Hybrid strategy integrating variable selection and a neural network for

Subscriber access provided by YORK UNIV

Process Systems Engineering

Hybrid strategy integrating variable selection and a neural network for fluid catalytic cracking modeling Jian Long, Tianyue Li, Ming-Lei Yang, Guihua Hu, and Weimin Zhong Ind. Eng. Chem. Res., Just Accepted Manuscript • DOI: 10.1021/acs.iecr.8b04821 • Publication Date (Web): 18 Dec 2018 Downloaded from http://pubs.acs.org on December 21, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

Hybrid strategy integrating variable selection and a neural network for fluid catalytic cracking modeling Jian Long, Tianyue Li, Minglei Yang, Guihua Hu, and Weimin Zhong* Key Laboratory of Advanced Control and Optimization for Chemical Processes, Ministry of Education, East China University of Science and Technology, Shanghai 200237, China ABSTRACT: Different from traditional modeling methods for maximizing iso-paraffins (MIP), a hybrid approach that integrates the least absolute shrinkage and selection operator (LASSO) method for variable selection and an output-focused back-propagation neural network (BPNN) method for predictive model construction is proposed in this paper. LASSO was used to reduce the dimensionality of the factors influencing the yield and property of products and eliminate the correlations among factors to obtain the feature variables that were used as BPNN input vectors. The combined LASSO-BPNN models were trained and tested using industrial production historical data and were further compared with principal component analysis (PCA)-BPNN models and BPNN models. The selection results of the LASSO method, which are superior to the results attained according to traditional knowledge and experience, quantitatively show that the production information varies in its effect on product yields and properties. The prediction results of a validation dataset analyzed by comparing model values with industrial values indicate that the intelligent LASSO-BP models have good prediction accuracy. Comparative results of the average relative errors show that the LASSO-BPNN models have the highest generalizability among LASSO-BP, PCA-BPNN and BPNN models. Sensitivity analysis results of feed carbon residue

ACS Paragon Plus Environment

1

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 2 of 43

content and operating conditions indicate that the combined LASSO-BP neural network models effectively predicted the yields and properties of MIP products and provided a good reference for industrial MIP process optimization.

1. INTRODUCTION The current economic, political, and regulatory climates place significant pressures on petroleum refiners to make significant investments.1 The use of advanced process engineering tools has become essential for refiners, not only for design but also in the tasks of process control, optimization, scheduling, and planning. Fluid catalytic cracking (FCC), producing gasoline, liquefied petroleum gas (LPG) and middle distillates, remains key in many refineries and takes a substantial percentage of profits. More than 400 FCC units are operated worldwide.2 Maximizing iso-paraffins (MIP) technology by dividing the riser reactor into two continuous parts, is one of fluid catalytic cracking process to reduce olefin content and increase iso-paraffin and aromatic contents in gasoline, increasing the liquid yield by 2–9% when compared to the traditional FCC plant.3 According to the reaction mechanism, high reaction temperature and short reaction time are employed in the first zone of the MIP riser reactor to crack the feedstock adequately, and the diameter of the second reaction zone is increased to create suitable conditions for reacting olefins.4 MIP technology has been applied to more than 50 sets of catalytic cracking units in China.5 Inferior crude oils and strict environmental regulations, such as China’s national standard VI for gasoline, bring significant challenges and attention to FCC technology, especially MIP technology, because there are less olefin components in gasoline. The FCC process, accompanied by complex chemical reactions and phase transitions,6 is a very large and complex industrial system with multiple variables, strong interference, large lag, and strong coupling.7 To apply process simulation and optimization technologies, it is important to determine whether the established


2

Page 3 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


process model can be accurate.8 Many researchers have proposed a variety of modeling methods, such as lumping kinetics modeling,9 molecular kinetics modeling,10 soft measurement modeling,11 neural network modeling,12 multi-scale computational fluid dynamics simulation.13 Research on reaction mechanism requires considerable manpower and material resources. The kinetics parameters of the model need to be adjusted with unit factors when they are used in the industrial unit.7 Moreover, the mechanism model is generally composed of algebraic equations, differential equations and partial differential equations.14 Therefore, the reaction model is complicated and requires significant calculation and analysis,15 which are time-consuming tasks to optimization and control. As a powerful data-driven computational tool, back-propagation neural network can capture the complicated underlying mechanism with high precision and thus serves as a powerful tool for classification problems.16 The BPNN model can further deal with complicated nonlinear problems and has good fault tolerance, the ability of strong associative memory and approximating any continuous function.17 A prediction model with 17 input variables based on generalized regression neural network and the adaptive boosting algorithm was proposed to predict the gasoline yield of an MIP process.18 A 19–24–4 back propagation (BP) neural network was established to predict the product distribution of an MIP unit using 19 input variables including feedstock and regenerated catalyst properties and operating variables.19 The input variables have a great impact on the accuracy and extrapolation capability of the model. In traditional BP-NN modeling for FCC processes, the determination of the input variables lacks relevant guiding theories and thus simply depends on personal knowledge and experience. Technological advances lead to increasingly larger quality-related datasets. In real-world production systems, causal relationships of resource inputs and production outputs are complicated


3


Page 4 of 43

and the collected data may be abnormal, which might cause unreliable results.20 Modern machine learning methods are increasingly applied to solve such problems.21-23 One such application involves variable selection (VS) methods for process modeling to reduce the complexity of the models without compromising accuracy.24 VS methods can identify a subset of variables that carries most of the relevant information contained in the complete dataset.25 Some common VS methods applied in industrial datasets are forward selection (FS), backward selection, data mining, partial least squares (PLS), principal component analysis (PCA), and clustering.26 Recently variable selection has been mainly carried out using the least absolute shrinkage and selection operator (LASSO) method.27 In this paper, a hybrid strategy is first proposed for FCC-MIP process modeling by integrating LASSO and BPNN for intelligent selection of production process data to ensure model accuracy and expand model functions to improve applications. The combined LASSO-BPNN models were trained and tested using industrial production historical data and further compared with PCABPNN models and BPNN models. Sensitivity analysis of the first riser outlet temperature, the second riser outlet temperature and the dense bed temperature of the regenerator on yield and product properties was investigated using the LASSO-BPNN models.

