CN106777466A - High sulfur-containing natural gas based on ST UPFNN algorithms purify the dynamic evolutionary modeling method of technique - Google Patents

Abstract

The high sulfur-containing natural gas based on ST UPFNN algorithms that the present invention is provided purify the dynamic evolutionary modeling method of technique, including choose the technological parameter of influence desulfuration efficiency and the performance indications of desulfurization unit；The technological parameter of Preset Time and the data of the performance indications are gathered, sample set is formed after rejecting error sample；Sample set is normalized, normalization sample set is formed, training sample and test sample are chosen from normalization sample set；Neural network model is built based on training sample and original state variable is determined；Estimate optimum state variable using ST UPFNN algorithms；Using optimum state variable as the connection weight and threshold value of neural network model, the neural network model after weight threshold updates is obtained；Test sample is input to the neural network model after updating, is predicted the outcome, be compared predicting the outcome with the reality output in test sample, if comparative result is less than preset error value, constructed neural network model is effective.

Description

High sulfur-containing natural gas based on ST-UPFNN algorithms purify the dynamic evolutionary modeling of technique Method

Technical field

The present invention relates to high sulfur-containing natural gas purification techniques field, more specifically, it is related to a kind of based on ST-UPFNN calculations The high sulfur-containing natural gas of method purify the dynamic evolutionary modeling method of technique.

Background technology

High sulfur-containing natural gas acidic components content is higher by several times than conventional gas, its sweetening process amine liquid internal circulating load is big, Technological process is complicated, high energy consumption.Statistics shows that desulfurization unit energy consumption accounts for high sulfur-containing natural gas purification plant total energy consumption more than 50%, Its unit comprehensive energy consumption is up to 1729.3MJt-1, belongs to highly energy-consuming unit.For large-scale purification plant, by desulfurization unit Optimize capable of reducing energy consumption 5%~10%.Additionally, high sulfur-containing natural gas acidic components concentration is high, by the product tolerance after purification Relative raw material throughput is decreased significantly.Therefore, carrying out process optimization to high sulfur-containing natural gas sweetening process, realize that energy-conservation drops Consumption, it is very necessary to improve yield and gas processing economic benefit.

The content of the invention

In view of the above problems, it is an object of the invention to provide a kind of high sulfur-containing natural gas purification based on ST-UPFNN algorithms The dynamic evolutionary modeling method of technique, to solve the problems, such as that above-mentioned background technology is proposed.

The high sulfur-containing natural gas based on ST-UPFNN algorithms that the present invention is provided purify the dynamic evolutionary modeling method of technique, Including：

Step S1：The technological parameter of selection influence desulfuration efficiency and the performance indications of desulfurization unit；Wherein, the technique ginseng Number includes poor amine flow quantity, poor amine flow quantity, unstripped gas treating capacity, the tail gas into two-level absorption tower into tail gas absorber Unit returns to half rich amine flow quantity of desulfurization unit, the amine liquid of first grade absorption tower enters tower temperature degree, the amine liquid of two-level absorption tower enters tower Temperature, flash tank pressure, the steam consumption of reboiler, the steam consumption of another reboiler and vapor preheater Steam consumption；The performance indications of the desulfurization unit include H in purified gas₂S and CO₂Concentration and purified gas yield；

Step S2：The technological parameter of Preset Time and the data of the performance indications are gathered, is formed after rejecting error sample Sample set [X, Y]；

Step S3：Sample set [X, Y] is normalized, normalization sample set is formedTake normalization sample setIn preceding 80% sample as training sample, and remaining 20% sample is used as test sample；

Step S4：The original state variable X of neural network model and neural network model is built based on training sample, with And, by training sampleAs the input of neural network model, by training sampleAs the defeated of neural network model Go out；

Wherein, neural network model is：

Wherein, I_kIt is the vector sample value of training sample, and as the input of neural network model,It is network input layer To the connection weight of the neuron of hidden layer,It is the threshold value of the neuron of network input layer to hidden layer,For hidden layer is arrived The connection weight of the neuron of network output layer,It is hidden layer to the threshold value of the neuron of network output layer, wherein, i=1, 2…S₀；J=1,2 ... S₁；K=1,2 ... S₂；S₀It is the quantity of the neuron of network input layer, S₁It is the neuron of network hidden layer Quantity, S₂It is the quantity of the neuron of network output layer；

Original state variable is：

Step S5：The optimum state variable of neural network model is estimated using ST-UPFNN algorithms；

Step S6：Using optimum state variable as neural network modelWithFormula (1) is carried out Update, obtain the neural network model after training sample updates；

Step S7：By in test sampleThe neural network model after updating is input to, is predicted the outcome, will predicted Reality output in result and test sampleIt is compared, if comparative result is less than preset error value, constructed nerve Network model is effective；Otherwise repeat the above steps S1-S7, untill comparative result is less than the preset error value.

