325W WO 2020147461 A1 WO2020147461 A1 WO 2020147461A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
model
content
dao
solvent
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/124325
Other languages
English (en)
French (fr)
Inventor
钱锋
钟伟民
隆建
杨明磊
杜文莉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
East China University of Science and Technology
Original Assignee
East China University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by East China University of Science and Technology filed Critical East China University of Science and Technology
Publication of WO2020147461A1 publication Critical patent/WO2020147461A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Definitions

  • the invention relates to a data-driven method for optimizing simulation and operating conditions of a solvent deasphalting device.
  • This method can be used to simulate the solvent deasphalting process, predict product yield and properties, and quickly and intelligently optimize the main operating conditions of the device.
  • the solvent deasphalting device is divided into two parts: the solvent extraction section and the solvent recovery section.
  • the principle of the core extraction section is the liquid-liquid extraction of the solvent-based on the solvent's ability to dissolve different chemical components in the residue to separate light and heavy Liquid phase.
  • the composition of residual oil is extremely complex.
  • Most of the existing studies use thermodynamic prediction of phase equilibrium data, supplemented by a small amount of experimental data as an extension, or divide new family components for each temperature fraction according to different K values.
  • the mechanism model mostly adopts trial and error method in the division of residue virtual family components, which requires continuous debugging based on experience, which has high complexity.
  • the invention aims to provide a data-driven simulation and optimization method for a solvent deasphalting device.
  • a data-driven modeling method for solvent deasphalting equipment includes:
  • the on-site data includes production and operation data of the solvent deasphalting device, raw material and product property analysis data;
  • PCA Principal component analysis
  • BP back propagation neural network
  • the preprocessing of the data includes data reconciliation, cleaning, transformation and reduction.
  • the preprocessing of the data includes after removing the abnormal data points, processing the DCS data into a 12h mean value form, and performing an interpolation operation on the default value of the property data for the corresponding time period and operation Combine the data and normalize it to the interval [-1, 1].
  • the characteristic variables include input variables for the DAO yield model and variables of the DAO desulfurization and carbon residue removal model;
  • the input variables of the DAO yield model include residual oil density ⁇ VR , kinematic viscosity ⁇ VR , residual carbon mass fraction w VR , sulfur content w S , iron content w Fe , nickel content w Ni , vanadium content w V.
  • Saturated content w SH aromatic content w AH , colloidal content w R , asphaltene content w A , extraction tower inlet temperature T R1I , extraction tower top temperature T R1T , extraction tower bottom temperature T R1B , The settling tower inlet temperature T R2I , the settling tower top pressure P R2T , the settling tower sub-solvent temperature T R2FRJ , the cold slag feed flow rate F VR and the solvent ratio R total 19-dimensional input characteristic variables.
  • the variables of the DAO desulfurization and decarbonization model include residual oil density ⁇ VR , kinematic viscosity ⁇ VR , sulfur content w S , iron content W Fe , nickel content w Ni , saturation content w SH , aromatic content content w AH , colloid content w R , asphaltene content w A , extraction tower inlet temperature T R1I , extraction tower top temperature T R1T , settling tower top pressure P R2T , settling tower top temperature T R2T , settling tower secondary solvent temperature T R2FRJ , cold slag feed flow rate F VR and solvent ratio R total 16-dimensional characteristic variables.
  • the PCA method is used to reduce the dimensionality of feature variables, and components with feature contribution rates greater than 99% are selected as modeling variables.
  • a data-driven method for optimizing a solvent deasphalting device includes using the method provided by the present invention to construct a data-driven solvent deasphalting device model to optimize the solvent deasphalting device
  • the optimization includes: using a particle swarm algorithm based on differential evolution and using a trained model to optimize the operating conditions of the solvent deasphalting device according to the production target.
  • the optimization target is defined as:
  • the decision variable X is the operating conditions of the solvent deasphalting device.
  • y DAO represents the output of the DAO yield model, which is the yield of the product deasphalted oil.
  • the DAO yield model is linked with the desulfurization and decarbonization model, and the device is operated through the DAO yield model according to the production goal of maximizing the yield of deasphalted oil
  • the conditions are optimized, and the operating conditions generated in each iteration are sent to the desulfurization and decarbonization model to obtain the sulfur and carbon residue content of the deasphalted oil, and determine whether the value meets the constraints, and set the temperature at the top of the extraction tower to be greater than The temperature at the bottom of the extraction tower to ensure the executable of the optimized results.
  • the present invention solves the problems in the prior art by providing a simulation method for a solvent deasphalting device.
  • Figure 1 is a schematic diagram of the solvent deasphalting process.
  • Figure 2 is a distribution diagram of the DAO sample yield.
  • Figure 3 is the PCA-BP model structure.
  • Figure 4 shows the effect of DAO yield model on yield.
  • Figure 5 shows the simulation effect of DAO desulfurization and carbon residue removal model on carbon residue.
  • Figure 6 is the simulation effect of the DAO desulfurization and decarbonization model on the sulfur content.
  • Figure 7 shows the effect of simulated operating conditions on the DAO yield.
  • Figure 8 shows the effect of simulated operating conditions on the DAO sulfur and carbon residue content.
  • the inventor provided a method for predicting the product yield and sulfur and residual carbon content of a solvent deasphalting device based on industrial data, and an optimization method for device operating conditions.
  • this method uses the methods of data reconciliation, cleaning, transformation, and reduction to preprocess the field data to eliminate fault points and smooth the data; after preprocessing
  • the characteristic variables are selected according to the process principle and the characteristics of the production process, and the characteristic variables that have a greater impact on the DAO yield, DAO sulfur content and residual carbon content are screened out; the principal component analysis method ( PCA) dimensionality reduction to reduce the impact of noise;
  • the dimensionality reduction data is input into the neural network model for training, and the DAO yield and DAO desulfurization and carbon residue removal model are obtained;
  • the particle swarm algorithm based on differential evolution is used to build The model optimizes the operating conditions, and obtains the operating conditions that the DAO
  • the data-driven simulation and optimization method of a solvent deasphalting device includes the following steps:
  • Step 1 Collect production operation data, raw material and product property analysis data of solvent deasphalting device