2. EXPERIMENTAL METHODS 2.1 Characteristics of MIP technology MIP technology has been widely used in newly built FCC plants in recent years. Derived from the traditional FCC process, MIP technology innovatively improved the riser reactor through its two-reaction-zone design. The diameter of the second reaction zone of the riser reactor is larger than that of the first reaction zone. The schematic diagram of MIP is shown in Figure 1.


4

Page 5 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


At the bottom of the riser, the preheated feed injection is lifted with the regenerated catalyst flow by the lift stream. The heat required to vaporize the feed and for the endothermic reactions is provided by the hot regenerated catalyst entering the riser. The outlet temperature of the first reaction zone is much lower than the inlet temperature. All cracking reactions take place in the ascendant flux. More endothermic cracking reactions than exothermic reactions occur in the first reaction zone of the MIP reactor. In the second reaction zone, exothermic reactions such as hydrogen transfer, isomerization and alkylation are enhanced. A large number of intermediate products such as olefins produced in cracking reactions, are converted to isoalkanes via olefin transformation reactions in the second reaction zone. Thus, the olefins content in the product is reduced and isoalkane content is increased without decreasing the gasoline yield of the MIP plants. Oil gas separated from the spent catalyst are further processed in the main fractionator. Spent catalyst after stripping with steam to remove remaining hydrocarbons is sent to the regenerator for regeneration by air combustion. The resulting flue gas is released at the top of the regenerator, and the regenerated catalyst is directed to the riser, where a new cracking cycle begins. 2.2 Data collection and processing In this study, the production data of a 1.2 million tons/year FCC industrial unit from July 2016 to March 2018 was collected and used for modeling. Sixty-eight data labels consisting of operating conditions, feed properties, and product flows and properties were established; data of 20 property items were collected from the Laboratory Information Management System (LIMS). The data in the Distributed Control System (DCS) were recorded every 10 seconds. The data of every 12-hour period were averaged. Data containing 48 measuring points in DCS were obtained. Whether the data are true affects the reliability of the model.


5


Page 6 of 43

FCC is a complex nonlinear system with dynamic changes. It is not only affected by feed flow rate, feedstock property structure and catalyst properties, but also by environmental factors (such as pressure, temperature and humidity.) and operating conditions. Generally, in the catalytic cracking unit with production in steady state, the flow rate of fresh feed in the riser determines the residence time of feedstock in the riser reactor, that is, the reaction time is determined and the conversion is affected. The preheating temperature of fresh feed affects the atomization of the feed oil and further influences the diffusion, adsorption and catalytic reaction of the feed oil molecules in the pore of the catalyst. Moreover, the feed temperature, outlet temperature of the risers and regeneration temperature are correlated. The distribution of catalyst in the prelift section is directly determined by the lift steam flow. Reaction temperature is a crucial and sensitive process parameter that directly affects the reaction rate and product distribution in the FCC process. In the MIP process, the outlet temperatures of the first riser and second riser reactors have an important effect on the chemical reactions in the two reactors. The reaction pressure is related to the catalytic cracking reaction. The stripping steam flow affects the stripping of hydrocarbons from the catalyst. The regenerator dense bed temperature balances the heat required for the reaction and the heat from coke burning in the regenerator, which determines the kinetic rate of coke combustion and the circulation rate of catalyst. Carbon residue is an important index for evaluating the coking tendency of the feedstock. Density is one of the most basic characteristics of oil. The composition of feedstock (saturates, aromatics, resins, asphaltenes) dictates its cracking performance. Sulfur content is also an important characteristic index. The contents of nickel and vanadium in the FCC feed are relatively high, and they can poison the catalysts to lose their activities. Microactivity evaluation is widely


6

Page 7 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


used in laboratories to evaluate catalytic cracking catalysts. The carbon content of the catalyst indicates that the regeneration effect of the catalyst will affect the catalyst activity. Neural network usually suffers from local minima, slow convergence, and the sensitive setting of the learning rate. Training a neural network usually takes a long time, mainly due to the large amount of points of the dataset. In addition, neural network can be inaccurate when limited training data are available.28 To develop efficient training methods with fast convergence and good generalizability, before training, the sample data usually needs to be preprocessed. The data processing process in this paper includes the following steps. 1) Data during instrument fault periods were eliminated. 2) Data with large fluctuations were deleted. 3) Twenty-two variables (of 68 data labels) were selected as the initial input variables of the models, including 10 operating variables, 10 feed properties and 2 catalyst characteristics. The 10 operating variables contained feed mass flow (FMF), feed temperature (FT), lift steam mass flow (LS), first riser outlet temperature (ROT1), second riser outlet temperature (ROT2), stripping steam mass flow (SSMF), reactor pressure (RP), pressure difference between the reactor and the regenerator (PDRR), regenerator dense bed temperature (RDBT) and air volume flow (AVF). The10 feed properties consisted of density, carbon residue content (CCR), the content of saturated hydrocarbons (SHs), aromatic hydrocarbons (AHs), resins (Rs) and asphaltenes (As), nickel content (Ni), vanadium content (V), sulfur content (Sulfur) and 50% recovered temperature (50% point). The 2 catalyst characteristics were regeneration catalyst carbon content (CC) and catalyst microactivity (CMAT). 4) The method used to calculate the product yield of dry gas, liquefied petroleum gas (LPG), gasoline and diesel oil is as follows:

The product yield =

product flow fresh feed flow + light sump oil flow + quenching oil flow

(1)


7


Page 8 of 43

Samples were not usually used directly for network training because they had different physical meanings and dimensions but were use in pretreatment for raw data. The experimental data contained some uncertain factors in neural network training.29 To eliminate the order of magnitude difference between inputs and output, normalization was used to preprocess each input and output before network learning. Therefore, input components of network training were on equal footing, i.e., inputs and outputs were constrained to certain intervals. The general formulas of normalization and anti-normalization are described below, respectively. 30 x  xmin  (b  a )  a xmax  xmin

(2)

X a  ( xmax  xmin )  xmin ba

(3)

X=

x=

where x is the raw data; X is the normalized data; xmax and xmin are the maximum and minimum values in the raw data; and a and b are the lower and upper limits of the normalized data range. In this study, all datasets were normalized to the interval [-1, 1]. 2.3 Cross-validation Cross-validation is a method of assessing the performance of a predictive model. Statistical analysis will generalize to an independent dataset. There are many types of cross-validation, such as repeated random sub-sampling validation, K-fold cross-validation, K×2 cross-validation, and leave-one-out cross-validation. In this study, we use K-fold cross-validation to select model parameters. K-fold cross-validation is a technique of dividing the original sample randomly into K sub-samples. Then, a single sub-sample is regarded as the validation data for testing the model, and the remaining K-1 sub-samples are used as training data. This process is repeated K times, and each of the K sub-samples is used exactly once as the validation data. The K results from the folds can then be averaged (or otherwise combined) to produce a single estimation.31