The high sulfur-containing natural gas based on ST-UPFNN algorithms that the present invention is provided purify the dynamic evolutionary modeling method of technique, Be capable of the whole optimal economic benefit route of effectively tracks of device, effectively overcome process to disturb, equipment performance change, economic effect The variation issue of benefit and productive target.

Brief description of the drawings

By reference to the explanation below in conjunction with accompanying drawing and the content of claims, and with to it is of the invention more comprehensively Understand, other purposes of the invention and result will be more apparent and should be readily appreciated that.In the accompanying drawings：

Fig. 1 a- Fig. 1 c are the fitting precision figure of training sample；

Fig. 2 a- Fig. 2 c are the measuring accuracy figure of test sample；

Fig. 3 is the trueness error figure of test sample and training sample.

Specific embodiment

Name Resolution

ST-UKFNN：Strong TrackUnscentedKalman FilterNeural Network, follow the trail of without mark by force Kalman filtering neutral net；

ST-UPFNN：Strong TrackUnscentedParticle FilterNeuralNetwork, follow the trail of without mark by force Particle filter neutral net, be combined for ST-UKFNN, particle filter (Particle Filter), BP neural network by it.

The high sulfur-containing natural gas based on ST-UPFNN algorithms that the present invention is provided purify the dynamic evolutionary modeling method of technique, Including：

Step S1：The technological parameter of selection influence desulfuration efficiency and the performance indications of desulfurization unit；Wherein, technological parameter bag Include into tail gas absorber poor amine flow quantity, poor amine flow quantity, unstripped gas treating capacity, tail gas unit into two-level absorption tower Half rich amine flow quantity of return desulfurization unit, first grade absorption tower amine liquid enter tower temperature degree, two-level absorption tower amine liquid and enter tower temperature degree, flash distillation The steam consumption of pressure tank, the steam consumption, the steam consumption of another reboiler and vapor preheater of reboiler Amount；The performance indications of desulfurization unit include H in purified gas₂S and CO₂Concentration and purified gas yield.As shown in table 1：

Table 1

Step S2：The technological parameter of Preset Time and the data of performance indications are gathered, sample is formed after rejecting error sample Collection [X, Y].Sample set [X, Y] is as shown in table 2 below：

Table 2

Step S3：Sample set [X, Y] is normalized, normalization sample set is formedTake normalization sample setIn preceding 80% sample as training sample, and remaining 20% sample is used as test sample.

Step S4：The original state variable X of neural network model and neural network model is built based on training sample, with And, by training sampleInput as neural network model isBy in training sampleAs neutral net The output of model is

Wherein, neural network model is：

Wherein, I_kIt is the vector sample value of training sample, and as the input of neural network model,It is network input layer To the connection weight of the neuron of hidden layer,It is the threshold value of the neuron of network input layer to hidden layer,For hidden layer is arrived The connection weight of the neuron of network output layer,It is hidden layer to the threshold value of the neuron of network output layer, wherein, i=1, 2…S₀；J=1,2 ... S₁；K=1,2 ... S₂；S₀It is the quantity of the neuron of network input layer, S₁It is the neuron of network hidden layer Quantity, S₂It is the quantity of the neuron of network output layer；

Original state variable is：

Step S5：The optimum state variable of neural network model is estimated using ST-UPFNN algorithms.

The present invention estimates the state variable of the neural network model using ST-UPFNN algorithms, with reach connection weight, The continuous adjustment of threshold value, requires until meeting.The optimum state variable that will be obtained is used as above-mentioned set up neural network model Connection weight, threshold value.It should be noted that the connection weight, threshold value be by ST-UPFNN algorithms adjust after connection weight, Threshold value, is also whole connection weights and threshold value of above-mentioned set up neural network model, includingWith

The process of the optimum state variable of neural network model is estimated using ST-UPFNN algorithms to be included：

Step S51：The number N of particle is set for particle filter, and it is rightSampling, obtains primary collectionAnd each particle for concentrating the primaryWeights be set to 1/N.

