CN102880905B - Online soft measurement method for normal oil dry point - Google Patents

Abstract

The invention relates to an online soft measurement method of a common-roof oil dry point; selecting and constructing key variables influencing the constant top oil dry point, namely constant pressure tower top temperature, constant pressure tower top pressure, constant pressure tower top reflux temperature, constant pressure tower top circulating brought energy, extraction ratio and constant pressure tower side line extraction total amount as input variables of a neural network model, taking the constant top oil dry point as corresponding output, establishing a soft instrument of the constant top oil dry point based on an artificial neural network model, correcting the model output through an artificial analysis value, and realizing real-time monitoring and control on the constant pressure tower top product quality; the method can guide an operator to adjust the operation process in time, and the dry point estimation value given by the soft instrument can provide a basis for process control, so that the optimization of the operation process is realized, and the crude oil extraction rate of the atmospheric and vacuum distillation unit is improved.

Description

Online soft measurement method for normal oil dry point

Technical Field

The invention belongs to the cross field of petroleum refining and intelligent control, and relates to an online soft measurement method for a normal pressure tower product-normal top oil dry point in a normal pressure and reduced pressure device in a petroleum refining process.

Background

The atmospheric tower is used as the main equipment of an atmospheric and vacuum device, processes crude oil from a primary distillation tower in a large amount, and consumes a large amount of energy. Therefore, the efficient and stable operation of the atmospheric tower is ensured, and the key points of energy conservation and consumption reduction of the atmospheric and vacuum device and improvement of the crude oil extraction rate are achieved.

The normal-top oil dry point is a main quality control index of a normal-top product and mainly reflects the light weight of an oil product to be produced, so the control quality of the normal-top oil dry point is not only related to the extraction rate of the crude oil in the normal pressure tower, but also influences the subsequent processing process. Currently, for dry spots, there is no suitable meter that can give measurements in real time, and most refineries rely on laboratory manual analysis of the values. The period of manual analysis is generally once every 4 hours or 8 hours, each analysis requires nearly two hours, and then the operation condition is adjusted according to the result of the manual analysis, so that the delay is large, the produced product is often too heavy and unqualified, or the crude oil extraction rate is finally influenced in order to ensure the product quality.

To address the above-mentioned problems, in process control, a number of soft-measurement methods have been developed that utilize various mathematical models to estimate the dry point values. And selecting variables closely related to the constant-top oil dry point value as auxiliary variables of the model, and then estimating the dry point value in real time according to the model input. The models used in soft measurement are numerous and can be roughly classified into three categories, namely mechanism models, statistical models and intelligent modeling methods. The mechanism model method and the statistical model method are traditional modeling methods, and have the common defect that the model precision is not high enough because the highly nonlinear system of the atmospheric and vacuum device cannot be accurately described. The intelligent modeling technology developed in recent years provides a new idea for nonlinear system modeling. The intelligent modeling method comprises an artificial neural network, a support vector machine and the like, and has the common characteristics of self-learning and self-organizing capabilities, and can give more accurate description to a nonlinear system, so that the method is widely applied to process modeling.

At present, in a soft measurement method for an atmospheric tower, no matter a traditional modeling method or intelligent modeling, the following problems exist:

(1) the calculation accuracy of the established model is gradually reduced along with time due to the frequent changes of the crude oil properties and the operation conditions, so that the effect exerted by the soft measurement model is limited.

(2) The atmospheric and vacuum distillation device is a highly nonlinear system, the variable coupling is serious, the number of selected variables in soft measurement modeling is often large, the variables are correlated with each other, the number of redundant variables is large, and the output stability of the model is reduced.

The method comprises the steps of firstly adopting AspenPlus to carry out flow simulation on the normal pressure tower to simulate the actual operating condition of the normal pressure tower, then carrying out sensitivity analysis on key operating variables of the normal pressure tower about a constant top oil dry point, selecting variables which have large influence on the constant top oil dry point as a candidate set of soft measurement model auxiliary variables according to the result of the sensitivity analysis, then carrying out correlation analysis on the candidate auxiliary variables based on the petroleum rectification principle, removing the redundant variables, and finally determining the input variables of the model. And according to the determined auxiliary variables, acquiring on-site historical production data, removing abnormal data and forming a training sample set. And then, training a 3-layer feedforward neural network by using a BP algorithm, and outputting model parameters after the error meets certain requirements.