8

Page 9 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


An example of estimating a turning parameter γ with K-fold cross-validation as follows:32 Step 1: Divide the data into K roughly equal parts; Step 2: For i=1,2,3,…,K, fit the model with parameter γ to the other K-1 parts to calculate 𝛼̂ −𝑘 (𝛾) and compute its error, i.e., the cross-validation error, in predicting the kth part; 2

Ek (γ)= ∑i ϵ kth part [yi -xi α̂ -k (γ)]

(4)

Step 3. Do this for many values of γ and choose the value of γ that minimizes 𝐶𝑉(𝛾). In this paper, we set K=10. 1

CV(γ)= K ∑Ki=1 Ei (γ)

(5)

2.4 LASSO method Two possible phenomena, i.e., large calculation cost and overlearning, occur while reducing the dimensionality of complex datasets. Massive high-dimensional datasets can easily have a large calculation cost. However, small high-dimensional sample datasets can easily result in overlearning.33 Therefore, effectively selecting features of high-dimensional datasets is an urgent problem solved in the study of feature selection. In 1996, Tibshirani proposed a new method called the LASSO method to enhance the prediction accuracy and interpretability of the statistical model it produced.34 The LASSO method uses an absolute value penalty function of the feature scores to automatically compress feature scores with small absolute values to 0, thereby simultaneously selecting significant variables and estimating corresponding parameters.35 LASSO is a good way to overcome the shortcomings of the manual feature selection. Therefore, LASSO has received great attention in the field of regression and classification. The basic framework is summarized as follows. Consider a sample consisting of n cases, each of which consists of p features (also known as covariates) and a single outcome. Each feature vector contains p features. There may be


9


Page 10 of 43

collinearity and other correlations between the p features. If all influence factors are used as the input of the BP neural network directly, there are too many input variables, which will complicate the neural network structure, intensify network training, produce local minima, and lead to poor generalizability. Let yi be the outcome, xi =(x1, x2,...xp)T be the covariate vector for the ith case, and

β =(β1, β2,... βp)T. The objective of LASSO is to solve the optimization problem. The basic form of LASSO is as follows.

arg min  n  ( y   1



0 , 

n

i

i 1

0

  xiT  )2  

(6)

p

s.t.

 j 1

s

j

where  j is the feature score of the jth variable. s≥0 is a prespecified free parameter that determines the amount of regularization. The value of s can be from 0 to infinity. When the value of s is small, some variables with low correlation are compressed to 0, and these variables are deleted to achieve feature selection. When s is large enough, the constraint is no longer applied, all attributes are selected and a variable selection sequence is formed. Supposed that X represents the n × p covariates matrix, n is the number of samples, p is the number of covariates, and y represents a response vector of excepted outputs. Formula (5) can be written more compactly as follows.36

arg min 0 ,   p

1 y  0 I  X  T n

s. t.  It is known that the standard Lp norm is Z

1

2

(7)

 t

= (  i 1 Z i )1/ p . Equation (7) represents the square n

p

2

p

of the L2 norm, that is, the meaning of the square loss function.


10

Page 11 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60




Since  0  y  x T  

yi   0  xiT   yi  ( y  x T  )  xiT   ( yi  y )  ( xi  x )T 

(8)

Equation (7) is rewritten as follows. 2 1 arg minp  y  X  T 2     n 

s. t. 

1

(9)

 t



LASSO estimator  can be represented by the following. 1 L(  ,  )  minp  y  X  T    n

2 2

    1 

(10)

K-fold cross-validation is generally used to estimate the regularization parameter λ of the LASSO regression model. In this study, we adopt 10-foldcross-validation to estimate the regularization parameter λ. The constant parameter λ is prespecified by 10-fold cross validation according to Eq. (10). Therefore, the feature score vector β is estimated by continuously reducing the residual error until it is either small enough or less than or equal to a constant. There is an application example of LASSO method in reference 36. Based on LASSO, a novel ship fuel consumption prediction model is proposed, and its superiority was confirmed through comparison with some existing methods. 2.5 BP neural network With the ability to approximate any continuous function and nonlinear mapping, BP neural network is a kind of typical multilayer feed-forward neural network whose topological structure includes an input layer, a hidden layer and an output layer. The basic structure of BP neural network is shown in Figure 2. The neurons in the same layer have no connections while adjacent layers are fully connected. The mapping relationship between input and output is attained by the network


11


Page 12 of 43

studying and adjusting the connecting weight and threshold value among neurons according to the input and output of given samples. During the training process, information positive transmission and error back propagation essentially occur. Complicated nonlinear problems can be dealt with using the BP neural network model because of its good fault tolerance and strong associative memory. Therefore, to improve simulation and optimization of FCC process, BP neural network combined with the LASSO method, as shown in Figure 2, was used to predict the yields and properties of FCC products. The symbols x1, x2,..., xn are the original n variables. After LASSO feature variable selection, the remaining m variable variables are used as the input variables of BP neural network models y by eliminating the feature whose score is 0.

3. RESULTS AND DISCUSSION 3.1 Feature variable selection and LASSO-BP Models 3.1.1 Feature variable selection. In this paper, the abovementioned LASSO feature selection method was introduced to improve the prediction accuracy and interpretability of the regression models. By altering the model fitting process, a subset of the provided covariates for use in the final FCC models was selected rather than using all 22 initial model input variables determined by the data processing steps. The regularization parameter λ of LASSO on the training set (0.00037) was employed with the ten-fold cross-validation method. When the feature scores of the input variables were close to 0, the variable was considered to have little influence on the output and would be eliminated. The feature scores of variables that affect the gasoline are shown in Figure 3. In the bar chart, factor effects are represented by horizontal bars. It can be seen that there are ten feature variables that affect the gasoline yield, which are ranked from high to low as follows: Feed CCR, SSMF, Rs, RP, ROT1, Density, SHs, PDRR, CMAT and RDBT. The scores of feed CCR and SSMF are 0.28 and 0.13, respectively, while those of PDRR,


