CN101930008A - Soft measurement method of 'ammonia net value' in ammonia synthesis process - Google Patents
Soft measurement method of 'ammonia net value' in ammonia synthesis process Download PDFInfo
- Publication number
- CN101930008A CN101930008A CN2010102499887A CN201010249988A CN101930008A CN 101930008 A CN101930008 A CN 101930008A CN 2010102499887 A CN2010102499887 A CN 2010102499887A CN 201010249988 A CN201010249988 A CN 201010249988A CN 101930008 A CN101930008 A CN 101930008A
- Authority
- CN
- China
- Prior art keywords
- ammonia
- soft
- neural network
- value
- convertor
- 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.)
- Granted
Links
Images
Landscapes
- Investigating Or Analyzing Non-Biological Materials By The Use Of Chemical Means (AREA)
- Organic Low-Molecular-Weight Compounds And Preparation Thereof (AREA)
- Treating Waste Gases (AREA)
Abstract
The invention relates to a soft measurement method of an 'ammonia net value' in an ammonia synthesis process by using a 'one-axle and three-radial beds quenching' ammonia synthesis tower. In the method, soft measured values of ammonia content in nitrogen and hydrogen mixed gas entering the tower and ammonia content in the 'gas' which is discharged out of the ammonia synthesis tower for the second time are obtained respectively through different BP (back propagation) neural network soft measurement models by selecting certain technological parameters in the ammonia synthesis process using the 'one-axle and three-radial beds quenching ammonia synthesis tower so as to obtain the on-line 'ammonia net value'. The soft measurement method provides favorable basis for product quality control of synthetic ammonia manufacturers: the neural network models are employed for pre-measuring, thus greatly improving the pre-measuring accuracy and fault tolerance performance of the ammonia content at the inlet and the outlet of the ammonia synthesis tower, and overcoming influence on the production process and the product quality caused by lagged measurement of the ammonia content at the inlet and the outlet of the ammonia synthesis tower.
Description
Technical field
The present invention relates to the flexible measurement method of " ammonia net value " in a kind of ammonia synthesis process, specifically, relate to the flexible measurement method of " ammonia net value " in the ammonia synthesis process of a kind of employing " one three footpath cold shock formula " ammonia convertor.
Background technology
Adopt the ammonia synthesis process of " one three footpath cold shock formula " ammonia convertor to have advantages such as low energy consumption and high yield with it, become the main flow technology of middle-size and small-size Ammonia Production, its schematic flow sheet as shown in Figure 1.
As shown in Figure 1: gaseous ammonia content is not more than 4% the tower nitrogen and hydrogen mixture of going into and divides one the tunnel as " one goes into " main line, enters ammonia synthesis Tata wall annular space, becomes " one goes out " gas after cooling off the ammonia convertor urceolus from top to bottom.Another road of going into the tower nitrogen and hydrogen mixture with enter gas-gas heat exchanger after the gas that " one goes out " comes out mixes.The gas that comes out through the gas-gas heat exchanger top, the one tunnel as " two go into " main line, enters between the Tube Sheet of Heat Exchanger of ammonia convertor bottom, and by the heating of the hot gas behind the inner reaction tube, the first axial layer that enters tower top through pipe core reacts; The two-way in addition of the gas that comes out through the gas-gas heat exchanger top enters synthetic tower from first radial layer and second radial layer respectively as quench gas (" quench gas ", " two quench gas ").Axially the reacted hot gas of layer and " quench gas " are mixed into first radial layer, after reacted gas goes out first radial layer, enter second radial layer, mix with " two quench gas " that be added into, enter second radial layer and react, reacted gas enters the 3rd radial layer and reacts.Through the reacted gaseous ammonia content of the 3rd radial layer is 12%~14%, enter the heat interchanger heat exchange of synthetic tower bottom after, go out for the second time synthetic tower (being called " two go out " gas).Because the content of ammonia directly influences " ammonia net value " (difference of ammonia content and " going into the tower nitrogen and hydrogen mixture " middle ammonia content in " two go out " gas) in " going into the tower nitrogen and hydrogen mixture " and " two go out " gas, be one of important indicator of being paid close attention to of each Ammonia Production merchant.
" ammonia net value " is the important reference of evaluation of catalyst activity.In time obtained " ammonia net value ", situation that can online judgement ammonia synthesis reaction, the attenuation of reflection catalyst activity, thereby in time adjust the scope of activities of reaction bed temperature, (if under the normal running parameter, " ammonia net value " is always than higher in the catalytic action of more effective performance catalyzer, temperature of reaction is higher simultaneously, this moment is rapid in order to prevent the catalyst deactivation that high temperature causes, and should suitably turn down the temperature of reaction of each beds, prolongs the serviceable life of catalyzer.If in one period, under the normal running parameter, ammonia net value continues to descend, illustrate that catalyst activity reduces, decline for the reaction velocity that causes of compensation loss of activity, should promote the temperature of reaction of each beds,, so just can obtain the high yield in the catalyzer life cycle) to improve " ammonia net value ".
At present, the mensuration of ammonia content in ammonia convertor in the ammonia synthesis process that adopts " one three footpath cold shock formula " ammonia convertor " going into the tower nitrogen and hydrogen mixture " and " two go out " gas is adopted hand sampling analysis (be to locate and " two go out " exit hand sampling of ammonia convertor " going into the tower nitrogen and hydrogen mixture " house steward in per 8 hours, and carry out the off-line analysis of ammonia content) more.Obviously, hysteresis quality is the greatest drawback of hand sampling analytic approach.The production of adopting " ammonia net value " of hysteresis quality to come " guidance " synthetic ammonia simultaneously can influence the catalyzer service efficiency and cause the fluctuation of product quality.