  • Step 2 Perform data reconciliation, cleaning, transformation and reduction of the missing and redundant parts of the field data
  • Step 3 Preliminarily screen the characteristic variables in the data obtained in Step 2 according to the process principle and process characteristics;
  • Step 4 Use principal component analysis (PCA) to reduce the dimensionality of the data set filtered by the feature variable in step 3 to eliminate abnormal noise;
  • PCA principal component analysis
  • Step 5 The data set after dimensionality reduction in step 4 is modeled by back propagation (BP) neural network, and the DAO yield model and the desulfurization and decarbonization model are established respectively;
  • BP back propagation
  • Step 6 Use the particle swarm algorithm based on differential evolution and use the trained model to optimize the operating conditions of the solvent deasphalting device according to the production target.
  • the optimization target is defined as:
  • the decision variable X is the operating conditions of the solvent deasphalting device.
  • y DAO represents the output of the DAO yield model, which is the yield of the product deasphalted oil.
  • step one the production operation data, raw material and product analysis data of the solvent deasphalting device are collected on site.
  • the production operation data is generally measured in seconds. Considering that there is a certain time lag in the industrial process, in one embodiment of the present invention, after removing data points with abnormal noise, the production operation data is processed into a 12h average form.
  • the analysis data usually has a daily analysis cycle, and the default value of the material analysis data corresponding to the production operation data is interpolated, and finally combined with the production operation data. Finally, a data set corresponding to the properties of raw materials, operating conditions and product properties is obtained and normalized to the interval [-1, 1].
  • modeling variables are screened for DAO yield and sulfur and carbon residue indicators respectively.
  • the input variables of the DAO yield model include residual oil density ⁇ VR and kinematic viscosity ⁇ VR , residual carbon mass fraction w VR , sulfur content w S , iron content w Fe , nickel content w Ni , vanadium content w V , saturation content w SH , aromatic content w AH , colloidal content w R , asphaltene Content w A , extraction tower inlet temperature T R1I , extraction tower top temperature T R1T , extraction tower bottom temperature T R1B , settling tower inlet temperature T R2I , settling tower top pressure P R2T , settling tower secondary solvent temperature T R2FRJ ,
  • the cold slag feed flow rate F VR and the solvent ratio R are 19-dimensional input characteristic variables.
  • the DAO desulfurization and decarbonization model includes residual oil density ⁇ VR , kinematic viscosity ⁇ VR , sulfur content w S , iron content w Fe , and nickel content w Ni , saturation content w SH , aromatic content w AH , colloidal content w R , asphaltene content w A , extraction tower inlet temperature T R1I , extraction tower top temperature T R1T , settling tower top pressure P R2T , sedimentation The temperature at the top of the tower T R2T , the temperature of the secondary solvent in the settling tower T R2FRJ , the cold slag feed flow rate F VR and the solvent ratio R total 16-dimensional characteristic variables.
  • step four the principal component analysis method is used to reduce the dimension of the selected feature subset, compress it to a low-dimensional space, and remove the noise in the original feature.
  • PCA is used to reduce the dimensionality of the characteristic variables of the solvent deasphalting process, and components with a characteristic contribution rate greater than 99% are selected as modeling variables.
  • back-propagation neural network is used to establish the models of DAO yield, DAO sulfur content, and residual carbon content.
  • the input is 19-dimensional raw material properties and operating conditions, and the output is the DAO yield;
  • the input is 16-dimensional The nature of raw materials and operating conditions, the output is DAO carbon residue content and sulfur content.
  • the particle swarm optimization algorithm based on differential evolution is used to link the DAO yield model with the desulfurization and decarbonization model, and according to the production goal of maximizing the yield of deasphalted oil, through DAO
  • the yield model optimizes the operating conditions of the device, and sends the operating conditions generated in each iteration to the desulfurization and decarbonization model to obtain the sulfur and carbon residue content of the deasphalted oil, and determine whether the value meets the constraints and set
  • the temperature at the top of the extraction tower is greater than the temperature at the bottom of the extraction tower to ensure the executable of the optimized results.
  • step 6 a particle swarm algorithm based on differential evolution is used to optimize the model.
  • the optimization goal is defined as:
  • the decision variable X is the operating conditions of the solvent deasphalting unit, including extraction tower inlet temperature T R1I , extraction tower top temperature T R1T , extraction tower bottom temperature T R1B , settling tower inlet temperature T R2I , settling tower top pressure P R2T , the sub-solvent temperature T R2FRJ of the settling tower, the cold slag feed flow rate F VR and the solvent ratio R total 8 dimensional variables.
  • y DAO represents the output of the trained DAO yield model.
  • the method of the present invention is simple, fast and practical, and avoids the tediousness and uncertainty of artificially dividing the virtual components of the feed into groups and determining the content by trial and error.
  • the present invention uses a neural network to realize rapid prediction of product DAO yield and its sulfur and residual carbon content; adopts swarm intelligence optimization algorithm to optimize the operating conditions of the device, so that it can meet the demand for maximizing production benefits.
  • the process flow of the device is shown in Figure 1.
  • the whole device is mainly divided into two parts: subcritical extraction and supercritical solvent recovery.
  • the vacuum residue from the previous device is mixed with the solvent and then enters the extraction tower, and the heavy deasphalted solution from the top enters the colloid settling tower for further sedimentation and separation.
  • the DAO solution drawn from the top of the settling tower is heated and enters the stripping system for solvent recovery.
  • the colloidal solution drawn from the bottom of the settling tower and the DOA solution drawn from the bottom of the extraction tower are mixed and heated to enter the stripping tower to recover the solvent.
  • the recovered solvent is returned to the tower for recycling.
  • the sources of factory data are generally divided into two types, one is the device operating condition data collected by the distributed distributed control system (DCS), and the other is the raw material and product analysis data collected by the laboratory information management system (LIMS). Due to the limitation of the reliability of on-site testing instruments, the data directly obtained from DCS often has problems such as material imbalance and heat imbalance, so it cannot be directly used to establish device models. In order to ensure the accuracy of the model sample data, it is necessary to establish cleaning and reconciliation standards for the data collected in real time.
  • DCS distributed distributed control system
  • LIMS laboratory information management system
  • Normalization reduces the influence of the inconsistency of magnitude and dimension between features on the model, and speeds up the solution of the model.