12

Page 13 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


CMAT and RDBT are the smallest (nearly 0.02). The results for the industrial MIP unit, which are different from simple experience, indicate feed CCR content and SSMF have a more significant influence on gasoline yield than ROT1 or ROT2. In general, a high feed CCR means low hydrogen content and high molecular weight (MW) and density of the feed. It will cause decreasing catalytic cracking to gasoline and additional coke formation.37 In the stripping zone, some steam is added to further crack and remove the heavy hydrocarbons from the catalyst surface.38 High temperature and appropriate mass flow of stripping steam relate to further cracking into small molecules and desorption of the hydrocarbons on the catalyst. The conclusion illustrates that the composition and properties of the FCC feed should be first considered for optimization when maximizing the gasoline yield as the production target. In Figure 4, the feature scores of variables for diesel yield, which are useful for comparing the factor effects, are presented. According to the chart, the most important factors are RP, FT and SSMF. RP has the highest direct proportional influence on diesel yield, while CMAT and RDBT affect yield weakly. It is also observed that some effects of LS, PDRR, AHs, Sulfur and SHs are relevant but to a lesser extent. There are great differences, up to 10%, in prediction of yield to products if pressure balance is included. The most important fact is that reaction rates depend on partial pressures, which are consequence of the total pressure in the riser.39 Similarly, the feature variables effecting the yield of other products such as acidic gas, dry gas and LPG, as well as gasoline olefin content and gasoline research octane number (RON), were screened by the LASSO method. The results are shown in Figure 5. By analyzing the feature variables with the LASSO method, the influence of operating conditions, feedstock properties and catalyst characteristics on the MIP industrial unit were further refined. The results show that operating conditions and feedstock properties affect the yield of


13


Page 14 of 43

products. The distribution of variable feature scores for different modeling items were distinguishable. The results are elaborated in Table 1. It can be seen that, for an industrial catalytic cracking unit, operating conditions and feed properties have different effects on the product yields and properties. Most of the variables affecting the yields of gasoline and diesel oil are the operating conditions, and the variables affecting the gasoline RON and diesel CN are largely comprised of the properties of feed and catalysts. Table 1 Number of feature variables among operating conditions, feedstock properties (B) and catalyst characteristics for the eight modeling items. Number of feature operating

feedstock

catalyst

variables

conditions

properties

characteristics

Acid gas yield

8

7

2

LS, V, Density

Dry gas yield

8

6

1

FT, AVF, SHs

LPG yield

3

5

1

Top 3

CMAT, PDRR, V CCR, Gasoline yield

5

4

SSMF,

1 Rs

Diesel yield

8

5

1

RP, FT, SSMF AHs,

Gasoline RON

0

6

50%

1 point, Rs RDBT, ROT1,

Gasoline olefin

6

3

2 CMAT Sulfur,

Diesel CN

2

5

CCR,

2 CC


14

Page 15 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


3.1.2 LASSO-BP Models The significant variables were selected from the variables with a large number of correlations in the high-dimensional space, the redundant variables were eliminated and the input of the models was distinctly simplified using the LASSO method. The neural network prediction models of the yields and properties of FCC products were established using the selected feature variables as input. Neural networks have many advantages compared to other iterative methods of learning, such as low computational time required for model training and the ability to select the number of hidden neurons. The number of hidden neurons plays a crucial role in the appropriate training of neural networks. Neural networks with a small hidden layer will not be capable of generalizing the training data (underfitting). However, if the network has a large number of hidden neurons, it can cause overfitting.40 Nevertheless, the number of hidden neurons in many cases significantly impacts on the performance of the model. Therefore, a method of selecting the architecture of neural networks is desirable. To choose the number of hidden nodes for each example, we train the networks with different numbers of hidden nodes using the 10-fold cross validation method, and then choose the number of hidden nodes that achieves the best validation error. The gasoline yield model of the MIP process was constructed using the MATLAB environment. The 265 processed sample datasets were divided into two groups: 215 sets of data were used for training and verification, and 50 sets of data were used as testing samples. A LASSO-BP neural network model with a 10-13-1 structure was established by selecting 13 hidden layer nodes after ten-fold cross-validation. Figure 6 shows the gasoline yield comparison between the model value and the industrial value in the training and validation data and the testing date. For the training and verification data, the relative error between the model value and the industrial value less than 6% is up to 94%. The average relative error of the model values and the industrial value is 2.91%. On


15


Page 16 of 43

the testing set, the average relative error of the model values and the industrial value is 2.72%. The results show that the prediction value of the gasoline yield model with 10 input variables has little error from the actual industrial value, and the trend change of the prediction result is very similar to that of the actual production data. The stabilized gasoline RON is an important index for measuring the quality of gasoline. For the gasoline RON model, 47 sets of data were used to train the network and validation and 25 sets of data were used as testing samples. Figure 7 shows the gasoline RON comparison of model value and industrial value on the training and validation data and the testing date. It can be seen from Figure 7 (a) and (b) that, in the training and verification sets, the predicted value of the model and the actual value are very close, and the relative error is mostly within 0.2% in the training and verification set and 0.18% in the testing set. Similarly, models of predicted acid gas yield, dry gas yield, LPG yield, diesel oil yield and gasoline olefin and diesel CN were established with industrial data by LASSO-BP method. Comparisons of model value and industrial value on testing data are shown in Figure 8. 3.2 Contrast analysis with PCA-BP model and BP model Principal component analysis (PCA) is a basic technique of data dimension reduction in chemometrics. PCA can extract the major relevant information of large data sets and describe it using only a few new variables, i.e., principal components (PCs) that are linear combinations of the original variables. 41 PCA decomposes the data matrix R into

R=TPT +E

(11)

where R is a nobj × nvar data matrix, E a nobj × nvar residual matrix, T a nobj × npc score matrix, and P a nvar × npc loading matrix. The loadings represent the principle components of the matrix, whereas the scores are the projections of the matrix onto the loadings.42