Given this, how in time obtain to adopt " ammonia net value " in the ammonia synthesis process of " one three footpath cold shock formula " ammonia convertor just to become the technical issues that need to address of the present invention.
Summary of the invention
The objective of the invention is to, the online soft sensor method of " ammonia net value " in the ammonia synthesis process of a kind of employing " one three footpath cold shock formula " ammonia convertor is provided, overcomes the defective (mainly being) that exists in the prior art because the low and unstable product quality of catalyzer service efficiency that hysteresis quality caused of " ammonia net value ".
The online soft sensor method of " ammonia net value " in the ammonia synthesis process of the said employing of the present invention " one three footpath cold shock formula " ammonia convertor comprises the steps:
(1) choose in the ammonia synthesis process of employing " one three footpath cold shock formula " ammonia convertor, (brief note is H to go into hydrogen richness in the tower nitrogen and hydrogen mixture
2), go into the pressure (brief note for P) of tower nitrogen and hydrogen mixture, (brief note is CH to go into the temperature (brief note is T) of tower nitrogen and hydrogen mixture and the middle methane content of dropping a hint
4) the on-line measurement value, with its normalized, go into the input variable of BP (Back Propagation) neural network of ammonia content in the tower nitrogen and hydrogen mixture as ammonia convertor, the output variable of BP neural network promptly gets ammonia content in the tower nitrogen and hydrogen mixture after anti-normalized (brief note is NH
3in) soft measured value;
(2) choose in the ammonia synthesis process of employing " one three footpath cold shock formula " ammonia convertor, (brief note is H to go into hydrogen richness in the tower nitrogen and hydrogen mixture
2, with (1) with), go into the pressure (brief note is for P, with (1) with) of tower nitrogen and hydrogen mixture, go into the flow (brief note is F) of tower nitrogen and hydrogen mixture, (brief note is F to the flow of a quench gas
1), (brief note is F to the flow of two quench gas
2) and the middle methane content of dropping a hint (brief note is for CH
4, with (1) with) the on-line measurement value and go into ammonia content in the tower nitrogen and hydrogen mixture (brief note be for NH
3in) soft measured value (by in the step (1)), with its normalized, (brief note is NH as ammonia content in ammonia convertor " two go out " gas
3out) the input variable of BP neural network, the output variable of BP neural network after anti-normalized in " two go out " gas of ammonia convertor ammonia content (brief note is NH
3out) soft measured value;
(3) calculating is by ammonia content (NH in " two go out " gas of the ammonia convertor of step (2) gained
3out) soft measured value with go into ammonia content (NH in the tower nitrogen and hydrogen mixture by step (1) gained
3in) the difference of soft measured value, online " ammonia net value " (" the ammonia net value "=NH in the ammonia synthesis process of " one three footpath cold shock formula " ammonia convertor
3out-NH
3in).
Wherein: the schematic flow sheet of the ammonia synthesis process of said employing " one three footpath cold shock formula " ammonia convertor as shown in Figure 1, concrete technology can be consulted: (Shen Jun etc. " synthetic ammonia " Chemical Industry Press, 2001, p815).
Description of drawings
Fig. 1 is the schematic flow sheet of the ammonia synthesis process of employing of the present invention " one three footpath cold shock formula " ammonia convertor.
Fig. 2 is inlet ammonia content BP of the present invention (Back Propagation) neural network synoptic diagram.
Fig. 3 is outlet ammonia content BP of the present invention (Back Propagation) neural network synoptic diagram.
Fig. 4 is the synoptic diagram of inlet ammonia content of the present invention, outlet ammonia content soft-sensing model on-line prediction and model tuning.
Embodiment
The online soft sensor method of " ammonia net value " in the ammonia synthesis process of the said employing of the present invention " one three footpath cold shock formula " ammonia convertor comprises the steps:
(1) NH
3inThe acquisition of soft measured value:
(1-1) selection of auxiliary variable:
According to the analysis of the ammonia synthesis process that adopts " one three footpath cold shock formula " ammonia convertor to having now, the present inventor has selected H
2, P, T and CH
4The on-line measurement value be NH
3inThe input variable of soft-sensing model (BP neural network);
(1-2) preconditioning technique of process data:
The preconditioning technique of process data comprises check and the correction that the data of gathering are carried out.To take into full account validity, sequential, the integrality of data in the image data process, to eliminate the influence of stochastic error and human error to measured value.The preconditioning technique of the process data that the present invention relates to is as follows:
1. reject the data of apparent error according to production operation experience and 3 σ criterions:
Analyze by the historical data to actual production, each auxiliary variable in the soft-sensing model (is H
2, P, T and CH
4) all approximate Normal Distribution, promptly satisfy formula (1) (rejecting obvious irrational measured value):
|ξ
in-μ
in|<3σ
in (1)
In the formula (1): ξ
InFor certain input variable of BP neural network (is H
2, P, T and CH
4) measured value, μ
InBe mathematical expectation, σ
In 2Be variance.
The present invention has the data of apparent error to remove soft-sensing model in the newly-built or used historical data of timing by the method.