Disclosure of Invention

The invention aims to provide an artificial neural network soft measurement method for a normal-pressure tower normal-top oil dry point. First, by Aspenplus sensitivity analysis, the temperature at the top of the atmosphere (x) was selected₁DEG C), normal top pressure (x)₂Kpa), constant top reflux temperature (x)₃DEG C), the energy (x) of the normal top circulation₄DEG C. t/h), production ratio (x)₅) And side line total amount (x) of atmospheric tower₆T/h) as input variables for the neural network model. Wherein x₁，x₂，x₃Can be measured directly by a meter, x₄，x₅，x₆The method can be obtained by calculation on the basis of the measured value of the instrument, and the manual analysis value of the atmospheric tower atmospheric top oil dry point at the previous day t' is input through a human-computer interface of the distributed control system; and correcting the output value of the model through online calculation of the model according to the relative error between the output value of the model and the input manual analysis value at the last day t', so as to realize online giving of the estimated value of the constant top oil dry point.

1. Selection or construction of auxiliary variables of soft measurement model

The operation variables of the atmospheric tower are many, and all the variables influence each other and are seriously coupled. The influencing variables for the common top oil dry point are: normal top oil extraction amount, normal first line oil extraction amount, normal second line extraction amount, normal third line extraction amount, normal fourth line extraction amount, normal pressure tower total feeding amount, feeding temperature, heating furnace outlet temperature, normal top pressure, normal top circulation heat extraction amount, normal first heat extraction amount, normal second heat extraction amount, normal first line stripping heating load, normal second line stripping steam amount, normal third line stripping steam amount and normal pressure tower kettle stripping steam amount. The accurate calculation of variables such as a constant top circulation heat extraction quantity, a constant one heat extraction quantity, a constant two heat extraction quantity, a constant one-line steam stripping heating load and the like relates to the physical property data of the crude oil, and the heat extraction quantity is simply represented by the product of heat exchange temperature difference and flow because the property of the crude oil changes frequently and accurate physical property data cannot be obtained in time. Sensitivity analysis is performed on the above variables one by Aspenplus, and as shown in fig. 1, the sensitivity analysis result of the outlet temperature of the heating furnace on the constant top oil dry point is given, and according to this example, sensitivity analysis can be performed on the other variables one by one. And selecting the operating variables which have obvious influence on the dry points as a candidate set of auxiliary variables of the soft measurement model according to the sensitivity analysis result.

The invention has the characteristics that:

(1) sensitivity analysis is carried out on the operation variables by utilizing Aspenplus, variables which have obvious influence on the common-top oil dry point are selected as candidate sets of auxiliary variables of the soft measurement model, and then redundant variables are removed based on the petroleum rectification principle to determine the auxiliary variables of the model.

(2) Three auxiliary variables of the energy brought out by the constant top circulation, the extraction ratio and the side line extraction amount of the normal pressure tower are constructed to respectively represent the heat taking amount of the constant top circulation, the extraction ratio of the constant top oil and the sum of the extraction amounts of the four side lines of the normal pressure tower.

(3) And correcting the soft measurement model output of the next day by utilizing the manual analysis value of the time t' of the day before the constant oil-top dry point so as to solve the problem of gradual reduction of the model precision caused by frequent change of crude oil properties.

Therefore, auxiliary variables selected by the normal-pressure tower normal-top oil dry point neural network soft measurement model are as follows:

① ordinary top temperature (x)₁，℃)

② Normal pressure (x)₂，Kpa)

③ constant top reflux temperature (x)₃，℃)

④ constant top cycle energy (x)₄)

⑤ production ratio (x)₅)

⑥ side line output (x) of atmospheric tower₆，t/h)

Wherein, the calculation formulas of the normal top circulating brought-out energy, the extraction ratio and the normal pressure tower side line extraction amount are constructed as follows:

x₄＝F₁×(t₁-t₂)(1)

x₅＝F₂/F(2)

x₆＝F₃+F₄+F₅+F₆(3)

wherein, F₁Is the constant top circulation flow (t/h), t₁、t₂The temperature (. degree. C.) of the withdrawal and return of the atmospheric top cycle, F₂The normal top gasoline extraction amount (t/h), F the normal pressure tower feeding amount (t/h), F₃，F₄，F₅，F₆Respectively the normal first line, the normal second line, the normal third line and the normal fourth line.