16

Page 17 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


The specific ability of PCA to determine correlations among variables is a prime motivation for contrast in this work. A combined PCA and BP neural network method (PCA-BPNN) is established to predict the product yields and quality of the MIP process with the same samples previously used for LASSO method. PCA is used to reduce the dimensionality of the factors influencing product yields and quality and eliminate correlations among factors. The cumulative contribution rate of a principal component is based on the number of principal components. To retain more information from the original data, the contribution rate is generally more than 85%. Thus, the original dataset can be adequately described using a few orthogonal principal components instead of the original variables with no significant loss of information. In this paper, principal components are obtained according to cumulative contribution rate of principal components (threshold >90%), as calculated by the method in reference17. Accordingly, the obtained principal components are used as BP neural network input vectors. To better reflect the advantages of the LASSO-BP method described in this article, the BPNN and PCA-BPNN methods were used to construct models of acid gas yield, dry gas yield, LPG yield, gasoline yield, diesel yield and gasoline RON, gasoline olefin and diesel CN. The model established by the LASSO-BP method is further compared with the PCA-BP and BPNN models. The relative errors between the model value of the three methods and the industrial value on the training dataset and the testing dataset are shown in Table 2. As seen from the simulation results of the eight items, compared with the BP and PCA-BP models, the LASSO-BP models generally have lower average relative errors on the training and validation data (ARE1) and average relative errors on the testing set (ARE2), which means better accuracy for industry FCC process prediction and a more satisfactory modeling method. Similarly, the results in Table 2 indicate that models established by PCA-BP are superior to that built by BPNN. Notably, PCA tends to select


17


Page 18 of 43

significantly fewer variables. In cases where it selects a similar amount, the performance is not noticeably worse than that of the LASSO method; an example is the LPG yield: when PCA selects 8 components, while LASSO selects 9 variables, the performance difference is 2.68% (PCA) vs. 2.42% (LASSO). Table 2 Model comparison of ARE1 and ARE2 on BPNN, PCA-BPNN and LASSO-BPNN. (Network structure: Initial input variables–feature variables–hidden layers–output) Models

Acid gas yield

Dry gas yield

LPG yield

Gasoline yield

Diesel yield

Gasoline RON

ARE1 (%)

ARE2 (%)

Network structure

BP

11.82

12.31

22-22-31-1

PCA-BP

9.42

9.25

22-6-25-1

LASSO-BP

9.19

9.62

22-17-19-1

BP

8.12

8.27

22-22-28-1

PCA-BP

7.28

7.64

22-6-25-1

LASSO-BP

6.79

7.03

22-14-29-1

BP

4.38

4.31

22-22-33-1

PCA-BP

2.52

2.68

22-8-32-1

LASSO-BP

2.05

2.42

22-9-27-1

BP

6.71

6.58

22-22-13-1

PCA-BP

3.87

4.26

22-6-16-1

LASSO-BP

2.91

2.72

22-10-13-1

BP

10.22

9.84

22-22-26-1

PCA-BP

7.84

8.38

22-6-24-1

LASSO-BP

5.19

5.42

22-13-29-1

BP

0.99

1.14

22-22-30-1


18

Page 19 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Gasoline olefin

Diesel CN

PCA-BP

0.78

0.93

22-5-30-1

LASSO-BP

0.16

0.18

22-7-19-1

BP

6.06

6.35

22-22-33-1

PCA-BP

4.79

5.02

22-3-33-1

LASSO-BP

1.93

2.77

22-11-22-1

BP

0.68

0.74

2222-26-1

PCA-BP

0.57

0.53

22-6-26-1

LASSO-BP

0.36

0.41

22-9-18-1

3.3 Sensitivity analysis of operating variables To further verify the prediction performance of the model, the effect of representative variables among operating variables and feed and catalyst properties on the yields and properties of products was investigated. Reaction temperature (ROT1) simulations were carried out at various temperatures (490–530°C with a 4°C interval) by increasing ROT2 with ROT1 and other parameters held constant. For the outlet temperature of the second riser (ROT2), the simulation of executable ROT2 (with a 2°C interval, 6–20 °C lower than ROT1 (at 520°C)) was studied to investigate the effect of ROT2 on the yield of dry gas, LPG, gasoline and gasoline olefin content. For the dense bed temperature of the regenerator (RDBT), different RDBT values (660–720°C with a 6°C interval) were tested to show the effect of RDBT on the yields and properties of products while keeping the other parameters constant. For carbon residue content of feed (CCR), various runs of feed CCR (1.0–5.0% with a 0.4% interval) were carried out to determine the effect of feedstock coking on the yields and properties of products; other parameters were kept constant.


19


Page 20 of 43

Figure 9 shows the model prediction of the product yield and property distribution of the FCC MIP reaction as a function of ROT1. As the temperature increased from 490°C to 530°C, the yields of dry gas and LPG increased with increased temperature and that of diesel oil decreased. The yield of gasoline increases initially and decreases thereafter; the maximum yield is achieved at 522°C, and there is slight difference in the gasoline yields at 510°C and 522°C. The diesel CN decreases with increasing temperature, while gasoline RON is improved. Such behavior can be attributed to the fact that, at high temperature, reaction rate constants are effectively influenced, resulting in an increase in feed conversion depending on the Arrhenius equation, where the rate constants are a function of reaction temperature and activation energy. An increase in reaction temperature leading to accelerate intermolecular motion results in a higher conversion of reactants to new components. Better feed vaporization can be obtained at high temperatures. The diffusion of feed compositions can be enhanced under higher temperatures. With an increase in reaction temperature, over-cracking of gasoline occurs. Clearly, there is a good match between FCC reaction mechanism knowledge and model prediction, which is supported by the work of Jarullah, A. T., et al.43 The reaction temperature in the first riser should not be too high or too low. An appropriate temperature should be controlled to obtain higher gasoline yield and lower gas and coke yields. In Figure 10, a comparison of ROT2 is presented. As shown, the yields of dry gas and LPG significantly increase from 3.8 to 4.1% and from 12.8 to 13.5%, respectively, when increasing ROT2 from 500 to 514°C. The yield of gasoline decreases from 55.5 to 54.6%. The olefin content in gasoline decreases due to the intensification of the olefin cracking reactions, which is consistent with the work of Ivanchina, E., et al.44 More endothermic cracking reactions than exothermic reactions occur in the first reaction zone of the MIP reactor, resulting in a much lower outlet