2. disturb for reducing, carry out The disposal of gentle filter for real-time process The data digital filtering method:
The function of digital filtering is the Serial No. that the Serial No. of one group of input is changed into another group output after by certain calculation.The present invention adopts the single order digital filtering technique, for example for n group (corresponding n sampling instant) list entries X
In(n), it through filtered output sequence is:
Wherein
a
In>0, b
In>0, a
In+ b
In=1, a
In=0.9, b
In=0.1.
(1-3) NH
3inThe foundation of soft-sensing model:
The present invention adopts the feed-forward type neural network of error backpropagation algorithm, is called for short BP (Back Propagation) network (seeing also Fig. 2).It contains input layer, hidden layer and output layer, is 3 layers of feedforward network structure.This BP network has 4 input node (H
2, CH
4, P, T).Output layer has 1 node, and its corresponding output variable is: (seeing also Fig. 1) goes into (the brief note NH of ammonia content in the tower nitrogen and hydrogen mixture
3in) [0,1] scope in soft measured value.
The hidden layer of described BP network has 4 nodes.
1. (be H to input variable
2, P, T or CH
4The value of on-line measurement value after the single order digital filtering is handled) carry out normalized by formula (2), as the input variable of BP neural network model.
nX
in=(X
in-minX
in)/(maxX
in-minX
in) (2)
X in the formula (2)
InFor certain input variable (is H
2, P, T or CH
4) the value of on-line measurement value after the single order digital filtering is handled, nX
InBe the later numerical value of normalization, [minX
In, maxX
In] be that certain input variable (is H
2, P, T or CH
4The on-line measurement value) variation range.
2. the feedforward of BP network is calculated
Learning phase at this network of training is provided with N training sample, then use p (p=1,2 ..., N) organize the input data inX of sample
pWith output data ind
p(adopt historical NH during newly-built BP Network Soft Sensor Model
3inThe manual analysis value; During soft survey model tuning set by step in (1-5) mode of soft-sensing model online " rolling " carry out value) network is trained, the feedforward of BP network is calculated as follows:
Being input as of j node of hidden layer
Wherein j is the index of hidden layer node;
It is the input of the hidden layer of p group training data;
Be output (output of input layer equals its input) from i node of input layer;
Be the weights that are connected between i node of input layer and j node of hidden layer; θ
1jThreshold value for j node of hidden layer.
J node of hidden layer is output as
Being input as of output layer node
Wherein
Input for output layer; η
jBe the weights that are connected between j node of hidden layer and the output layer node; θ
2Threshold value for the output layer node.
The output of output layer node, promptly the BP network is output as
3. BP network weight and threshold value determines
If the right quadratic form error function of input and output mode of each group sample p is defined as:
The average error cost function of N group data is:
The array format of neural network weight and threshold value is
, its total quantity is dim=4 * 4+4+4+1=25.
Adopting the BP Neural Network Toolbox of Matlab7.0 software to be optimized to the BP neural network soft sensor model of going into ammonia content in the tower nitrogen and hydrogen mixture finds the solution, obtains the weights and the threshold value of BP neural network.
(1-4) use soft-sensing model and carry out on-line prediction:
With input variable (is H
2, P, T or CH
4The on-line measurement value) through the single order digital filtering and by formula after (2) normalized in the BP neural network soft sensor model of substitution (1-3), to obtain soft-sensing model [0,1] output valve of scope, (8) anti-normalization by formula obtains ammonia content (NH in the tower nitrogen and hydrogen mixture
3in) soft measured value.
Y
in=nY
in×(maxY
in-minY
in)+minY
in (8)
Y in the formula (8)
InFor going into ammonia content (NH in the tower nitrogen and hydrogen mixture
3in) soft measured value, nY
InFor going into ammonia content (NH in the tower nitrogen and hydrogen mixture
3in) the BP neural network at the output variable value of [0,1], [maxY
In, maxY
In] for going into ammonia content (NH in the tower nitrogen and hydrogen mixture
3in) variation range.
(1-5) on-line correction of soft-sensing model:
Because the factors such as imperfection of the time variation of soft measuring object, non-linear and model, must consider the on-line correction of model, could adapt to new operating mode.For this reason, according to per 8 hours once go into tower nitrogen and hydrogen mixture ammonia content (NH
3in) the manual analysis value soft-sensing model carried out online " rolling " (that is, judge NH
3inThe soft-sensing model output valve whether exceed restriction with the relative error of corresponding manual analysis value; If transfinite, then with NH
3inThe soft-sensing model output valve be changed to NH
3inThe manual analysis value, and by input data of the history in previous 8 hours periods and corresponding NH
3inSoft measurement data is to NH
3inSoft-sensing model is trained again, obtains new model parameter) optimize and proofread and correct, make soft-sensing model adapt to the variation of industrial process operating characteristic and the migration of production status.
On-line measurement value (the H of the existing instrument that the present invention reads in the DCS of Ammonia Production factory (Distributed Control Systems) in real time
2, CH
4, P, T) and go on the basis of tower nitrogen and hydrogen mixture ammonia content manual analysis value, can infer in real time into tower nitrogen and hydrogen mixture ammonia content (NH
3in).
(2) NH
3outThe acquisition of soft measured value:
(2-1) selection of auxiliary variable:
According to the analysis of the ammonia synthesis process that adopts " one three footpath cold shock formula " ammonia convertor to having now, the present inventor has selected H
2, P, F, F
1, F
2And CH
4On-line measurement value and NH
3inSoft measured value be NH
3outThe input variable of soft-sensing model (BP neural network);
Wherein: read in real time among the DCS (Distributed Control Systems) of said on-line measurement value by Ammonia Production factory, and with step (1) in go into ammonia content (NH in the tower nitrogen and hydrogen mixture
3in) the relevant input variable value correspondence of soft-sensing model.