  • the purpose of introducing principal component analysis is to use as few input variables as possible to complete the estimation of the variables to be measured, remove the interference information in the original variables, retain useful information, and simplify the input of the model.
  • ⁇ is the sample mean and ⁇ is the sample standard deviation.
  • the variance contribution rate of all variables is usually used to determine the principal elements that should be retained. Define the variance contribution rate of the first k inputs:
  • a principal element with a contribution rate greater than 99% is selected as the input of the subsequent neural network.
  • the data after dimensionality reduction by PCA is used as the input of BP neural network to establish DAO yield and DAO sulfur content and residual carbon content models respectively.
  • the present invention adopts back propagation (BP) neural network, which is a neural network structure most widely used in process control.
  • BP neural network is shown in Figure 3.
  • FS Feature Selection
  • the n features with the highest importance are selected as the new feature subset
  • X FS ⁇ x 1 , x 2 ,..., x n ⁇ .
  • X FS selects the first k principal elements P 1 , P 2 ,..., P k with the largest contribution after PCA dimensionality reduction as the input of the BP neural network.
  • the neural network is composed of an input layer, a hidden layer and an output layer: the input layer has a total of k neurons, the hidden layer has a total of q nodes, and the output layer has p outputs.
  • the connection weights between the input layer and the hidden layer, and between the hidden layer and the output layer are W k ⁇ q and V q ⁇ p respectively .
  • the hidden layer and output layer neurons of the neural network adopt the Sigmoid nonlinear activation function.
  • the purpose of learning is to find a series of weights so that after each set of input vectors of the sample acts on the network, the actual output vector of the network and the expected output of the sample The vectors are consistent.
  • the whole process is to adjust the connection weight of each neuron in the network to minimize the error energy function of the network:
  • I the predicted output of the input sample x k .
  • the error backpropagation (BP) algorithm is used for training, and the Levenberg-Marquart algorithm with faster search speed is used to replace the traditional gradient descent method to minimize the error square sum.
  • H is the number of neurons in the hidden layer
  • m is the number of neurons in the input layer
  • n is the number of neurons in the output layer
  • L is a constant between 1 and 10.
  • the PCA-BP models of DAO yield and sulfur and carbon residue content were established respectively.
  • the network structure of the DAO yield model is: 13-7-1
  • Figure 4 is the comparison between the simulated results of the test data on the yield model and the actual value
  • the network structure of the DAO desulfurization and carbon residue removal model is 13-6-2, Figure 4 5.
  • Figure 6 shows the comparison between the simulated results of the residual carbon content and the sulfur content and the actual values
  • Figures 7 and 8 show the effect of changes in operating conditions on the yield and properties of DAO.
  • Table 1 shows the prediction accuracy of the model. It can be seen that the model has a high simulation accuracy for DAO yield and carbon residue and sulfur content.
  • the production optimization direction of solvent deasphalting equipment is usually to find the best operating conditions that maximize the yield of deasphalted oil under the condition that the properties of deasphalted oil meet the constraint requirements.
  • the intelligent search algorithm based on the built DAO yield and property simulation model, the key operating conditions of the solvent deasphalting process can be optimized according to this production demand.
  • PSO Particle Swarm Optimization
  • PSO organically combines the individuality and sociality of the particles in the group to guide the search. It has the characteristics of simple and easy implementation and fast convergence speed, but it also has the problem of easily falling into a local optimal solution.
  • Differential evolution algorithm (Differential Evolution, DE), like PSO algorithm, is also an optimization algorithm based on the theory of swarm intelligence. The competition and cooperation between populations are the basis for optimized search among individuals within a group.
  • the mutation mechanism of the algorithm that is, the method of generating offspring is:
  • r′ r 1 +F(r 2 -r 3 ) (7)
  • r 1 , r 2 and r 3 are 3 different individuals randomly selected from the evolutionary group. r'is a new individual generated from three randomly selected individuals.
  • DE algorithm does not depend on specific problems and has the ability to memorize individual optimal solutions. At the same time, DE algorithm also has problems such as slow convergence speed in the later stage and insufficient robustness.
  • the present invention introduces a differential evolution correction strategy on the basis of PSO, that is, performs a differential after each iteration.
  • the evolution operator is based on the crossover and mutation of the individual's current optimal solution. Replace the current optimal solution with the optimal solution with better fitness after evolution, guide the PSO to search for the optimal solution, and increase its ability to jump out of the local optimal solution.
  • the industry generally hopes to maximize the yield of DAO while ensuring that the production index meets a certain requirement.
  • carbon residue content and sulfur content are more important indicators to measure the yield of DAO.
  • the DE-PSO algorithm is now used to optimize the main operating conditions of the solvent deasphalting unit.
  • the operating conditions generated by each iteration of the DAO yield model are sent to the desulfurization and decarbonization model, and then the deasphalted oil under the corresponding operating conditions is obtained.
  • the sulfur content and residual carbon value are compared with the set deasphalted oil sulfur and residual carbon constraint indicators to determine the final executable optimal operating conditions.
  • the given quality index is CCR ⁇ 7%, S ⁇ 1.8%.
  • the model prediction result is used as the individual fitness value:
  • F i is the fitness value of individual i
  • y DAOi is the model prediction value of the yield of individual i.
  • the optimization goal is defined as:
  • the decision variable X is the operating conditions of the solvent deasphalting unit, including extraction tower inlet temperature T R1I , extraction tower top temperature T R1T , extraction tower bottom temperature T R1B , settling tower inlet temperature T R2I , settling tower top pressure P R2T , the sub-solvent temperature T R2FRJ of the settling tower, the cold slag feed flow rate F VR and the solvent ratio R total 8 dimensional variables.
  • y DAO represents the output of the DAO yield model.
  • Three sets of LIMS raw material data are selected as shown in Table 2. Under the condition that the properties of the raw materials remain unchanged, the operating conditions that maximize F i are found. In order to obtain the optimal operating conditions within the set carbon residue index range and conform to the principle of the extraction process, set the constraint condition as T R1T >T R1B , that is to ensure that the temperature at the top of the extraction tower is greater than the temperature at the bottom of the tower during the optimization process . In each iteration process, the value obtained by the yield model is sent to the desulfurization and decarbonization model to obtain the residual carbon content, and the particles that do not meet the requirements are punished, as shown in formula (10).