Wherein,Represent with x₀It is average, P₀For the normal distribution of variance is adopted.

Step S52：Obtaining the observational variable value at (k+1) momentAfterwards, using ST-UKFNN algorithms to each particleThe state estimation at (k+1) moment is carried out, optimal State Estimation value is obtainedAnd covariance

Using ST-UKFNN algorithms to each particleThe process of state estimation is carried out, including：

Step S521：Sigma samplings are carried out to the original state variable X, 2n+1 sampled point, initialization control is obtained 2n+1 the distribution parameter alpha of sampled point, parameter κ to be selected, and non-negative right factor beta, to the original state variable X's Sigma samplings are as follows：

Step S522：The weight of each sampled point is calculated, the weight of each sampled point is as follows：

Wherein, W_cTo calculate the weight of the covariance of state variable, W_mPower during to calculate state estimation and observation prediction Weight,It isFirst row,It isFirst row.

Step S523：By the state equation of Discrete time Nonlinear Systems by the optimal shape at the k moment of each sampled point The state estimation of state variable is transformed to the state estimation of the state variable at (k+1) momentAnd, by merging (k+1) The vector of the state estimation at moment, obtains the state prior estimate of the state variable at (k+1) momentWith covariance P_k+1|k； Wherein,

(k+1) state estimation of the state variable at momentFor：

Wherein,Represent the optimal estimation at k moment, w_kIt is process noise, its covariance matrix Q_kIt is cov (w_k,w_j)= Q_kδ_kj,

(k+1) the state prior estimate of the state variable at momentFor：

(k+1) the covariance P of the state variable at moment_k+1|kFor：

Step S524：The state variable at (k+1) moment that will be obtained by the observational equation of Discrete time Nonlinear Systems State estimationIt is converted into the observation prediction at (k+1) moment

Wherein, ν_kIt is observation noise, its covariance matrix R_kIt is cov (v_k,v_j)=R_kδ_kj,

Step S525：By and estimate (k+1) moment observation predictVector, obtain (k+1) moment priori Observation predictionAnd prediction is observed according to prioriEstimate the covariance of (k+1) moment observation prediction

(k+1) the priori observation prediction at momentFor：

(k+1) the prediction covariance of moment observation predictionFor：

Wherein, strong tracing algorithm, i.e. fading factor λ are introduced herein_k+1Strengthen the trace ability of model to improve model essence Degree；

N_k+1=V_k+1-βR_k+1 (12)

Wherein, β is the reduction factor, β >=1；

Step S526：Calculate the state prior estimate of the state variable at (k+1) momentWith the priori at (k+1) moment Observation predictionBetween state covarianceState covarianceFor：

Step S527：By setting up state covarianceWith prediction covarianceRelation, update (k+1) when The state estimation and covariance of the state variable at quarter, obtain the optimal State Estimation value at (k+1) moment respectivelyAnd covariance

Wherein, the state covariance of foundationWith prediction covarianceRelation be：

Wherein, K_k+1It is gain matrix, and,

The state estimation of the state variable at (k+1) moment after renewalFor：

The covariance P of the state variable at (k+1) moment after renewal_k+1For：

The state estimation of the state variable at (k+1) moment after by renewalWith covariance P_k+1During respectively as (k+1) The optimal State Estimation value at quarterAnd covariance

Step S53：By optimal State Estimation valueWith the covariance after initializationAs the importance density of particle Function is sampled, and obtains new particleBy all new particlesThe particle collection of composition In each new particleNormal distribution probability density value it is as follows：

(rand is normal distribution random error)

Density Function of Normal Distribution：

Wherein, N is each new particleNormal distribution, x, μ, σ are respectively three variables of normal distyribution function, p It is each new particleConditional probability；

In formula (2), x, μ, σ is corresponded respectively

In formula (3), x, μ, σ is corresponded respectively

In formula (4), x, μ, σ is corresponded respectively

Step S54：To new particleWeights be updated, and be normalized；Wherein,

Right value update formula is：

Weights normalize formula：

Step S55：According to the weights and resampling strategy after new particle normalized to particle collection Resampling is carried out, new particle collection is obtainedAnd ask for new particle collection In each new particleState estimation

If variable u, orderTake u₁∈(0,1)

Step S56：Using the number N of particle as the calculating process of cycle-index circulation step S51- steps S55, will be last Once estimate to obtain system state variables as the optimum state variable using ST-UPFNN algorithms estimation neural network model；Its In, by new particleState estimationPaid as the optimal estimation at this momentCarry out the shape of subsequent time State is estimated.