(2-2) preconditioning technique of process data:
The preconditioning technique of process data comprises check and the correction that the data of gathering are carried out.To take into full account validity, sequential, the integrality of data in the image data process, to eliminate the influence of stochastic error and human error to measured value.The preconditioning technique of the process data that the present invention relates to is as follows:
1. reject the data of apparent error according to production operation experience and 3 σ criterions:
Analyze by the historical data to actual production, each auxiliary variable in the soft-sensing model (is H
2, P, F, F
1, F
2, CH
4Or NH
3in) all approximate Normal Distribution, promptly satisfy formula (9) (rejecting obvious irrational measured value)
|ξ
out-μ
out|<3σ
out (9)
In the formula (9): ξ
OutCertain input variable (H for the BP neural network
2, P, F, F
1, F
2, CH
4Measured value or NH
3inSoft measured value), μ
OutBe mathematical expectation, σ
Out 2Be variance.
The present invention has the data of apparent error to remove soft-sensing model in the newly-built or used historical data of timing by the method.
2. disturb for reducing, carry out The disposal of gentle filter for real-time process The data digital filtering method:
The function of digital filtering is the Serial No. that the Serial No. of one group of input is changed into another group output after by certain calculation.The present invention adopts the single order digital filtering technique, for example for n group list entries (corresponding n sampling instant) X
Out(n), it through filtered output sequence is:
Wherein
a
Out>0, b
Out>0, a
Out+ b
Out=1, a
Out=0.9, b
Out=0.1.
(2-3) NH
3outThe foundation of soft-sensing model:
Ammonia content (NH in ammonia convertor among the present invention " two go out " gas
3out) soft-sensing model adopt the feed-forward type neural network of error back propagation, be called for short BP (Back Propagation) network.As shown in Figure 3: it contains input layer, hidden layer and output layer, is 3 layers of feedforward network structure.This BP network has 7 input node (H
2, CH
4, NH
3in, P, F, F
1And F
2).Output layer has 1 node, and its corresponding output variable is: (the brief note NH of ammonia content in " two go out " gas of (referring to Fig. 1) ammonia convertor
3out) the interior soft measured value of [0,1] scope.
The hidden layer of described BP network has 7 nodes.ω
IjBe the weights that are connected between i node of input layer and j node of hidden layer; b
1jThreshold value for j node of hidden layer; υ
jBe the connection weights between j node of hidden layer and the output layer node; b
2Threshold value for the output layer node.The input of hidden layer and output layer node is the weighted sum of the output of last node layer, and the incentive degree of each node is decided by the Sigmoid excitation function.
1. (be H to input variable
2, CH
4, P, F, F
1, F
2On-line measurement value and NH
3inThe value of soft measured value after the single order digital filtering is handled) carry out normalized by formula (10), as the input variable of BP neural network model.
nX
out=(X
out-minX
out)/(maxX
out-minX
out) (10)
X in the formula (10)
OutFor certain input variable (is H
2, CH
4, P, F, F
1, F
2On-line measurement value and NH
3inThe value of soft measured value after the single order digital filtering is handled), nX
OutBe the later numerical value of normalization, [minX
Out, maxX
Out] be that certain input variable (is H
2, P, F, F
1, F
2, CH
4On-line measurement value or NH
3inSoft measured value) variation range.
2. the feedforward of the soft measurement BP network model of ammonia content is calculated in ammonia convertor " two go out " gas
Learning phase at this network of training is provided with N training sample, then use p (p=1,2 ..., N) organize the input data X of sample
pWith output data d
p(adopt historical NH during newly-built BP Network Soft Sensor Model
3outThe manual analysis value; During soft survey model tuning set by step in (2-5) mode of soft-sensing model online " rolling " carry out value) network is trained, the feedforward of BP network is calculated as follows:
Being input as of j node of hidden layer
Wherein j is the index of hidden layer node;
It is the input of the hidden layer of p group training data;
Be output (output of input layer equals its input) from i node of input layer; ω
IjBe the weights that are connected between i node of input layer and j node of hidden layer; b
1jThreshold value for j node of hidden layer.
J node of hidden layer is output as
Being input as of output layer node
Wherein
Input for output layer; υ
jBe the weights that are connected between j node of hidden layer and the output layer node; b
2Threshold value for the output layer node.
The output of output layer node, promptly the BP network is output as
Wherein
Output for output layer.
3. the right quadratic form error function of input and output mode of establishing each group sample p of determining of the soft measurement BP network model weights of ammonia content and threshold value is defined as in ammonia convertor " two go out " gas:
The average error cost function of N group data is:
The present invention uses particle swarm optimization algorithm (PSO) optimization to obtain the weights and the threshold value of BP neural network.In the PSO algorithm, each particle position correspondence of population be a candidate vector of neural network weight and threshold value, array format is (ω
11, ω
12..., ω
17, ω
21, ω
22..., ω
27, ω
31, ω
32..., ω
37, ω
41, ω
42..., ω
47, ω
51, ω
52..., ω
57, ω
61, ω
62..., ω
67, ω
71, ω
72..., ω
77, υ
1, υ
2..., υ
7, b
11, b
12..., b
17, b
2), its total quantity is dim=7 * 7+7+7+1=64.