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Production Of Liquid Hydrocarbon Mixture For Refining Petroleum (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于数据驱动的溶剂脱沥青装置建模及优化方法。所述基于数据驱动的溶剂脱沥青装置建模方法包括：(1)现场数据的预处理，所述现场数据包括溶剂脱沥青装置的生产操作数据、原料和产品性质分析数据；(2)筛选特征变量，所述筛选根据工艺原理和过程进行；(3)对特征变量采用主成分分析法(PCA)进行降维；和(4)采用反向传播(BP)神经网络就降维后的数据集进行建模，分别建立脱沥青油(DAO)的收率模型和脱硫脱残炭模型。所述优化方法包括利用构建得到的基于数据驱动的溶剂脱沥青装置模型优化溶剂脱沥青装置操作条件的步骤。

Description

一种基于数据驱动的溶剂脱沥青装置建模及优化方法 技术领域
本发明涉及一种基于数据驱动的溶剂脱沥青装置模拟与操作条件的优化方法。该方法可用于模拟溶剂脱沥青过程，预测产品收率和性质，并且快速智能地优化装置的主要操作条件。
背景技术
随着现代工业的迅速发展，石油作为国家战略物资在国民经济中发挥着至关重要的作用。由于原油质量逐渐劣化，炼油过程中的渣油再利用问题得到了重视。溶剂脱沥青(solvent deasphalting，SDA)工艺是生产催化裂化或加氢裂化原料的一个重要环节，增加了重油的加工深度和附加值。作为炼油企业重要的加工单元之一，对溶剂脱沥青装置进行模拟和优化，使其适应不同工况以满足后续加工单元的要求至关重要。
溶剂脱沥青装置分为溶剂抽提段和溶剂回收段两部分，其核心抽提段的原理是溶剂的液液萃取——基于溶剂对渣油中不同化学族组分的溶解能力不同来分离轻重液相。渣油的成分极其复杂，现有的研究大多采用热力学预测相平衡数据，辅以少量实验数据作为延伸，或者按照K值不同给每一个温度馏分划分新的族组分。机理模型在对渣油虚拟族组分的划分上多采用试凑法，需要凭经验不断调试，具有很高的复杂性。
因此，本领域迫切需要一种溶剂脱沥青装置模拟方法，用以解决上述技术问题。
发明内容
本发明旨在提供一种基于数据驱动的溶剂脱沥青装置模拟及优化方法。
在本发明的第一方面，提供一种基于数据驱动的溶剂脱沥青装置建模方法，所述方法包括：
(1)现场数据的预处理，所述现场数据包括溶剂脱沥青装置的生产操作数据、原料和产品性质分析数据；
(2)筛选特征变量，所述筛选根据工艺原理和过程特性进行；
(3)对特征变量采用主成分分析法(PCA)进行降维；和
(4)采用反向传播(BP)神经网络就降维后的数据集进行建模，分别建立脱沥青油(DAO)的收率模型和脱硫脱残炭模型。
在另一优选例中，所述数据的预处理包括进行数据调和、清洗、变换和归约。
在另一优选例中，所述数据的预处理包括在剔除掉异常数据点后，将DCS数据处理成12h的均值形式，并对相应时间段的性质数据的缺省值进行插值操作后与操作数据组合，并将其归一化到[-1，1]区间。
在另一优选例中，所述特征变量包括针对DAO收率模型的输入变量，以及DAO脱硫脱残炭模型的变量；
更佳地，所述DAO收率模型的输入变量包括渣油密度ρ VR、运动粘度μ VR、残炭质量分数w VR、硫含量w S、铁含量w Fe、镍含量w Ni、钒含量w V、饱和分含量w SH、芳香分含量w AH、胶质含量w R、沥青质含量w A、抽提塔入口温度T R1I、抽提塔顶部温度T R1T、抽提塔底部温度T R1B、沉降塔入口温度T R2I、沉降塔顶部压力P R2T、沉降塔副溶剂温度T R2FRJ、冷渣进料流量F VR和溶剂比R等共计19维输入特征变量。
更佳地，所述DAO脱硫脱残炭模型的变量包括渣油密度ρ VR、运动粘度μ VR、硫含量w S、铁含量W Fe、镍含量w Ni、饱和分含量w SH、芳香分含量w AH、胶质含量w R、沥青质含量w A、抽提塔入口温度T R1I、抽提塔顶部温度T R1T、沉降塔顶部压力P R2T、沉降塔顶部温度T R2T、沉降塔副溶剂温度T R2FRJ、冷渣进料流量F VR和溶剂比R等共计16维特征变量。
在另一优选例中，采用PCA方法对特征变量进行降维，选取特征贡献率大于99％的成分作为建模变量。
在本发明的第二方面，提供一种基于数据驱动的溶剂脱沥青装置优化方法， 所述方法包括利用上述本发明提供的方法构建得到的基于数据驱动的溶剂脱沥青装置模型优化溶剂脱沥青装置的步骤，其中，所述优化包括：使用基于差分进化的粒子群算法，利用训练完毕的模型，根据生产目标，对溶剂脱沥青装置的操作条件进行优化，优化目标定义为：
max f(X)＝y DAO
其中，决策变量X为溶剂脱沥青装置的操作条件。y DAO表示DAO收率模型的输出，为产品脱沥青油的收率。
在另一优选例中，通过基于差分进化的粒子群优化算法，利用DAO收率模型和脱硫脱残炭模型联动，根据脱沥青油收率最大化的生产目标，通过DAO收率模型对装置操作条件进行优化，并将每一次迭代生成的操作条件送入脱硫脱残炭模型中求得脱沥青油的硫、残炭含量，并判断该值是否满足约束，同时设定抽提塔顶部温度大于抽提塔底部温度，以此保证优化结果的可执行性。
据此，本发明通过提供一种溶剂脱沥青装置模拟方法，解决了现有技术中的问题。
附图说明
图1是溶剂脱沥青工艺过程示意图。
图2是DAO样本收率分布图。
图3是PCA-BP模型结构。
图4是DAO收率模型对收率的效果。
图5是DAO脱硫脱残炭模型对残炭的模拟效果。
图6是DAO脱硫脱残炭模型对硫含量的模拟效果。
图7是模拟的操作条件变化对DAO收率的影响。
图8是模拟的操作条件变化对DAO硫和残炭含量的影响。
具体实施方式
发明人经过深入研究，提供了一种基于工业数据的溶剂脱沥青装置产品收率和硫、残炭含量预测方法及装置操作条件的优化方法。该方法在获得炼厂溶剂脱沥青装置运行数据的基础上，采用数据调和、清洗、变换、归约的 方法对现场数据进行预处理，以剔除故障点和对数据进行平滑处理；对预处理后的数据根据工艺原理和生产过程特性进行特征变量的选择，分别筛选出对DAO收率和DAO硫含量和残炭含量影响较大的特征变量；对于特征选择后的数据集采用主元分析法(PCA)对其进行降维，减少噪声的影响；降维后的数据输入到神经网络模型中进行训练，获得DAO收率和DAO脱硫脱残炭模型；采用基于差分进化的粒子群算法对所建模型进行操作条件的优化，求得在给定原料性质下产品DAO收率尽可能大且满足残炭质量指标约束的操作条件，为生产优化提供方案。
本发明提供的基于数据驱动的溶剂脱沥青装置模拟及优化方法，所述方法包括以下步骤：
步骤一：采集溶剂脱沥青装置的生产操作数据、原料和产品性质分析数据；
步骤二：对现场数据中缺失和冗余的部分进行数据调和、清洗、变换和归约等预处理环节；
步骤三：根据工艺原理和过程特性，对步骤二获得的数据中的特征变量进行初步筛选；
步骤四：对步骤三中经过特征变量筛选的数据集采用主成分分析法(PCA)对其进行降维，剔除异常噪声；