2. Data pre-processing

Collecting historical production data of the atmospheric tower, including x, according to the auxiliary variables of the model determined above₁，x₂，x₂And calculating x₄，x₅，x₆The required variables are: constant top circulation flow F₁The withdrawal and return temperatures t of the common top cycle₁、t₂Constant top gasoline extraction F₂The feeding quantity F of the atmospheric tower and the production quantities F of the first line, the second line, the third line and the fourth line₃，F₄，F₅，F₆And the dry point manual analysis value at the corresponding moment. The collected data range should cover a wide operating conditionMeanwhile, data under the conditions of starting, stopping and abnormal working conditions are avoided to be collected, and gross errors of the collected data are removed and detected to form a data sample set.

Because the magnitude of each input variable of the model is greatly different, the input variables of the model are normalized according to the following formula:

${\mathrm{sx}}_i=\frac{x_i-x_i^{\min }}{x_i^{\max }-x_i^{\min }}(b-a)+a,i=\mathrm{1,2},...,6---\left(4\right)$

in the formula, x_iIs a measured value of the ith input variable, sx, of the neural network model_iRepresenting the input value of the neural network after the ith input variable is subjected to normalization processing,andrepresenting the lower limit and the upper limit of the variation range of the ith acquired input variable, and a and b representing the lower limit and the upper limit of the input variable after normalization processingHere, b is 1 and a is 1.

The output variables of the neural network model are normalized using the following equation:

$\mathrm{sy}=\frac{y-y^{\min }}{y^{\max }-y^{\min }}(d-c)+c---\left(5\right)$

wherein y is the actual analysis value of the constant top oil dry point of the output variable, y^min，y^maxThe lower limit and the upper limit of the actual analysis value of the output variable are shown, the value after normalization is sy, and c and d represent the lower limit and the upper limit of the output value of the neural network model after normalization processing. Here, the values after normalization (sx) are taken such that d is 0.9 and c is 0.1₁，sx₂Lsx₆Sy) as a neural network training sample set.

3. Establishment of artificial neural network model

The artificial neural network employs a three-layer feedforward network, as shown in FIG. 2. The activation functions all use tansig functions, an error back propagation algorithm (BP algorithm) is adopted to train the network, the number of nodes of an input layer is 6, the number of nodes of an implicit layer is determined to be 5-10, the number of nodes of an output layer is 1, the network is trained by utilizing a training sample set, and the training error function MSE is as follows:

$\mathrm{MSE}=\frac{1}{n}{(y^{\′}-\mathrm{sy})}^2---\left(6\right)$

wherein y' is a calculated value of the neural network, and sy is an artificial analysis value of the dry point after normalization.

After a certain training algebra, when the error is less than a certain requirement, stopping training, and outputting the weight and threshold of the neural network, namely obtaining the sx described by the neural network₁，sx₂Lsx₆The approximate relationship to sy can be written as:

sy≈y′＝f(sx₁，sx₂Lsx₆)(7)

wherein the neural network model f (-) describes sx₁，sx₂Lsx₆The function relation with y ' is that y ' is used to approximate the true value sy, then y ' is inverse normalized to obtain the calculated value of the modelNamely, it is

$\tilde{y}=\frac{y^{\′}-c}{d-c}(y^{\max }-y^{\min })+y^{\min }---\left(8\right)$

4. On-line correction of soft measurement models

The method comprises the steps of establishing points in a distributed control system, writing the neural network model into a DCS, acquiring values of 6 input variables in real time or calculating, then giving a model calculation value of a common top oil dry point through model calculation, comparing the model calculation value with a manual analysis value at a corresponding moment, correcting the model calculation value by using errors of the model calculation value and the model calculation value, and finally outputting a correction value.

The model was calibrated as follows:

since the dry spots were analyzed manually every 8 hours, the model was corrected every 8 hours. The calculation formula is as follows:

$\hat{y}\left(t\right)=\tilde{y}\left(t\right)+\mathrm{\α}(y_a-{\tilde{y}}_a)---\left(9\right)$

wherein, y_aFor the value of the normal top oil dry point manual analysis,is a neural netThe corrected output value of the network model,for the output value of the model that is not corrected,α is the weighted value of the error between the artificial analysis value and the neural network model output value, and can be taken according to the confidence of the results given by the artificial analysis value and the neural network model output value.