The step of wherein said PSO algorithm optimization neural network weight and threshold value is as follows:
A) initialization of PSO algorithm parameter: population scale Popsize=50; Maximum iteration time MaxIter=2000; The maximal value of inertial factor is 0.9, and minimum value is 0.2.Two study operators are 2.0.
B) particle position in the population and speed are carried out initialization, the various piece of particle position is the initial weight and the threshold value of network, and its scope is [2.0,2.0], and the various piece of particle rapidity then is to get random number in [2.0,2.0] scope.
C) adaptive value function f itness is the average error cost function of N group data, promptly
D) utilize the adaptive value function to estimate the adaptive value of each particle.
E) adaptive value of each particle and the desired positions pbest of particle self are compared.When the current adaptive value of particle is better than self best adaptive value, then upgrade pbest.
F) pbest of each particle and the optimal location gbest of whole population are compared.When the best adaptive value of particle self is better than the optimal-adaptive value of population, then upgrade gbest.
G) calculate current inertial factor, and particle's velocity and position are upgraded.
Wherein w is an inertial factor; Iter is the current iteration number of times; I represents i individual index in the population; V is a particle's velocity; c
1, c
2Be respectively two study operators; r
1, r
2Be respectively two equally distributed random numbers; p
iI individual desired positions pbest in the expression population; p
gThe optimal location gbest that represents whole population; x
iI individual position in the expression population.
H) judge whether current iteration number of times iter reaches MaxIter.Then do not go to step d); Otherwise the pairing particle position of output optimal-adaptive value promptly is neural network weight and the threshold value that is obtained by the PSO algorithm optimization.
(2-4) use soft-sensing model and carry out on-line prediction:
With input variable (is H
2, CH
4, P, F, F
1, F
2On-line measurement value and NH
3inSoft measured value) through the single order digital filtering and by formula after (10) normalized in the BP neural network soft sensor model of substitution (2-3), the soft-sensing model output valve that obtains is obtained ammonia content (NH in ammonia convertor " two go out " gas by the anti-normalization of formula (17)
3out) soft measured value.
Y
out=nY
out×(maxY
out-minY
out)+minY
out (17)
Y in the formula (17)
OutBe ammonia content (NH in ammonia convertor " two go out " gas
3out) soft measured value, nY
OutBe the output variable value in [0,1] of BP neural network, [minY
Out, maxY
Out] be ammonia content (NH in ammonia convertor " two go out " gas
3out) variation range.
(2-5) NH
3outThe on-line correction of soft-sensing model:
Because the factors such as imperfection of the time variation of soft measuring object, non-linear and model, must consider the on-line correction of model, could adapt to new operating mode.For this reason, go into tower nitrogen and hydrogen mixture ammonia content (NH according to per 8 hours ammonia convertors once
3in) the manual analysis value and ammonia convertor " two go out " gas in ammonia content (NH
3out) the manual analysis value soft-sensing model carried out online " rolling " (that is, judge NH
3inAnd NH
3outWhether the soft-sensing model output valve exceeds restriction with the relative error of corresponding manual analysis value; If wherein have any one to transfinite, then with NH
3in, NH
3outThe soft-sensing model output valve be changed to corresponding manual analysis value, and import data and corresponding NH by the history in previous 8 hours periods
3in, NH
3outSoft measurement data, to NH
3in, NH
3outSoft-sensing model train again, obtain new model parameter) optimize to proofread and correct, make soft-sensing model adapt to the variation of industrial process operating characteristic and the migration of production status.
(3) acquisition of online " ammonia net value ":
The present invention is at the on-line measurement value (H that utilizes existing instrument
2, CH
4, P, F, F
1And F
2) and go into the soft measured value (NH of ammonia content in the tower nitrogen and hydrogen mixture
3in) the basis on, can infer the ammonia content (NH in ammonia convertor " two go out " gas in real time
3out), thereby calculate ammonia net value (=NH
3out-NH
3in).
The present invention provides favourable foundation for the production quality control of Ammonia Production factory: owing to adopt neural network model to predict, improved the precision of prediction and the fault freedom of ammonia convertor inlet, outlet ammonia content greatly, overcome because the measurement of ammonia convertor inlet, outlet ammonia content lags behind to the influence of production run and product quality.Ammonia convertor inlet, outlet ammonia content soft-sensing model are carried out the flow process of on-line prediction and model tuning referring to Fig. 4.
The present invention is further elaborated by the following examples, and its purpose only is better to understand content of the present invention.
Embodiment 1
Go into the ammonia content (NH in the tower nitrogen and hydrogen mixture
3in) soft measurement:
Go into tower nitrogen and hydrogen mixture hydrogen richness according to what from the DCS system of Ammonia Production device, read in real time, the methane content of dropping a hint, go into tower nitrogen and hydrogen mixture pressure, go into to choose in the tower nitrogen and hydrogen mixture temperature corresponding above-mentioned variable constantly as the input of neural network, the manual analysis value of going into ammonia content in the tower nitrogen and hydrogen mixture in previous 8 hours periods is as the foundation of judging whether soft-sensing model is proofreaied and correct, neural network training, thus trained and one group of weights that predicated error is less after obtain the real-time soft measurement predicted value that ammonia convertor is gone into ammonia content in the tower nitrogen and hydrogen mixture.