步骤五：步骤四中降维后的数据集采用反向传播(BP)神经网络进行建模，分别建立DAO收率模型和脱硫脱残炭模型；
步骤六：使用基于差分进化的粒子群算法，利用训练完毕的模型，根据生产目标，对溶剂脱沥青装置的操作条件进行优化，优化目标定义为：
max f(X)＝y DAO
其中，决策变量X为溶剂脱沥青装置的操作条件。y DAO表示DAO收率模型的输出，为产品脱沥青油的收率。
在本发明的一种实施方式中，步骤一中现场采集溶剂脱沥青装置的生产操作数据、原料和产品分析数据。
步骤二中由于现场数据的缺失和冗余，不能直接用来建模。因此需要对采集到的数据进行清洗，变换和归约。现场的数据存在着物料不平衡，仪表的某些读数故障等问题，需要在进行数据分析的基础上对缺失值和异常值进行处理。生产操作数据一般以秒计，考虑到工业流程存在一定的时滞，在本 发明的一种实施方式中，剔除存在异常噪声的数据点之后，将生产操作数据处理成12h的均值形式。分析数据通常分析周期为日，对与生产操作数据对应时间段的物料分析数据的缺省值进行插值操作，最后与生产操作数据组合。最终得到原料性质、操作条件和产品性质相对应的数据集，并将其归一化到[-1，1]区间。
根据工艺过程特性，在本发明的一种实施方式中，步骤三中针对DAO收率以及硫和残炭指标分别筛选建模变量，DAO收率模型的输入变量包含渣油密度ρ VR、运动粘度μ VR、残炭质量分数w VR、硫含量w S、铁含量w Fe、镍含量w Ni、钒含量w V、饱和分含量w SH、芳香分含量w AH、胶质含量w R、沥青质含量w A、抽提塔入口温度T R1I、抽提塔顶部温度T R1T、抽提塔底部温度T R1B、沉降塔入口温度T R2I、沉降塔顶部压力P R2T、沉降塔副溶剂温度T R2FRJ、冷渣进料流量F VR和溶剂比R等共计19维输入特征变量，DAO脱硫脱残炭模型包含渣油密度ρ VR、运动粘度μ VR、硫含量w S、铁含量w Fe、镍含量w Ni、饱和分含量w SH、芳香分含量w AH、胶质含量w R、沥青质含量w A、抽提塔入口温度T R1I、抽提塔顶部温度T R1T、沉降塔顶部压力P R2T、沉降塔顶部温度T R2T、沉降塔副溶剂温度T R2FRJ、冷渣进料流量F VR和溶剂比R等共计16维特征变量。
步骤四中，采用主成分分析法对选取的特征子集进行降维，将其压缩到低维空间，剔除原特征中的噪声。在本发明的一种实施方式中，步骤四采用PCA对溶剂脱沥青过程的特征变量进行降维，选取特征贡献率大于99％的成分作为建模变量。
根据步骤四中降维后的数据，采用反向传播神经网络，分别建立DAO收率和DAO硫含量、残炭含量的模型。在本发明的一种实施方式中，步骤五中对于DAO收率模型，输入为19维的原料性质和操作条件，输出为DAO的收率；对于DAO脱硫脱残炭模型，输入为16维的原料性质和操作条件，输出为DAO残炭含量和硫含量。
基于差分进化的粒子群优化算法，利用训练完毕的模型，在脱沥青油的残炭、硫含量满足限制指标下，优化得到脱沥青油收率最大对应的装置操作条件。在本发明的一种实施方式中，步骤六中通过基于差分进化的粒子群优化算法，利用DAO收率模型和脱硫脱残炭模型联动，根据脱沥青油收率最大化的生产目标，通过DAO收率模型对装置操作条件进行优化，并将每一次迭代生成的操作条件送入脱硫脱残炭模型中求得脱沥青油的硫、残炭含量，并判断该值是否满足约束，同时设定抽提塔顶部温度大于抽提塔底部温度，以此保证优化结果的可执行性。
在本发明的一种实施方式中，步骤六采用基于差分进化的粒子群算法对模型进行优化。优化目标定义为：
max f(X)＝y DAO
其中，决策变量X为溶剂脱沥青装置的操作条件，包括抽提塔入口温度T R1I，抽提塔顶部温度T R1T，抽提塔底部温度T R1B，沉降塔入口温度T R2I，沉降塔顶部压力P R2T，沉降塔副溶剂温度T R2FRJ，冷渣进料流量F VR和溶剂比R共计8维变量。y DAO表示训练好的DAO收率模型的输出。
本发明提到的上述特征，或实施例提到的特征可以任意组合。本案说明书所揭示的所有特征可与任何组合物形式并用，说明书中所揭示的各个特征，可以任何可提供相同、均等或相似目的的替代性特征取代。因此除有特别说明，所揭示的特征仅为均等或相似特征的一般性例子。
本发明的主要优点在于：
1、本发明方法简单、快速、实用，避免了人为对进料虚拟组分另外划分族组分并采用试差法确定其含量时的繁琐和不确定性。
2、本发明使用神经网络，可以实现对产品DAO收率和其硫、残炭含量的快速预测；采用群智能优化算法优化了装置的操作条件，使其能满足生产效益最大化的需求。
下面结合具体实施例，进一步阐述本发明。应理解，这些实施例仅用于说明本发明而不用于限制本发明的范围。下列实施例中未注明具体条件的实验方法，通常按照常规条件或按照制造厂商所建议的条件。除非另行定义，文中所使用的所有专业与科学用语与本领域熟练人员所熟悉的意义相同。此外，任何与所记载内容相似或均等的方法及材料皆可应用于本发明方法中。文中所述的较佳实施方法与材料仅作示范之用。
实施例1
下面结合图表具体介绍本发明的实现方法：
1.溶剂脱沥青工艺流程
装置的工艺流程如图1所示。整个装置主要分为亚临界抽提和超临界溶剂回收两个部分。上一个装置来的减压渣油和溶剂混合后进入抽提塔，从顶部出来的重脱沥青溶液进入胶质沉降塔进一步沉降分离。沉降塔顶部抽出的DAO溶液经过加热进入汽提系统进行溶剂回收。沉降塔底部抽出的胶质溶液和抽提塔底部抽出的DOA溶液进行混合后经加热进入汽提塔回收溶剂。回收后的溶剂返塔循环使用。
2.数据清洗和调和
由于工业现场情况复杂，生产过程受到多种因素的影响。工业现场的噪声和仪表自身的故障，采集到的数据存在着不完整和冗余，无法直接用来建模。工厂数据的来源一般分为两种，一种是分布式集散控制系统(DCS)采集的装置操作条件数据，一种是实验室信息管理系统(LIMS)采集的原料及产品分析数据。受现场检测仪表可靠性的局限，直接从DCS上获取到的数据往往存在物料不平衡、热量不平衡等问题，因此不能直接用于建立装置模型。为了确保模型样本数据的准确性，有必要对实时采集的数据建立清洗和调和标准，具体使用以下几种方法：(1)对仪表读数错误和缺失数据样本进行剔除(2)对DCS数据进行降采样，将其处理成12h的均值形式，以减小数据的波动。(3)对相应时间段的LIMS数据的缺省值进行插值操作后与DCS数据组合。图2为DAO收率在样本集上的分布，横坐标表示收率的范围，纵坐标表示各收率区间的频数。可知样本集呈现近似的正态分布，满足建模的前提条件。将清洗过后的数据集归一化到[-1，1]区间，具体公式为：
Figure PCTCN2019124325-appb-000001
归一化减小了特征之间数量级和量纲的不一致问题对模型带来的影响，加快了模型的求解速度。
3.采用PCA对输入数据降维
引入主元分析的目的是用尽可能少的输入变量完成对待测变量的预估，剔除原有变量中的干扰信息，保留有用信息，且简化了模型的输入。首先对所有输入样本进行标准化，使归一化后的样本均值为0，方差为1：
Figure PCTCN2019124325-appb-000002
式中，μ为样本均值，σ为样本标准差。