20

Page 21 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


temperature. In the second reaction zone, lower temperature inhibits endothermic reactions such as catalytic cracking, thermal cracking and dehydrogenation, while exothermic reactions such as hydrogen transfer, isomerization and alkylation are enhanced because a larger diameter causes the oil–gas mixture to stay in the second zone longer. In the industrial catalytic cracking unit, the temperature in the second reaction zone can be controlled flexibly by injecting the cold flow or hot catalyst to reach a desired reaction temperature. The temperature is decreased to produce FCC gasoline with lower olefin content by adding the cooled regenerated catalyst, spent catalyst, cold feed or other measures. More low-carbon olefins (e.g., ethylene, propylene and butene) are produced when the temperature is increased by supplementing hot regeneration catalyst. Figure 11 illustrates the product yield and property distribution as a function of RDBT. A very good match between the production knowledge and model prediction is observed in this figure. Increasing RDBT from 660°C to 720°C (the actual operating data of RBDT is 680–710°C), the yield and CN of diesel oil increase. Opposite phenomena can be observed for the yields of dry gas, LPG and gasoline. By increasing RDBT, the circulation rate of catalyst would be reduced to control the reaction temperature while keeping the feed mass flow constant. In other words, as the catalyst-to-oil ratio decreased, the concentration of the catalyst will decrease, decreasing the reaction rate of primary and secondary cracking. Such behavior weakens the total number of molecules cracked on the catalyst surface, increasing the amount of heavy components generated, such as diesel and coke. Moreover, decreasing the catalyst-to-oil ratio decreases the catalyst active site concentration, which contributes to the reaction rates of cracking reactions.45 Changing RDBT relates to coke combustion, flue gas combustion, regeneration catalyst microactivity, air volume flow and reaction temperature. Thus, a suitable RDBT positively affects FCC production.


21


Page 22 of 43

Figure 12 illustrates the product distribution of the FCC reaction at several feed CCR contents, where the yield and CN of diesel increase by increasing the feed CCR. The yields of dry gas, LPG and gasoline decrease as a result of increasing the feed CCR content, while gasoline RON has no significant change. The conversion of feed decreased and CCR content increased, indicating that the presence of more heavy components in the feed increases the possibility of coking by polycondensation, which inhibits catalyst activity by reducing the chemical reactions. In general, a high CCR causes additional coke formation and it creates unit processability requirements. In a commercial unit, a feed is preferable, which will give a higher conversion change with a small change in coke yield. 11, 37

4. CONCLUSIONS In this study, BP neural network technology coupled with a variable selection technique, i.e., the LASSO method, was introduced to model the relationships among 5 products yields, 3 product properties and 22 input variables of operating conditions, feed properties and catalyst properties to identify how some production problems (e.g. maximizing iso-paraffins, or producing more gasoline) regarding an industrial MIP unit is affected by an input set of process variables. The results show that feature variable selection using the LASSO method is an efficient statistical technique in the screening of significant factors for deterministic process models. Approximately ten variables, such as feed carbon residue content, stripping steam flow rate, and reactor pressure, strongly affect gasoline yield. Similarly, the feature variables of predicting the yield of other products, such as acidic gas, dry gas and LPG, as well as product quality, such as gasoline olefin, gasoline RON and diesel CN, can be selected intelligently using simple empirical knowledge. The distribution of the feature scores of variables demonstrates that the influences of operating conditions, feedstock properties, and catalyst characteristics on the modeling items of


22

Page 23 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


the MIP industrial unit differ. The main variables affecting the yields of gasoline and diesel oil are the operating conditions, and the main variables affecting the gasoline RON and diesel CN are the properties of feed and catalysts. With selected feature variables, simplified models were supplied to predict the yields and properties of the products of the MIP industrial unit. Comparison results with sufficient industrial data indicate the applicability of the method to MIP process simulation in a simple way: there is a good accuracy of less than 5% average relative error between model value and industrial data for most samples. The results of the average relative error contrast with the results of the models established by the BPNN and PCA-BPNN methods, reflecting the advantage of the LASSO-BP model described in this article. The dry gas and LPG yields and gasoline RON increased significantly with increasing riser temperature, while the yield and CN of the diesel oil decreased. The yields of gasoline increased initially and decrease thereafter with the maximum yield at ROT1 of 522°C. By increasing RDBT or feed CCR, the yield and CN of light diesel oil increased. Opposite phenomena can be observed for the yields of dry gas, LPG and gasoline. The effect of RDBT or feed CCR on gasoline RON is less pronounced. Several sensitivity tests allow for the conclusion that the results are in a good agreement with what is commonly encountered in the literature or based on the process mechanism. Therefore, the complicated relationship between the yields and properties of FCC products and the production information, key operating conditions, feedstock properties and catalyst characteristics can be identified comprehensively using these simplified models, which could be used in simulation and optimization applications for the MIP industrial unit because of good process representation.

Corresponding Author *Weimin. Zhong. E-mail: [email protected]; Tel: +86 021 64252440.


23


Page 24 of 43

ACKNOWLEDGMENTS This work was supported by the Major Projects National Natural Science Foundation of China (61333010), the National Natural Science Foundation Youth Project (21506050) and the Fundamental Research Funds for the Central Universities (222201814038).

REFERENCES (1)

Pashikanti, K.; Liu, Y. A. Predictive Modeling of Large-Scale Integrated Refinery Reaction

and Fractionation Systems from Plant Data. Part 3: Continuous Catalyst Regeneration (CCR) Reforming Process. Energy Fuels 2011, 25 (11), 5320-5344. (2)

Pinheiro, C. I. C.; Fernandes, J. L.; Domingues, L.; Chambel, A. J. S.; Graca, I.; Oliveira,

N. M. C.; Cerqueira, H. S.; Ribeiro, F. R. Fluid Catalytic Cracking (FCC) Process Modeling, Simulation, and Control. Ind. Eng. Chem. Res. 2012, 51 (1), 1-29. (3)

Liu, J. C.; Chen, H.; Pi, Z. P.; Liu, Y. F.; Sun, H.; Shen, B. X. Molecular-Level-Process

Model with Feedback of the Heat Effects on a Complex Reaction Network in a Fluidized Catalytic Cracking Process. Ind. Eng. Chem. Res. 2017, 56 (13), 3568-3577. (4)

Duan, L. W.; Zhang, L.; Sun, P.; Weng, H. X. Study on Modeling for a New FCC

Technique-MIP Riser Reactor. China Pet Process Pe 2012, 14 (1), 50-55. (5)

XU, Y. H. Advance in China Fluid Catalytic Cracking(FCC) Process. SciSn 2014, 44 (1),

13. (6)

Shah, M. T.; Utikar, R. P.; Pareek, V. K.; Evans, G. M.; Joshi, J. B. Computational Fluid

Dynamic Modelling of FCC Riser: A Review. Chem. Eng. Res. Des. 2016, 111, 403-448. (7)