Value (the H that reads in real time by the DCS system for certain Ammonia Production factory day below
2, P, T, CH
4) and NH
3in360 groups of data being formed of manual analysis value, by formula (1) through 3 σ criterions reject 18 groups of data of apparent error are arranged after, obtain 342 groups of data.Through the single order digital filtering, by formula after (2) normalization, wherein 242 groups of data are used for neural network training, are used to test the generalization ability of neural network with all the other 100 groups of data.BP neural network model input number of nodes is 4, and the middle layer node number is 4, and the output node number is 1, and iterations was 200 generations.The one group of neural network weight and the threshold value that obtain after training are as follows:
η
1=0.1546;η
2=-0.0845;η
3=0.0962;η
4=-0.1919;
θ
11=24.7610;θ
12=-29.4720;θ
13=11.0857;θ
14=-16.1016
θ
2=0.7729;
As the input variable H of the model of choosing
2=60.99%, CH
4=18.96%, P=27.71MPa, T=26.36 ℃, H
2Variation range be [60,65], CH
4Variation range be [16,20], the variation range of P is [25,30], the variation range of T is [25,30], through the single order digital filtering, by formula after (2) normalization, numerical value is respectively 0.1982,0.7391,0.5413,0.2727.
inO
11=1/(1+exp(-innet
11));
inO
12=1/(1+exp(-innet
12));
inO
13=1/(1+exp(-innet
13));
inO
14=1/(1+exp(-innet
14));
innet
2=inO
11*η
1+inO
12*η
2+inO
13*η
3+inO
14*η
4+θ
2;
inO
2=innet
2;
With weights and the formula above the threshold value substitution, the variation range of inlet ammonia content is [2,3], carries out anti-normalized by formula (8) and gets: inModelOut=inO
2* (3-2)+2, then inModelOut is the NH by the prediction of BP neural network soft sensor model
3inActual value.
Ammonia content (NH in ammonia convertor " two go out " gas
3out) soft measurement:
Go into tower nitrogen and hydrogen mixture hydrogen richness according to what from the DCS system of Ammonia Production device, read in real time, the methane content of dropping a hint, go into tower nitrogen and hydrogen mixture pressure, go into tower nitrogen and hydrogen mixture flow, quench gas 1 flow, quench gas 2 flow values and the soft measured value of going into tower nitrogen and hydrogen mixture ammonia content, the manual analysis value of utilizing ammonia content in ammonia convertor " two go out " exit gas in previous 8 hours periods is as the foundation of judging whether soft-sensing model is proofreaied and correct, neural network training, thus trained and one group of weights that predicated error is less after obtain the real-time soft measurement predicted value of ammonia convertor " two go out " outlet ammonia content.
Be (corresponding) value (H that reads in real time by the DCS system on the same day of same Ammonia Production factory below with the sampling instant of inlet ammonia content soft-sensing model
2, CH
4, F, F
1, F
2, P), NH
3inSoft measured value and NH
3out360 groups of data that the manual analysis value constitutes, by formula (9) through 3 σ criterions reject 18 groups of data of apparent error are arranged after, with 342 groups of remaining data through the single order digital filtering, by formula after (10) normalization, wherein 242 groups are used for neural network training, with the generalization abilities of 100 groups of test neural networks.The one group of neural network weight and the threshold value that obtain after training are as follows:
ω
11=-0.0805;ω
12=-0.3182;ω
13=0.1523;ω
14=-1.5755;ω
15=0.7675;ω
16=1.0215;ω
17=-1.3626;
ω
21=-1.4075;ω
22=-1.1437;ω
23=-1.2269;ω
24=0.7893;ω
25=0.4678;ω
26=-1.3268;ω
27=-1.5907;
ω
31=0.8332;ω
32=0.6604;ω
33=0.9781;ω
34=0.5199;ω
35=0.8261;ω
36=1.6343;ω
37=-0.7909;
ω
41=-0.5343;ω
42=1.5865;ω
43=0.8559;ω
44=0.2185;ω
45=-0.9341;ω
46=0.3234;ω
47=-0.8237;
ω
51=-1.2447;ω
52=-0.2329;ω
53=1.4728;ω
54=-0.8013;ω
55=-0.9915;ω
56=1.4455;ω
57=0.9004;
ω
61=-1.0656;ω
62=1.5108;ω
63=0.9078;ω
64=-0.8394;ω
65=-0.9500;ω
66=1.6077;ω
67=1.9750;
ω
71=-1.3288;ω
72=-0.1280;ω
73=0.3306;ω
74=-1.3689;ω
75=1.0085;ω
76=-0.1122;ω
77=1.7784;
υ
1=-0.3326;υ
2=0.7010;υ
3=0.4200;υ
4=-0.9251;υ
5=0.0503;υ
6=0.6925;υ
7=1.9929;
b
11=-0.5939;b
12=0.7180;b
13=-2.1196;b
14=0.9492;b
15=0.0869;b
16=-0.5680;b
17=-0.6722;
b
2=-0.8439;
As the input variable H of the model of choosing
2=63.02%, CH
4=19.04%, F=187969.91m
3/ h, F
1=34867.51m
3/ h, F
2=32236.09m
3/ h, P=27.71MPa, NH
3in=2.7%, H
2Variation range be [60,65], CH
4Variation range be [16,20], the variation range of F is [150000,200000], F
1Variation range be [30000,40000], F
2Variation range be [250000,30000], the variation range of P is [25,30], NH
3inVariation range be [2,3].Through the single order digital filtering, by formula after (10) normalization, numerical value is respectively 0.6042,0.7592, and 0.7594,0.4868,0.7236,0.5418,0.7.