通常利用所有变量的方差贡献率来确定应该保留下来的主元。定义前k个输入的方差贡献率为：
Figure PCTCN2019124325-appb-000003
本例中选取贡献率大于99％的主元作为后续神经网络的输入。
4.利用BP神经网络建立产品收率和性质模型
通过PCA降维后的数据作为BP神经网络的输入，分别建立DAO收率和DAO硫含量以及残炭含量模型。
本发明采用反向传播(Back propagation，BP)神经网络，这是一种在过程控制中应用最广泛的神经网络结构。BP神经网络结构如图3所示。最初的原始特征集为X＝{x 1，x 2，…，x m}，经过特征选择(Feature Selection，FS)之后选择了重要度靠前的n个特征作为新的特征子集X FS＝{x 1，x 2，…，x n}。X FS经过PCA降维后选取了贡献度最大的前k个主元P 1，P 2，…，P k作为BP神经网络的输入。神经网络的构成为输入层、隐含层和输出层：输入层共k个神经元，隐层节点共有q个，输出层有p个输出。输入层与隐含层、隐含层与输出层间的连接权值分别为W k×q和V q×p。
神经网络的隐层和输出层神经元采用Sigmoid非线性激活函数，学习目的是找出一系列权值，使样本的每组输入向量作用于网络后，其网络的实际输出向量与样本的期望输出向量一致。整个过程是调整网络中各神经元的连接权值，使网络的误差能量函数达到最小：
Figure PCTCN2019124325-appb-000004
式中，
Figure PCTCN2019124325-appb-000005
是输入样本x k的预测输出。
采用误差反向传播(BP)算法进行训练，采用搜索速度较快的Levenberg-Marquart算法取代传统的梯度下降方法使误差平方和最小。
神经网络的一项重要任务是确定隐含层结构，其神经元个数对网络预测精度有较大影响。神经元个数的确定没有固定规则，这里根据经验公式(5)在一定合理范围内进行计算：
Figure PCTCN2019124325-appb-000006
式中：H为隐含层神经元个数；m为输入层神经元个数；n为输出层神经元个数；L为1～10之间的常数。考虑到神经网络的非线性动态特性，可以在(5)式确定的范围下进行适当外延。
根据上述神经网络结构，分别建立DAO收率和硫、残炭含量的PCA-BP模型。DAO收率模型的网络结构为：13-7-1，图4为测试数据在收率模型上的模拟结果与实际值对比；DAO脱硫脱残炭模型的网络结构为13-6-2，图5、图6分别为残炭含量和硫含量的模拟结果与实际值对比；图7、图8为操作条件的变化对DAO收率和性质的影响。表1为模型的预测精度。可以看出模型对DAO收率和残炭、硫含量均有着较高的模拟精度。
表1：模型预测精度
Figure PCTCN2019124325-appb-000007
5.溶剂脱沥青装置操作条件的优化
溶剂脱沥青装置生产优化方向，通常是希望在脱沥青油性质指标符合约束要求下找到使脱沥青油收率最大的最佳操作条件。通过智能搜索算法，根据所建的DAO收率和性质模拟模型，可以根据这一生产需求对溶剂脱沥青过程关键操作条件进行优化。
粒子群优化算法(Particle Swarm Optimization，PSO)是一种基于随机群体的智能优化算法。粒子通过个体极值和群体极值更新自身的速度和位置，更新式为：
Figure PCTCN2019124325-appb-000008
其中，w为惯性权重；d＝1，2，…，D为当前的维度；i＝1，2，…，n当前第i个个体；k为当前迭代次数；V id为粒子的速度；c 1和c 2为非负的常数，称为加速度因子；r 1和r 2为分布于[0，1]的随机数。
PSO将群体中粒子的个体性和社会性有机结合来指导搜索，具有简单易实现、收敛速度快等特点，但同时也存在易于陷入局部最优解的问题。差分进化算法(Differential Evolution，DE)与PSO算法一样，也是基于群体智能理论的优化算法，种群间的竞争与合作是群体内个体间优化搜索的基础。算法的变异机制，即生成子代的方法为：
r′＝r 1+F(r 2-r 3) (7)
其中，r 1，r 2和r 3为从进化群体中随机选取的互不相同的3个互不相同的个体。r′为从三个随机选取的个体中产生的新个体。
DE算法具有不依赖于具体问题、具有记忆个体最优解的能力，但同时DE算法也存在着后期收敛速度较慢，表现不够稳健等问题。
为了改善PSO算法在面对高维多模态函数时因后期微粒多样性的降低导致的早熟收敛问题，本发明在PSO的基础上引入差分进化修正策略，即在每次迭代结束后进行一次差分进化算子基于个体当前最优解的交叉、变异。用进化之后适应度更好的最优解替换当前最优解，引导PSO向最优解方向搜索，增加其跳出局部最优解的能力。
工业上一般希望在保证生产指标达到某一要求的情况下最大化DAO的收率。通常残炭含量和硫含量是衡量DAO收率较为重要的指标。现采用 DE-PSO算法对溶剂脱沥青装置主要操作条件进行优化，将DAO收率模型每一次迭代生成的操作条件送入脱硫脱残炭模型中，进而求得相应操作条件下的脱沥青油的硫含量和残炭值，并与设定的脱沥青油硫、残炭约束指标进行比较，进而确定最终的可执行性的优化操作条件。给定的质量指标为CCR≤7％、S≤1.8％。
在收率迭代寻优过程中设置种群个数N＝300，最大迭代次数iter＝100，DE变异算子F＝0.5，交叉算子CR＝0.1。在计算过程中把模型预测结果作为个体适应度值：
F i＝y DAOi(i＝0，1，2，…，N)  (8)
式中：F i为个体i的适应度值；y DAOi为个体i收率的模型预测值。
优化目标定义为：
max f(X)＝y DAO    (9)
其中，决策变量X为溶剂脱沥青装置的操作条件，包括抽提塔入口温度T R1I，抽提塔顶部温度T R1T，抽提塔底部温度T R1B，沉降塔入口温度T R2I，沉降塔顶部压力P R2T，沉降塔副溶剂温度T R2FRJ，冷渣进料流量F VR和溶剂比R共计8维变量。y DAO表示DAO收率模型的输出。
选择三组LIMS原料数据如表2所示，在保持原料性质不变的情况下，寻找使F i最大的操作条件。为使求得的最优操作条件在设定的残炭指标范围之内且符合萃取过程原理，设置约束条件为T R1T＞T R1B，即保证寻优过程中抽提塔顶温度大于塔底温度。每次迭代过程中，将收率模型求得的值送入脱硫脱残炭模型中求得残炭含量，对于不符合要求的粒子进行惩罚，如公式(10)所示。
Figure PCTCN2019124325-appb-000009
优化过后的操作条件见表3。可以看出，DAO的收率均有不同程度的提高。
表2：所选LIMS原料性质
Figure PCTCN2019124325-appb-000010
表3：操作条件优化的结果
Figure PCTCN2019124325-appb-000011
以上所述仅为本发明的较佳实施例而已，并非用以限定本发明的实质技术内容范围，本发明的实质技术内容是广义地定义于申请的权利要求范围中，任何他人完成的技术实体或方法，若是与申请的权利要求范围所定义的完全相同，也或是一种等效的变更，均将被视为涵盖于该权利要求范围之中。