Amblard, B.; Singh, R.; Gbordzoe, E.; Raynal, L. CFD Modeling of the Coke Combustion

in an Industrial FCC Regenerator. Chem. Eng. Sci. 2017, 170, 731-742. (8)

John, Y. M.; Patel, R.; Mujtaba, I. M. Modelling and Simulation of an Industrial Riser in


24

Page 25 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Fluid Catalytic Cracking Process. Comput. Chem. Eng. 2017, 106, 730-743. (9)

Naik, D. V.; Karthik, V.; Kumar, V.; Prasad, B.; Garg, M. O. Kinetic Modeling for Catalytic

Cracking of Pyrolysis Oils with VGO in a FCC Unit. Chem. Eng. Sci. 2017, 170, 790-798. (10)

Cruz, P. L.; Montero, E.; Dufour, J. Modelling of Co-Processing of HDO-Oil with VGO in

a FCC Unit. Fuel 2017, 196, 362-370. (11)

Dasila, P. K.; Choudhury, I. R.; Saraf, D. N.; Kagdiyal, V.; Rajagopal, S.; Chopra, S. J.

Estimation of FCC Feed Composition from Routinely Measured Lab Properties through ANN Model. Fuel Process. Technol. 2014, 125, 155-162. (12)

Wang, Z. W.; Yang, B. L.; Chen, C.; Yuan, J.; Wang, L. Y. Modeling and Optimization for

the Secondary Reaction of FCC Gasoline Based on the Fuzzy Neural Network and Genetic Algorithm. Chem. Eng. Process. 2007, 46 (3), 175-180. (13)

Lu, B.; Wang, W.; Li, J.; Wang, X.; Gao, S.; Lu, W.; Xu, Y.; Long, J. Multi-Scale CFD

Simulation of Gas–Solid Flow in MIP Reactors with a Structure-Dependent Drag Model. Chem. Eng. Sci. 2007, 62 (18), 5487-5494. (14)

Li, Y.; Chu, J. Z.; Zhang, J. R. A New Generic Reaction for Modeling Fluid Catalytic

Cracking Risers. Chin. J. Chem. Eng. 2017, 25 (10), 1449-1460. (15)

Amikiya, A. E.; Banda, M. K. Modelling and Simulation of Reactive Transport Phenomena.

Journal of Computational Science 2018, 28, 155-167. (16)

Li, J. C.; Zhao, D. L.; Ge, B. F.; Yang, K. W.; Chen, Y. W. A Link Prediction Method for

Heterogeneous Networks Based on BP Neural Network. PhyA 2018, 495, 1-17. (17)

He, F.; Zhang, L. Prediction Model of End-Point Phosphorus Content in Bof Steelmaking

Process Based on PCA and BP Neural Network. Journal of Process Control 2018, 66, 51-58. (18)

Jiang, H.; Tang, J.; Ouyang, F. A New Method for the Prediction of the Gasoline Yield of


25


Page 26 of 43

the MIP Process. Pet. Sci. Technol. 2015, 33 (20), 1713-1720. (19)

Ouyang, F. S.; Zhang, J. H.; Fang, W. G. Optimizing Product Distribution in the Heavy Oil

Catalytic Cracking (MIP) Process. Pet. Sci. Technol. 2017, 35 (13), 1315-1320. (20)

Silva, A. F.; Vercruysse, J.; Vervaet, C.; Remon, J. P.; Lopes, J. A.; De Beer, T.; Sarraguça,

M. C. In-Depth Evaluation of Data Collected During a Continuous Pharmaceutical Manufacturing Process: A Multivariate Statistical Process Monitoring Approach. J. Pharm. Sci. 2018. DOI: https://doi.org/10.1016/j.xphs.2018.07.033 (21)

Spesivtsev, P.; Sinkov, K.; Sofronov, I.; Zimina, A.; Umnov, A.; Yarullin, R.; Vetrov, D.

Predictive Model for Bottomhole Pressure Based on Machine Learning. Journal of Petroleum Science and Engineering 2018, 166, 825-841. (22)

Ahmad, T.; Chen, H.; Huang, R.; Yabin, G.; Wang, J.; Shair, J.; Azeem Akram, H. M.;

Hassnain Mohsan, S. A.; Kazim, M. Supervised Based Machine Learning Models for Short, Medium and Long-Term Energy Prediction in Distinct Building Environment. Energy 2018, 158, 17-32. (23)

Liu, L.; Silva, E. A.; Wu, C.; Wang, H. A Machine Learning-Based Method for the Large-

Scale Evaluation of the Qualities of the Urban Environment. Comput. Environ. Urban Syst. 2017, 65, 113-125. (24)

Cai, J.; Luo, J.; Wang, S.; Yang, S. Feature Selection in Machine Learning: A New

Perspective. Neurocomputing 2018, 300, 70-79. (25)

Mohan, S.; Buchanan, B. R.; Wollenberg, G. D.; Igne, B.; Drennen, J. K.; Anderson, C. A.

Variable Selection Optimization for Multivariate Models with Polar Qualification System. Chemometrics Intellig. Lab. Syst. 2018, 180, 1-14. (26)

Sanchez-Pinto, L. N.; Venable, L. R.; Fahrenbach, J.; Churpek, M. M. Comparison of


26

Page 27 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Variable Selection Methods for Clinical Predictive Modeling. Int. J. Med. Inf. 2018, 116, 10-17. (27)

Peres, F. A. P.; Fogliatto, F. S. Variable Selection Methods in Multivariate Statistical

Process Control: A Systematic Literature Review. Comput Ind Eng 2018, 115, 603-619. (28)

Bataineh, M.; Marler, T. Neural Network for Regression Problems with Reduced Training

Sets. Neural Networks 2017, 95, 1-9. (29)

Jing, G.; Du, W.; Guo, Y. Studies on Prediction of Separation Percent in Electrodialysis

Process Via BP Neural Networks and Improved BP Algorithms. Desalination 2012, 291, 78-93. (30)

Zheng, D.; Qian, Z.-d.; Liu, Y.; Liu, C.-b. Prediction and Sensitivity Analysis of Long-

Term Skid Resistance of Epoxy Asphalt Mixture Based on GA-BP Neural Network. Construction and Building Materials 2018, 158, 614-623. (31)

Rohani, A.; Taki, M.; Abdollahpour, M. A Novel Soft Computing Model (Gaussian Process