Drawings

FIG. 1 is a graph of sensitivity analysis of furnace exit temperature versus common top oil dry point

FIG. 2 is a block diagram of a neural network soft measurement model for a common-roof oil dry point

Detailed Description

1. Data acquisition and preprocessing

Collecting more than 400 groups of historical production data from an industrial field, removing gross errors and abnormal data, selecting 400 groups for training a neural network, and carrying out normalization processing on input variables by using a formula (4), wherein x is₁Has a variation range of [100.5, 114.86 ]]，x₂Has a variation range of [32.31, 55.35 ]]，x₃Has a variation range of [69.21, 86.67 ]]，x₄Has a variation range of [2747.49, 4427.83 ]]，x₅Has a variation range of [0.0145, 0.0279 ]]，x₆Has a variation range of [97.9, 123.1 ]]Here, a is-1 and b is 1. The input variable normalization calculation formula is as follows:

${\mathrm{sx}}_1=\frac{x_1-114.86}{114.86-100.5}\×2-1$

${\mathrm{sx}}_2=\frac{x_2-32.31}{55.35-32.31}\×2-1$

${\mathrm{sx}}_3=\frac{x_3-69.21}{86.67-69.21}\×2-1$

${\mathrm{sx}}_4=\frac{x_4-2747.49}{4427.83-2747.49}\×2-1$

${\mathrm{sx}}_5=\frac{x_5-0.0145}{0.0279-0.0145}\×2-1$

${\mathrm{sx}}_6=\frac{x_6-97.9}{123.1-97.9}\×2-1$

the dependent variable is normalized by using the formula (5), the range of the normalized common top oil is [135.4, 154.6], the normalized common top oil is normalized to [0.1, 0.9], and d is 0.9, c is 0.1, and the calculation formula is as follows:

$\mathrm{sy}=\frac{y-135.4}{154.6-135.4}(0.9-0.1)+0.1$

2. establishment of soft measurement model

The number of nodes of the input layer of the neural network is 6, the number of nodes of the middle layer is 8, and the number of nodes of the output layer is 1. And taking 320 groups of normalized data as training samples, training the 3-layer feedforward neural network by using a BP algorithm, taking the remaining 80 groups of samples as test data, checking the model, and outputting parameters of the neural network after the average value of model training errors and the average prediction error are not more than 2 ℃ to obtain the common-ejection oil dry point soft measurement model. The parameters related to the constant oil dry point neural network are as follows:

$w1=\left[\begin{array}{cccccc}-0.341& 0.6661& -1.6805& 0.2676& -2.7082& 1.9736\\ 9.8793& -202.701& -2.6907& -113.45& 321.4433& -395.88\\ -5.4458& 0.3865& 3.559& -4.5623& -7.0847& -2.4882\\ 4.9235& 0.5815& -3.9916& 4.1895& -0.1155& 3.3565\\ -1.5312& -0.9696& 1.5297& 3.5879& 0.6568& 0.7242\\ -21.5012& 2.1147& -44.9359& -14.2282& -0.5703& 70.2191\\ -5.0854& 1.1167& 1.2024& 2.668& -0.7608& -3.1868\\ 0.3375& -0.6293& 1.6411& -0.2988& 2.6869& -1.9619\end{array}\right]$

w2＝[18.88560.0618-0.1445-0.13670.2957-0.1004-0.169819.0212]

b1＝[1.5976-127.5591.4681-0.50360.019117.9897-3.2703-1.6006]

b2＝0.6431

wherein w1 and w2 are connection weights of the neural network, and b1 and b2 are thresholds of the hidden layer node and the output node, respectively. Therefore, the neural network model of the common top oil dry point is:

o^h＝t(sx·w1+b1)(10)

y′＝t(oh·w2+b2)(11)

wherein,oh is the hidden layer output value and y' is the output value of the neural network.