net
11=H
2*ω
11+CH
4*ω
21+F*ω
31+F
1*ω
41+F
2*ω
51+P*ω
61+NH
3in*ω
71+b
11;
net
12=H
2*ω
12+CH
4*ω
22+F*ω
32+F
1*ω
42+F
2*ω
52+P*ω
62+NH
3in*ω
72+b
12;
net
13=H
2*ω
13+CH
4*ω
23+F*ω
33+F
1*ω
43+F
2*ω
53+P*ω
63+NH
3in*ω
73+b
13;
net
14=H
2*ω
14+CH
4*ω
24+F*ω
34+F
1*ω
44+F
2*ω
54+P*ω
64+NH
3in*ω
74+b
14;
net
15=H
2*ω
15+CH
4*ω
25+F*ω
35+F
1*ω
45+F
2*ω
55+P*ω
65+NH
3in*ω
75+b
15;
net
16=H
2*ω
16+CH
4*ω
26+F*ω
36+F
1*ω
46+F
2*ω
56+P*ω
66+NH
3in*ω
76+b
16;
net
17=H
2*ω
17+CH
4*ω
27+F*ω
37+F
1*ω
47+F
2*ω
57+P*ω
67+NH
3in*ω
77+b
17;
O
11=1/(1+exp(-net
11));
O
12=1/(1+exp(-net
12));
O
13=1/(1+exp(-net
13));
O
14=1/(1+exp(-net
14));
O
15=1/(1+exp(-net
15));
O
16=1/(1+exp(-net
16));
O
17=1/(1+exp(-net
17));
net
2=O
11*υ
1+O
12*υ
2+O
13*υ
3+O
14*υ
4+O
15*υ
5+O
16*υ
6+O
17*υ
7+b
2;
O
2=1/(1+exp(-net
2));
With weights and formula above the threshold value substitution, the variation range of outlet ammonia content is [12,14], carries out anti-normalized by formula (17) and gets: ModelOut=O
2* (14-12)+12, then ModelOut is for to record NH in advance by the BP neural network soft sensor model
3outActual value.
Current time " ammonia net value " predicted value=NH
3out(ModelOut)-NH
3in(inModelOut).
The condition of above-mentioned requirements all can satisfy, so this invention has universality in most middle-size and small-size Ammonia Production device.
Claims (7)
1. the online soft sensor method of " ammonia net value " in the ammonia synthesis process of an employing " one three footpath cold shock formula " ammonia convertor comprises the steps:
(1) choose in the ammonia synthesis process of employing " one three footpath cold shock formula " ammonia convertor, go into hydrogen richness in the tower nitrogen and hydrogen mixture, go into the pressure of tower nitrogen and hydrogen mixture, (brief note is CH to go into the temperature (brief note is T) of tower nitrogen and hydrogen mixture and the middle methane content of dropping a hint
4) the on-line measurement value, with its normalized, go into the input variable of the BP neural network of ammonia content in the tower nitrogen and hydrogen mixture as ammonia convertor, the output variable of BP neural network promptly gets ammonia content in the tower nitrogen and hydrogen mixture after anti-normalized (brief note is NH
3in) soft measured value;
(2) choose in the ammonia synthesis process of employing " one three footpath cold shock formula " ammonia convertor H
2, P, CH
4, go into the flow (brief note for F) of tower nitrogen and hydrogen mixture, (brief note is F to the flow of a quench gas
1) and the flow of two quench gas (brief note is F
2) the on-line measurement value and the NH that obtains by step (1)
3inSoft measured value, with its normalized, (brief note is for NH as ammonia content in ammonia convertor " two go out " gas
3out) the input variable of BP neural network, the output variable of BP neural network after anti-normalized the NH of ammonia convertor
3outSoft measured value;
(3) calculating is by the NH of step (2) gained
3outSoft measured value with by the NH of step (1) gained
3inThe difference of soft measured value, online " ammonia net value " in the ammonia synthesis process of " one three footpath cold shock formula " ammonia convertor.
2. online soft sensor method as claimed in claim 1 is characterized in that, ammonia convertor described in the step (1) is gone in the BP neural network model of ammonia content in the tower nitrogen and hydrogen mixture, the node number of input layer is 4, the hidden layer number of plies in middle layer is 1, and the number of hidden nodes is 4, and output layer node number is 1.
3. online soft sensor method as claimed in claim 2 is characterized in that, adopts the single order digital filtering technique to carry out The disposal of gentle filter to the on-line measurement value of the input variable of BP neural network in the step (1).
4. online soft sensor method as claimed in claim 3 is characterized in that, wherein, per 8 hours, adopts NH
3inThe manual analysis value soft-sensing model of BP neural network in the step (1) carried out online " rollings " optimize correction once.
5. online soft sensor method as claimed in claim 1 is characterized in that, in the BP neural network model to ammonia content in the gas of ammonia convertor " two go out " described in the step (2), the node number of input layer is 7, the hidden layer number of plies in middle layer is 1, and the number of hidden nodes is 7, and output layer node number is 1.
6. online soft sensor method as claimed in claim 5 is characterized in that, adopts the single order digital filtering technique to carry out The disposal of gentle filter to the on-line measurement value of the input variable of BP neural network in the step (2).