Regression with K-Fold Cross Validation) for Daily and Monthly Solar Radiation Forecasting (Part: I). Renewable Energy 2018, 115, 411-422. (32)

Jiang, P.; Chen, J. Displacement Prediction of Landslide Based on Generalized Regression

Neural Networks with K-Fold Cross-Validation. Neurocomputing 2016, 198, 40-47. (33)

Sheikholeslami, R.; Razavi, S.; Gupta, H. V.; Becker, W.; Haghnegahdar, A. Global

Sensitivity Analysis for High-Dimensional Problems: How to Objectively Group Factors and Measure Robustness and Convergence While Reducing Computational Cost. Environ. Model. Software 2019, 111, 282-299. (34)

Tibshirani, R. Regression Shrinkage and Selection Via the Lasso: A Retrospective. J R Stat

Soc B 2011, 73, 273-282. (35)

Sermpinis, G.; Tsoukas, S.; Zhang, P. Modelling Market Implied Ratings Using Lasso

Variable Selection Techniques. J. Empirical Finance 2018, 48, 19-35.


27


(36)

Page 28 of 43

Wang, S.; Ji, B.; Zhao, J.; Liu, W.; Xu, T. Predicting Ship Fuel Consumption Based on

Lasso Regression. Transportation Research Part D: Transport and Environment 2018, 65, 817824. (37)

Lappas, A. A.; Iatridis, D. K.; Papapetrou, M. C.; Kopalidou, E. P.; Vasalos, I. A. Feedstock

and Catalyst Effects in Fluid Catalytic Cracking – Comparative Yields in Bench Scale and Pilot Plant Reactors. Chem. Eng. J. 2015, 278, 140-149. (38)

Pashikanti, K.; Liu, Y. A. Predictive Modeling of Large-Scale Integrated Refinery Reaction

and Fractionation Systems from Plant Data. Part 2: Fluid Catalytic Cracking (FCC) Process. Energy Fuels 2011, 25 (11), 5298-5319. (39)

Jiménez-García, G.; Aguilar-López, R.; Maya-Yescas, R. The Fluidized-Bed Catalytic

Cracking Unit Building Its Future Environment. Fuel 2011, 90 (12), 3531-3541. (40)

Henríquez, P. A.; Ruz, G. A. A Non-Iterative Method for Pruning Hidden Neurons in Neural

Networks with Random Weights. Applied Soft Computing 2018, 70, 1109-1121. (41)

Stanimirovic, O.; Hoefsloot, H. C. J.; De Bokx, P. K.; Smilde, A. K. Variable Selection

Methods as a Tool to Find Sensor Locations for Distributed Parameter Systems. Ind. Eng. Chem. Res. 2008, 47 (4), 1184-1191. (42)

Brewer, J.; Huang, Z.; Singh, A. K.; Misra, M.; Hahn, J. Sensor Network Design Via

Observability Analysis and Principal Component Analysis. Ind. Eng. Chem. Res. 2007, 46 (24), 8026-8032. (43)

Jarullah, A. T.; Awad, N. A.; Mujtaba, I. M. Optimal Design and Operation of an Industrial

Fluidized Catalytic Cracking Reactor. Fuel 2017, 206, 657-674. (44)

Ivanchina, E.; Ivashkina, E.; Nazarova, G. Mathematical Modelling of Catalytic Cracking

Riser Reactor. Chem. Eng. J. 2017, 329, 262-274.


28

Page 29 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


(45)

Quintana-Solórzano, R.; Thybaut, J. W.; Galtier, P.; Marin, G. B. Simulation of an

Industrial Riser for Catalytic Cracking in the Presence of Coking Using Single-Event Microkinetics. Catal. Today 2010, 150 (3), 319-331.


29


Page 30 of 43

FIGURE CAPTIONS Figure 1. Reactor and regenerator unit of the MIP process. Figure 2. Basic structure of the BP neural network in combination with LASSO method. Figure 3. Feature scores of variables (>0) upon gasoline yield. Figure 4. Feature scores of variables (>0) affect the diesel yield. Figure 5. Feature scores of variables (>0) on acid gas yield, dry gas yield, LPG yield, gasoline RON, gasoline olefin and diesel CN. Figure 6. Gasoline yield comparison of model value and industrial value on (a) training and verification data, (b) testing data. Figure 7. Gasoline RON comparison of model value and industrial value on (a) training and verification data, (b) testing data. Figure 8. Comparisons of model value and industrial value on testing data. Figure 9. The product yield and property distribution as a function of ROT1. Figure 10. Product yield and property distribution as a function of ROT2. Figure 11. Product yield and property distribution as a function of RDBT. Figure 12. Product distribution of FCC reaction at several feed CCR content.


30

Page 31 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 1. Reactor and regenerator unit of the MIP process 84x47mm (600 x 600 DPI)



Figure 2. Basic structure of the BP neural network in combination with LASSO method 84x47mm (600 x 600 DPI)


Page 32 of 43

Page 33 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 3. Feature scores of variables (0) upon gasoline yield 84x59mm (600 x 600 DPI)



Figure 4. Feature scores of variables (0) affect the diesel yield 84x59mm (600 x 600 DPI)


Page 34 of 43

Page 35 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 5. Feature scores of variables (0) on acid gas yield, dry gas yield, LPG yield, gasoline RON, gasoline olefin and diesel CN 145x101mm (600 x 600 DPI)



Figure 6. Gasoline yield comparison of model value and industrial value on (a) training and verification data, (b) testing data 84x59mm (600 x 600 DPI)


Page 36 of 43

Page 37 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 7. Gasoline RON comparison of model value and industrial value on (a) training and verification data, (b) testing data 84x59mm (600 x 600 DPI)



Figure 8. Comparisons of model value and industrial value on testing data 145x101mm (600 x 600 DPI)


Page 38 of 43

Page 39 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 9. Product yield and property distribution as a function of ROT1 84x59mm (600 x 600 DPI)



Figure 10. Product yield and property distribution as a function of ROT2 84x59mm (600 x 600 DPI)


Page 40 of 43

Page 41 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 11. Product yield and property distribution as a function of RDBT 84x59mm (600 x 600 DPI)



Figure 12. Product distribution of FCC reaction at several feed CCR content 84x59mm (600 x 600 DPI)


Page 42 of 43

Page 43 of 43 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


For Table of Contents Only 84x47mm (600 x 600 DPI)


Hybrid strategy integrating variable selection and a neural network for

Recommend Documents