The structural parameters of the optimum state variable of neural network model are as follows：

Step S6：Using optimum state variable as neural network modelWithFormula (1) is carried out Update, obtain the neural network model after weight threshold updates.

Step S7：By in test sampleThe neural network model after updating is input to, is predicted the outcome, will predicted Reality output in result and test sampleIt is compared, if comparative result is less than preset error value, constructed nerve Network model is effective；Otherwise repeat the above steps S1-S7, untill comparative result is less than preset error value.

The present invention obtains following technique effect by several groups of tests：

Fig. 1 a- Fig. 1 c are the fitting precision figure of training sample, wherein, the technological parameter of Fig. 1 a influence desulfuration efficiencies is purification Gas H₂S concentration, the technological parameter of Fig. 1 b influence desulfuration efficiencies is purified gas CO₂Concentration, Fig. 1 c influence the technological parameter of desulfuration efficiency It is purified gas yield.

Fig. 2 a- Fig. 2 c are the measuring accuracy figure of test sample, wherein, the technological parameter of Fig. 2 a influence desulfuration efficiencies is purification Gas H₂S concentration, the technological parameter of Fig. 2 b influence desulfuration efficiencies is purified gas CO₂Concentration, Fig. 2 c influence the technological parameter of desulfuration efficiency It is purified gas yield.

Purified gas H₂S concentration, purified gas CO₂Within 2%, error is less than for concentration and purified gas yield relative error 10%, therefore model effectively.

The above, specific embodiment only of the invention, but protection scope of the present invention is not limited thereto, and it is any Those familiar with the art the invention discloses technical scope in, change or replacement can be readily occurred in, should all contain Cover within protection scope of the present invention.Therefore, protection scope of the present invention described should be defined by scope of the claims.

Claims (4)

1. a kind of high sulfur-containing natural gas based on ST-UPFNN algorithms purify the dynamic evolutionary modeling method of technique, including：

Step S1：The technological parameter of selection influence desulfuration efficiency and the performance indications of desulfurization unit；Wherein, the technological parameter bag Include poor amine flow quantity, poor amine flow quantity, unstripped gas treating capacity, the tail gas unit into two-level absorption tower into tail gas absorber Return desulfurization unit half rich amine flow quantity, the amine liquid of first grade absorption tower enters tower temperature degree, the amine liquid of two-level absorption tower enters tower temperature degree, The steam of flash tank pressure, the steam consumption, the steam consumption of another reboiler and vapor preheater of reboiler Consumption；The performance indications of the desulfurization unit include H in purified gas₂S and CO₂Concentration and purified gas yield；

Step S2：The technological parameter of Preset Time and the data of the performance indications are gathered, is formed after rejecting error sample Sample set [X, Y]；

Step S3：Sample set [X, Y] is normalized, normalization sample set is formedTake the normalization sample set In preceding 80% sample as training sample, and remaining 20% sample is used as test sample；

Step S4：Neural network model is built based on training sample training and original state variable X is determined according to model, will In the training sampleAs the input of the neural network model, by the training sampleAs the nerve The output of network model；

Wherein, the neural network model is：

$Y=\underset{j=1}{\overset{s_2}{\mathrm{\Σ}}}\left(f\right(\underset{i=1}{\overset{S_1}{\mathrm{\Σ}}}{w}_{ik}^1{I}_k+b_i^1\left)\right)\·w_{kj}^2+b_j^2---\left(1\right)$

Wherein, I_kIt is the vector sample value of the training sample, and as the input of the neural network model,For network is defeated Enter layer to the connection weight of the neuron of hidden layer,It is network input layer to the threshold value of the neuron of the hidden layer,For The hidden layer to the neuron of network output layer connection weight,It is the hidden layer to the nerve of the network output layer The threshold value of unit, wherein, i=1,2 ... S₀；J=1,2 ... S₁；K=1,2 ... S₂；S₀It is the number of the neuron of the network input layer Amount, S₁It is the quantity of the neuron of the network hidden layer, S₂It is the quantity of the neuron of the network output layer；

The original state variable is：

$X={\left[\begin{array}{cccccccc}{w}_{11}^{1}L& w_{s_0{s}_1}^1& b_1^{1}L& b_{s_1}^1& w_{11}^{2}L& w_{s_1{s}_2}^2& b_1^{2}L& b_{s_2}^2\end{array}\right]}^T---\left(2\right)$

Step S5：The optimum state variable of the neural network model is estimated using ST-UPFNN algorithms；

Step S6：Using the optimum state variable as the neural network modelWithFormula (1) is entered Row updates, and obtains the neural network model after weight threshold updates；

Step S7：By in the test sampleThe neural network model after updating is input to, is predicted the outcome, will be described Predict the outcome and the reality output in the test sampleIt is compared, if comparative result is less than preset error value, institute's structure The neural network model built is effective；Otherwise repeat the above steps S1-S7, until the comparative result is less than the preset error value Untill.