Claims (9)

  1. 一种基于数据驱动的溶剂脱沥青装置建模方法，其特征在于，所述方法包括：
    (1)现场数据的预处理，所述现场数据包括溶剂脱沥青装置的生产操作数据、原料和产品性质分析数据；
    (2)筛选特征变量，所述筛选根据工艺原理和过程特性进行；
    (3)对特征变量采用主成分分析法(PCA)进行降维；
    (4)采用反向传播(BP)神经网络就降维后的数据集进行建模，分别建立脱沥青油(DAO)的收率模型和脱硫脱残炭模型。
  2. 如权利要求1所述的建模方法，其特征在于，所述数据的预处理包括进行数据调和、清洗、变换和归约。
  3. 如权利要求1所述的建模方法，其特征在于，所述数据的预处理包括在剔除掉异常数据点后，将DCS数据处理成12h的均值形式，并对相应时间段的性质数据的缺省值进行插值操作后与操作数据组合，并将其归一化到[-1，1]区间。
  4. 如权利要求1所述的建模方法，其特征在于，所述特征变量包括针对DAO收率模型的输入变量，以及DAO脱硫脱残炭模型的变量。
  5. 如权利要求4所述的建模方法，其特征在于，所述DAO收率模型的输入变量包括渣油密度ρ VR、运动粘度μ VR、残炭质量分数w VR、硫含量w S、铁含量w Fe、镍含量w Ni、钒含量w V、饱和分含量w SH、芳香分含量w AH、胶质含量w R、沥青质含量w A、抽提塔入口温度T R1I、抽提塔顶部温度T R1T、抽提塔底部温度T R1B、沉降塔入口温度T R2I、沉降塔顶部压力P R2T、沉降塔副溶剂温度T R2FRJ、冷渣进料流量F VR和溶剂比R等共计19维输入特征变量。
  6. 如权利要求4所述的建模方法，其特征在于，所述DAO脱硫脱残炭模型的变量包括渣油密度ρ VR、运动粘度μ VR、硫含量w S、铁含量w Fe、镍含量w Ni、饱和分含量w SH、芳香分含量w AH、胶质含量w R、沥青质含量w A、抽提塔入口温度T R1I、抽提塔顶部温度T R1T、沉降塔顶部压力P R2T、沉降塔顶部温度T R2T、沉降塔副溶剂温度T R2FRJ、冷渣进料流量F VR和溶剂比R等共计16维特征变量。
  7. 如权利要求1所述的建模方法，其特征在于，采用PCA方法对特征变量进行降维，选取特征贡献率大于99％的成分作为建模变量。
  8. 一种基于数据驱动的溶剂脱沥青装置优化方法，其特征在于，所述方法包括利用权利要求1-7中任一项所述的方法构建得到的基于数据驱动的溶剂脱沥青装置模型优化溶剂脱沥青装置的步骤，其中，所述优化包括：使用基于差分进化的粒子群算法，利用训练完毕的模型，根据生产目标，对溶剂脱沥青装置的操作条件进行优化，优化目标定义为：
    max f(X)＝y DAO
    其中，决策变量X为溶剂脱沥青装置的操作条件。y DAO表示DAO收率模型的输出，为产品脱沥青油的收率。
  9. 如权利要求8所述的优化方法，其特征在于，通过基于差分进化的粒子群优化算法，利用DAO收率模型和脱硫脱残炭模型联动，根据脱沥青油收率最大化的生产目标，通过DAO收率模型对装置操作条件进行优化，并将每一次迭代生成的操作条件送入脱硫脱残炭模型中求得脱沥青油的硫、残炭含量，并判断该值是否满足约束，同时设定抽提塔顶部温度大于抽提塔底部温度，以此保证优化结果的可执行性。
PCT/CN2019/124325 2019-01-18 2019-12-10 一种基于数据驱动的溶剂脱沥青装置建模及优化方法 Ceased WO2020147461A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910048915.2A CN109657411B (zh) 2019-01-18 2019-01-18 一种基于数据驱动的溶剂脱沥青装置建模及优化方法
CN201910048915.2 2019-01-18

Publications (1)

Publication Number Publication Date
WO2020147461A1 true WO2020147461A1 (zh) 2020-07-23

Family

ID=66120448

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/124325 Ceased WO2020147461A1 (zh) 2019-01-18 2019-12-10 一种基于数据驱动的溶剂脱沥青装置建模及优化方法

Country Status (2)

Country Link
CN (1) CN109657411B (zh)
WO (1) WO2020147461A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113780626A (zh) * 2021-08-16 2021-12-10 中冶南方工程技术有限公司 一种基于进化算法的炉渣粘度预测方法
US20220012309A1 (en) * 2020-07-10 2022-01-13 Nice Ltd. Systems and methods for applying semi-discrete calculus to meta machine learning
CN115725316A (zh) * 2021-08-27 2023-03-03 中国石油化工股份有限公司 一种高等级道路沥青的制备方法
CN117467466A (zh) * 2022-07-22 2024-01-30 中国石油化工股份有限公司 确定溶剂脱沥青的温度与平衡分配性质的关系的方法、介质及设备
CN117682598A (zh) * 2023-11-24 2024-03-12 航天国华生态环境有限公司 一种废水中氯酚类化合物无害化处理装置及控制方法
WO2025255043A1 (en) * 2024-06-03 2025-12-11 Marathon Petroleum Company Lp Systems, analyzers, controllers, and associated methods to enhance fluid production of refining operations
US12599848B2 (en) 2024-06-03 2026-04-14 Marathon Petroleum Company Lp Systems, analyzers, controllers, and associated methods to enhance fluid separation for distillation operations