7. online soft sensor method as claimed in claim 6 is characterized in that, wherein, per 8 hours, adopts NH
3inAnd NH
3outThe manual analysis value soft-sensing model of BP neural network in the step (2) carried out online " rollings " optimize correction once.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN 201010249988 CN101930008B (en) | 2010-08-11 | 2010-08-11 | Soft measurement method of 'ammonia net value' in ammonia synthesis process |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN 201010249988 CN101930008B (en) | 2010-08-11 | 2010-08-11 | Soft measurement method of 'ammonia net value' in ammonia synthesis process |
Publications (2)
Publication Number | Publication Date |
---|---|
CN101930008A true CN101930008A (en) | 2010-12-29 |
CN101930008B CN101930008B (en) | 2013-05-15 |
Family
ID=43369302
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN 201010249988 Expired - Fee Related CN101930008B (en) | 2010-08-11 | 2010-08-11 | Soft measurement method of 'ammonia net value' in ammonia synthesis process |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN101930008B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102184452A (en) * | 2011-03-04 | 2011-09-14 | 华东理工大学 | Soft measurement method for component gas of 'Texaco synthesis gas' |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20040226392A1 (en) * | 2003-03-10 | 2004-11-18 | Sensor Wireless Incorporated | Apparatus for detecting and reporting environmental conditions in bulk processing and handling of goods |
CN101158674A (en) * | 2007-11-15 | 2008-04-09 | 天津市市政工程设计研究院 | Method for predicting chlorophyll a concentration in water based on BP nerval net |
-
2010
- 2010-08-11 CN CN 201010249988 patent/CN101930008B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20040226392A1 (en) * | 2003-03-10 | 2004-11-18 | Sensor Wireless Incorporated | Apparatus for detecting and reporting environmental conditions in bulk processing and handling of goods |
CN101158674A (en) * | 2007-11-15 | 2008-04-09 | 天津市市政工程设计研究院 | Method for predicting chlorophyll a concentration in water based on BP nerval net |
Non-Patent Citations (4)
Title |
---|
刘卓倩,等: "基于智能集成优化的合成塔入口氨含量软测量", 《化工学报》 * |
宋哲英,等: "氨合成塔温度分级递阶智能控制", 《仪器仪表学报》 * |
李振民,等: "基于BP网络的大型氨厂转化系统操作模拟", 《大氮肥》 * |
王晓晔等: "神经网络自学习模糊控制及其在合成氨生产中的应用", 《控制与决策》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102184452A (en) * | 2011-03-04 | 2011-09-14 | 华东理工大学 | Soft measurement method for component gas of 'Texaco synthesis gas' |
CN102184452B (en) * | 2011-03-04 | 2016-02-24 | 华东理工大学 | The flexible measurement method of the component gas of " Texaco syngas " |
Also Published As
Publication number | Publication date |
---|---|
CN101930008B (en) | 2013-05-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110368808B (en) | Ammonia spraying amount control method and system for SCR flue gas denitration system | |
CN109681299B (en) | Engine control method of Urea-SCR system of diesel engine | |
Zhang et al. | A weighted auto regressive LSTM based approach for chemical processes modeling | |
CN101625353B (en) | Soft measurement method of outflow water quality of sewage treatment and on-line intelligent detecting instrument | |
CN104651559B (en) | Blast furnace liquid iron quality online forecasting system and method based on multivariable online sequential extreme learning machine | |
CN103534453A (en) | Method for model-based determination of the temperature distribution of an exhaust gas post-treatment unit | |
EP2350748A1 (en) | Optimizing refinery hydrogen gas supply, distribution and consumption in real time | |
CN102254057A (en) | Method for predicting rolling off-line mechanical property of thin plate | |
CN101176063A (en) | Facility monitoring method | |
CN104750902B (en) | Molten steel quality dynamic soft-measuring method based on multi output support vector regression | |
Zahedi et al. | Hybrid artificial neural network—First principle model formulation for the unsteady state simulation and analysis of a packed bed reactor for CO2 hydrogenation to methanol | |
CN102213949A (en) | Method for optimizing value of ethylene device | |
CN102830625A (en) | Process control system and method based on neural network predictive control | |
CN107045575A (en) | Aero-engine performance model modelling approach based on self-adjusting Wiener model | |
CN109670625A (en) | NOx emission concentration prediction method based on Unscented kalman filtering least square method supporting vector machine | |
CN108985574A (en) | A kind of polypropylene melt index flexible measurement method based on selective ensemble extreme learning machine | |
CN105425583A (en) | Control method of penicillin production process based on cooperative training local weighted partial least squares (LWPLS) | |
CN108304886A (en) | Aeroengine Performance Monitoring method | |
CN101930008B (en) | Soft measurement method of 'ammonia net value' in ammonia synthesis process | |
CN102967350A (en) | Calibration device and calibration method for automobile air mass flow meter | |
CN106706491B (en) | Intelligent detection method for membrane bioreactor-MBR water permeability | |
CN102129242A (en) | Product quality control method during batch processing production process based on two-layer hybrid intelligent optimization | |
KR20080033419A (en) | Real-time operating optimized method of multi-input and multi-output continuous manufacture procedure | |
Glielmo et al. | A machine learning approach to modeling and identification of automotive three-way catalytic converters | |
CN102759602A (en) | Method for predicting faults in cascade polymerization reaction process of high density polyethylene (HDPE) |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20130515 Termination date: 20150811 |
|
EXPY | Termination of patent right or utility model |