2. the high sulfur-containing natural gas based on ST-UPFNN algorithms as claimed in claim 1 purify the dynamic evolutionary modeling side of technique Method, wherein, the step S5 includes：

Step S51：The number N of particle is set for particle filter, it is rightIt is sampled, obtains primary collectionAnd each particle for concentrating the primaryWeights be set to 1/N；

Wherein,Represent with x₀It is average, P₀For the normal distribution of variance is sampled；

Step S52：Obtaining the observational variable value at (k+1) momentAfterwards, using ST-UKFNN algorithms to each particle The state estimation at (k+1) moment is carried out, optimal State Estimation value is obtainedAnd covariance

Step S53：By the optimal State Estimation valueWith the covariance after the initializationImportance as particle is close Degree function is sampled, and obtains new particleBy all new particlesThe particle collection of composition In each new particleNormal distribution probability density value it is as follows：

$X_{k+1}^{\′j}={\hat{X}}_{k+1}^j+rand---\left(3\right)$

Wherein, rand is normal distribution random error；

Density Function of Normal Distribution：

$p\left(X_{k+1}^{\′j}\right|{\hat{X}}_{k+1}^j,Y_{k+1})=N(X_{k+1}^{\′j},{\hat{X}}_{k+1}^j,P_{k+1}^j)---\left(5\right)$

$p\left(Y_{k+1}\right|X_{k+1}^{\′j})=N(Y_{ek,k+1}^j,h(X_{k+1}^{\′j},I_{k+1}),1)---\left(6\right)$

$p\left(X_{k+1}^{\′j}\right|{\hat{X}}_{k+1}^j)-N(X_{k+1}^{\′j},{\hat{X}}_{k+1}^j,1)---\left(7\right)$

Wherein, N is each new particleNormal distribution, x, μ, σ be respectively three variables of normal distribution, and p is new for each ParticleConditional probability；

Step S54：To new particleWeights be updated, and be normalized；Wherein,

Right value update formula is：

Weights normalize formula：

Step S55：According to the weights and resampling strategy after new particle normalized to particle collection Resampling is carried out, new particle collection is obtainedAnd ask for new particle collection In each new particleState estimation

If variable u, orderTake u₁∈(0,1)

$X_{k+1}^j=\left\{\begin{array}{c}{X}_{k+1}^{\&prime;c}u\left(j\right)<\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{k+1}^i,c=1,2L,N\\ X_{k+1}^{\&prime;j},else\end{array}\right.,j=1,2L,N---\left(10\right)$

${\hat{X}}_{k+1}=\left(\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{X}_{k+1}^j\right)/N---\left(11\right)$

Step S56：Using the number N of particle as the calculating process of cycle-index circulation step S51- steps S55, will last time The optimum state variable that estimation is obtained is used as the optimum state variable using the ST-UPFNN algorithms estimation neural network model； Wherein, by new particleState estimationPaid as the optimal estimation at this momentCarry out subsequent time State estimation.

3. the high sulfur-containing natural gas based on ST-UPFNN algorithms as claimed in claim 2 purify the dynamic evolutionary modeling side of technique Method, wherein, using ST-UKFNN algorithms to each particleThe process of state estimation is carried out, including：

Step S521：Sigma samplings are carried out to the original state variable X, 2n+1 sampled point, initialization control 2n+1 is obtained The distribution parameter alpha of individual sampled point, parameter κ to be selected, and non-negative right factor beta, to the Sigma of the original state variable X Sampling is as follows：

$\left\{\begin{array}{c}\begin{array}{cc}{\hat{X}}_{k|k}^{\left(i\right)}=X_k+\sqrt{(n+\mathrm{\&lambda;})p_k}& i=1:n\end{array}\\ \begin{array}{cc}{\hat{X}}_{k|k}^{\left(i\right)}=X_k-\sqrt{(n+\mathrm{\&lambda;})p_k}& i=n+1:2n\end{array}\\ \mathrm{\&lambda;}=a^2(n+\mathrm{\&kappa;})-n\end{array}\right.---\left(12\right)$