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109657411B (zh) * 2019-01-18 2023-08-04 华东理工大学 一种基于数据驱动的溶剂脱沥青装置建模及优化方法
CN115830201B (zh) * 2022-11-22 2024-05-24 光线云(杭州)科技有限公司 一种基于聚簇的粒子系统优化渲染方法和装置
CN116523388B (zh) * 2023-04-17 2023-11-10 无锡雪浪数制科技有限公司 一种基于工业互联网平台的数据驱动的质量建模方法
CN116300774B (zh) * 2023-05-23 2023-08-08 蓝星智云(山东)智能科技有限公司 基于主元分析与核密度估计的间歇过程可视化监控方法
CN116448209A (zh) * 2023-06-15 2023-07-18 合肥市富林物联网技术股份有限公司 一种基于压力传感器的液位识别方法、装置及电子设备

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104765346A (zh) * 2015-03-26 2015-07-08 华东理工大学 一种炼油过程全流程建模方法
CN104915522A (zh) * 2015-07-01 2015-09-16 华东理工大学 结合过程先验和数据驱动模型的混合建模方法及系统
CN109657411A (zh) * 2019-01-18 2019-04-19 华东理工大学 一种基于数据驱动的溶剂脱沥青装置建模及优化方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104463343A (zh) * 2014-10-27 2015-03-25 中国石油大学(北京) 一种预测催化裂化轻质油产率的方法
CN104789256B (zh) * 2015-03-26 2016-07-06 华东理工大学 一种催化裂化装置的收率实时预测方法
WO2017117685A1 (en) * 2016-01-08 2017-07-13 Mcmaster University Disease models and prognostic tools for pre-leukemia and leukemia

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104765346A (zh) * 2015-03-26 2015-07-08 华东理工大学 一种炼油过程全流程建模方法
CN104915522A (zh) * 2015-07-01 2015-09-16 华东理工大学 结合过程先验和数据驱动模型的混合建模方法及系统
CN109657411A (zh) * 2019-01-18 2019-04-19 华东理工大学 一种基于数据驱动的溶剂脱沥青装置建模及优化方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LONG XUN-JIAN,QIAN JU, LIANG CHUAN: "Water Demand Forecast Model of BP Neutral Networks based on Principle Component Analysis", JOURNAL OF CHENGDU UNIVERSITY OF TECHNOLOGY (SCIENCE & TECHNOLOGY EDITION), vol. 2, no. 37, 30 April 2010 (2010-04-30), pages 206 - 210, XP009522012 *
RONG, GANG ET AL.: "A Robust Engineering Strategy for Scheduling Optimization of Refinery Fuel Gas System", INDUSTRIAL & ENGINEERING CHEMISTRY RESEARCH, vol. 57, no. 5, 31 December 2018 (2018-12-31), pages 1547 - 1559, XP055719239 *

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220012309A1 (en) * 2020-07-10 2022-01-13 Nice Ltd. Systems and methods for applying semi-discrete calculus to meta machine learning
US12481890B2 (en) * 2020-07-10 2025-11-25 Nice Ltd. Systems and methods for applying semi-discrete calculus to meta machine learning
CN113780626A (zh) * 2021-08-16 2021-12-10 中冶南方工程技术有限公司 一种基于进化算法的炉渣粘度预测方法
CN115725316A (zh) * 2021-08-27 2023-03-03 中国石油化工股份有限公司 一种高等级道路沥青的制备方法
CN117467466A (zh) * 2022-07-22 2024-01-30 中国石油化工股份有限公司 确定溶剂脱沥青的温度与平衡分配性质的关系的方法、介质及设备
CN117682598A (zh) * 2023-11-24 2024-03-12 航天国华生态环境有限公司 一种废水中氯酚类化合物无害化处理装置及控制方法
WO2025255043A1 (en) * 2024-06-03 2025-12-11 Marathon Petroleum Company Lp Systems, analyzers, controllers, and associated methods to enhance fluid production of refining operations
WO2025255046A1 (en) * 2024-06-03 2025-12-11 Marathon Petroleum Company Lp Systems, analyzers, controllers, and associated methods to enhance solvent deasphalting operations
WO2025255051A1 (en) * 2024-06-03 2025-12-11 Marathon Petroleum Company Lp Systems, analyzers, controllers, and associated methods to enhance fluid production of refining operations
US12599848B2 (en) 2024-06-03 2026-04-14 Marathon Petroleum Company Lp Systems, analyzers, controllers, and associated methods to enhance fluid separation for distillation operations

Also Published As

Publication number Publication date
CN109657411A (zh) 2019-04-19
CN109657411B (zh) 2023-08-04

Similar Documents

Publication Publication Date Title
CN109657411B (zh) 一种基于数据驱动的溶剂脱沥青装置建模及优化方法
Yuan et al. Quality prediction modeling for industrial processes using multiscale attention-based convolutional neural network
CN109814513B (zh) 一种基于数据模型的催化裂化装置优化方法
CN104765346B (zh) 一种炼油过程全流程建模方法
CN111026058B (zh) 基于瓦瑟斯坦距离和自编码器的半监督深度学习故障诊断方法
CN112560338B (zh) 基于自适应深度学习的复杂工业系统智能预报方法、装置、设备及存储介质
CN102626557B (zh) 基于ga-bp算法的分子蒸馏工艺参数优化方法
CN103440368A (zh) 一种多模型动态软测量建模方法
Li et al. Data-driven modeling and operation optimization with inherent feature extraction for complex industrial processes
CN113780420A (zh) 基于gru-gcn的变压器油中溶解气体浓度预测方法
CN109102002A (zh) 结合卷积神经网络和概念机递归神经网络的图像分类方法
CN104463343A (zh) 一种预测催化裂化轻质油产率的方法
CN106354017A (zh) 一种稀土萃取分离过程组分含量区间控制方法
CN106842948A (zh) 基于bp网络的hdp分子蒸馏系统的最优控制方法
CN116662925A (zh) 一种基于加权稀疏神经网络工业过程软测量方法
CN104765347A (zh) 一种渣油延迟焦化过程中收率实时预测方法
CN108062566A (zh) 一种基于多核潜在特征提取的智能集成软测量方法
Zhu et al. Data-driven models of crude distillation units for production planning and for operations monitoring
CN110927478B (zh) 一种确定电力系统变压器设备的状态的方法和系统
CN115374858B (zh) 基于混合集成模型的流程工业生产品质的智能诊断方法
Pater Application of artificial neural networks and genetic algorithms for crude fractional distillation process modeling
CN106777466A (zh) 基于st‑upfnn算法的高含硫天然气净化工艺的动态演化建模方法
Alhajree et al. Modeling and optimization of an industrial hydrocracker plant
Zhu et al. Multiple input self-organizing-map ResNet model for optimization of petroleum refinery conversion units
CN113420498B (zh) 常减压装置的ai建模方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19909802

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19909802

Country of ref document: EP

Kind code of ref document: A1