$o_1^h=t(\underset{i=1}{\overset{6}{\mathrm{\Σ}}}w1(1,i){\mathrm{gsx}}_i+b1\left(\mathrm{1,1}\right))$

$o_2^h=t(\underset{i=1}{\overset{6}{\mathrm{\Σ}}}w1(2,i){\mathrm{gsx}}_i+b1\left(\mathrm{1,2}\right))$

$o_3^h=t(\underset{i=1}{\overset{6}{\mathrm{\Σ}}}w1(3,i){\mathrm{gsx}}_i+b1\left(\mathrm{1,3}\right))$

$o_4^h=t(\underset{i=1}{\overset{6}{\mathrm{\Σ}}}w1(4,i){\mathrm{gsx}}_i+b1\left(\mathrm{1,4}\right))$

$o_5^h=t(\underset{i=1}{\overset{6}{\mathrm{\Σ}}}w1(5,i){\mathrm{gsx}}_i+b1\left(\mathrm{1,5}\right))$

$o_6^h=t(\underset{i=1}{\overset{6}{\mathrm{\Σ}}}w1(6,i){\mathrm{gsx}}_i+b1\left(\mathrm{1,6}\right))$

$o_7^h=t(\underset{i=1}{\overset{6}{\mathrm{\Σ}}}w1(7,i){\mathrm{gsx}}_i+b1\left(\mathrm{1,7}\right))$

$o_8^h=t(\underset{i=1}{\overset{6}{\mathrm{\Σ}}}w1(8,i){\mathrm{gsx}}_i+b1\left(\mathrm{1,8}\right))$

$y^{\′}=t(o_1^h\·w2\left(\mathrm{1,1}\right)+o_2^h\·w2\left(\mathrm{1,2}\right)+o_3^h\·w2\left(\mathrm{1,3}\right)+o_4^h\·w2\left(\mathrm{1,4}\right)+o_5^h\·w2\left(\mathrm{1,5}\right)+o_6^h\·w2\left(\mathrm{1,6}\right)+o_7^h\·w2\left(\mathrm{1,7}\right)+o_8^h\·w2\left(\mathrm{1,8}\right)+b2)$

Wherein,representing the output value of the ith node of the hidden layer. w1(i, j) represents the j-th dimension variable of the ith row in the weight matrix w1, b1(i, j) represents the j-th dimension variable of the ith row in the threshold matrix b1, and w2(i, j) represents the j-th dimension variable of the ith row in the weight matrix w 2.

y' is a calculated value of the neural network, and a predicted value of the neural network can be obtained through inverse normalization, wherein the range of the constant top oil normalization is [135.4, 154.6], the normalization is performed to [0.1, 0.9], d is 0.9, and c is 0.1, so that the inverse normalization formula is as follows:

$\tilde{y}=\frac{y^{\′}-0.1}{0.9-0.1}(154.6-135.4)+135.4$

the above describes a distributed control system-based ordinary-top oil dry-point neural network model, and next, a correction method of the model is described. Taking the calculated value of the t model at a certain time as an example:

the acquisition or calculation values of 6 input variables at time t are:

$\left[\begin{array}{c}{x}_1\\ x_2\\ x_3\\ x_4\\ x_5\\ x_6\end{array}\right]=\left[\begin{array}{c}107.85\\ 49.02\\ 81.77\\ 3723.35\\ 0.016\\ 114.27\end{array}\right]$

after normalization, [ sx ] is obtained₁，sx₂Lsx₆]＝[0.024，0.450，0.438，0.162，-0.772，0.307]Substituting the neural network model to obtain the model, wherein the calculated value of the model is as follows:since the previous manual analysis value corresponding to the time t is 142 and the current soft measurement model output value is 141.18, the following formula is used for correction:

the corrected output value of the model at time t here is taken to be α at 0.2.

Claims (1)

1. An online soft measurement method for a common oil dry point is characterized by comprising the following steps: the soft measurement method comprises the following steps:

(1) simulating the operating condition of the atmospheric tower by utilizing Aspenplus, carrying out sensitivity analysis on a key operating variable of the atmospheric tower about a constant-top oil dry point, and selecting the operating variable which has a significant influence on the constant-top oil dry point as a candidate set of auxiliary variables of a constant-top oil dry point soft measurement model according to the result of the sensitivity analysis;

(2) correlation between variables for auxiliary variables selected by sensitivity analysisAnalyzing, removing redundant variables, and finally selecting the constant top temperature (x)₁DEG C), normal top pressure (x)₂Kpa), constant top reflux temperature (x)₃DEG C), the energy (x) of the normal top circulation₄DEG C. t/h), production ratio x₅And side line total amount (x) of atmospheric tower₆T/h) as an input variable of the neural network model;

(3) selecting a common-top oil dry point as an output value of the network model;

(4) the auxiliary variable x is obtained by directly measuring or indirectly calculating by using a measuring instrument of a corresponding data acquisition point or a Distributed Control System (DCS)₁～x₆A value of (d);

(5) the method comprises the steps of preprocessing acquired field data to form a neural network training sample set, training a three-layer feedforward neural network by using an error Back Propagation (BP) algorithm, outputting parameters of a neural network model, namely a connection weight and a threshold value after an error meets a certain requirement, and obtaining the neural network model about a common-top oil dry point;

(6) building points in the DCS according to input and output variables of the neural network model, writing the model into the DCS, calculating and outputting the soft measurement value (y, DEG C) of the dry point in real time through model calculation according to real-time data of the input variables, and further guiding an operator to adjust the production process in time;

the auxiliary variable x₁,x₂,x₃Measured directly by the meter, said variable x₄,x₅,x₆Indirectly calculated by the following formula:

x₄＝F₁×(t₁-t₂)

x₅＝F₂/F

x₆＝F₃+F₄+F₅+F₆

wherein, F₁Is the constant top circulation flow (t/h), t₁、t₂The temperature of the outlet and inlet of the constant top circulation is in DEG C, F₂The normal top gasoline extraction amount (t/h), F the normal pressure tower feeding amount (t/h), F₃,F₄,F₅,F₆Respectively a normal line, a,Normal four line production (t/h);

the input and output variables of the neural network model are normalized;

the input variables of the neural network model are normalized according to the following formula:

${\mathrm{sx}}_i=\frac{x_i-x_i^{\min }}{x_i^{\max }-x_i^{\min }}(b-a)+a,i=1,2,...,6$

in the formula, x_iIs a measured value of the ith input variable, sx, of the neural network model_iThe ith input variable is used as the input value of the neural network after being normalized,andrepresenting the lower limit and the upper limit of the variation range of the ith acquired input variable, and a and b representing the lower limit and the upper limit of the input variable after normalization processing; taking b as 1 and a as-1;

the dependent variable of the neural network model is normalized by the following formula:

$sy=\frac{y-y^{\min }}{y^{\max }-y^{\min }}(d-c)+c$

wherein y is the actual analysis value of the constant top oil dry point of the output variable, y^min，y^maxThe lower limit and the upper limit of an actual analysis value of an output variable constant top oil dry point are provided, the value after normalization is sy, and c and d represent the lower limit and the upper limit of the output value of the neural network model after normalization processing;

input and output data are normalized to obtain a training sample set (sx)₁，sx₂…sx₆Sy), then training a three-layer feedforward neural network by adopting a BP algorithm, outputting model parameters after a training error meets a certain requirement, and obtaining a mapping relation of model input and model output:

y′＝f(sx₁，sx₂…sx₆)

the obtained model output needs to be subjected to inverse normalization processing to obtain the final constant top oil dry point value, and the inverse normalization is calculated according to the following formula:

$\tilde{y}=\frac{y^{\′}-c}{d-c}(y^{\max }-y^{\min })+y^{min}$

in the formula,is the output value y after inverse normalization of the normal top oil dry point soft measurement model^min，y^maxThe lower limit and the upper limit of an actual analysis value of a constant-top oil dry point of an output variable of the neural network are set, y' is a calculated value of the neural network, and c and d represent the lower limit and the upper limit of output value normalization processing of the neural network model;

model calculation value before next artificial analysis value is generated by using dry point artificial analysis value at time t every dayCorrection is carried out, and the correction formula is as follows:

$\hat{y}\left(t\right)=\tilde{y}\left(t\right)+\mathrm{\α}(y_a-{\tilde{y}}_a)$

wherein, y_aFor the value of the normal top oil dry point manual analysis,is the corrected output value of the neural network model,for the output value of the model that is not corrected,and α is a weighted value of the error between the artificial analysis value and the output value of the neural network model, and the value is taken according to the trust degree of the results given by the artificial analysis value and the output value of the neural network model.