Step S522：The weight of each sampled point is calculated, the weight of each sampled point is as follows：

$\left\{\begin{array}{c}{W}_m^{\left(0\right)}=\mathrm{\&lambda;}/(n+\mathrm{\&lambda;})\\ W_c^{\left(0\right)}=\mathrm{\&lambda;}/(n+\mathrm{\&lambda;})+(1-{\mathrm{\&alpha;}}^2+\mathrm{\&beta;})\\ \begin{array}{cc}{W}_m^{\left(i\right)}=W_c^{\left(i\right)}=\mathrm{\&lambda;}/(2\&times;(n+\mathrm{\&lambda;}\left)\right)& i=1:2n\end{array}\end{array}\right.---\left(13\right)$

Wherein, W_cTo calculate the weight of the covariance of state variable, W_mWeight during to calculate state estimation and observation prediction,It isFirst row, W_c ⁽⁰⁾It is W_c ⁽ⁱ⁾First row；

Step S523：The optimum state at the k moment of each sampled point is become by the state equation of Discrete time Nonlinear Systems The state estimation of amount is transformed to the state estimation of the state variable at (k+1) momentAnd by merging the shape at (k+1) moment The vector that state is estimated, obtains the state prior estimate of the state variable at (k+1) momentWith covariance P_k+1|k；Wherein,

The state estimationFor：

$X_{k+1|k}^{\left(i\right)}=F\left({\hat{X}}_{k|k}^{\left(i\right)}\right)+w_k---\left(14\right)$

Wherein,Represent the optimal estimation at k moment, w_kIt is process noise, its covariance matrix Q_kIt is cov (w_k,w_j)=Q_kδ_kj,

The state prior estimateFor：

${\hat{X}}_{k+1|k}=\underset{i=0}{\overset{2n}{\mathrm{\&Sigma;}}}{W}_m^{\left(i\right)}\&CenterDot;X_{k+1|k}^{\left(i\right)}---\left(15\right)$

The covariance P of the state variable_k+1|kFor：

$P_{k+1|k}=\underset{i=0}{\overset{2n}{\mathrm{\&Sigma;}}}{W}_c^{\left(i\right)}\&CenterDot;\&lsqb;X_{k+1|k}^{\left(i\right)}-{\hat{X}}_{k+1|k}^{\left(i\right)}\&rsqb;{\&lsqb;X_{k+1|k}^{\left(i\right)}-{\hat{X}}_{k+1|k}^{\left(i\right)}\&rsqb;}^T+Q_k---\left(16\right)$

Step S524：The shape of the state variable at (k+1) moment that will be obtained by the observational equation of Discrete time Nonlinear Systems State is estimatedObservation with (k+1) moment is predictedSet up contact：

$Y_{k+1|k}^{\left(i\right)}=g\left(\underset{k=1}{\overset{s_2}{\mathrm{\&Sigma;}}}\right(f\left(\underset{j=1}{\overset{S_1}{\mathrm{\&Sigma;}}}{w}_{ji}^1{X}_{k+1|k}^{\left(i\right)}+b_j^1\right))\&CenterDot;w_{kj}^2+b_k^2)+v_k---\left(17\right)$

Wherein, ν_kIt is observation noise, its covariance matrix R_kIt is cov (v_k,v_j)=R_kδ_kj,

Step S525：Predicted by the observation for estimating (k+1) momentVector, the priori observation for obtaining (k+1) moment is pre- SurveyAnd prediction is observed according to prioriEstimate the covariance of (k+1) moment observation prediction

(k+1) the observation prediction at momentFor：

${\hat{Y}}_{k+1|k}=\underset{i=0}{\overset{2n}{\mathrm{\&Sigma;}}}{W}_m^{\left(i\right)}\&CenterDot;Y_{k+1|k}^{\left(i\right)}---\left(18\right)$

(k+1) covariance of moment observation predictionFor：

$P_{y_{k+1}}={\mathrm{\&lambda;}}_{k+1}\underset{i=0}{\overset{2n}{\mathrm{\&Sigma;}}}{W}_c^{\left(i\right)}\&CenterDot;(Y_{k+1|k}^{\left(i\right)}-{\hat{Y}}_{k+1|k}){(Y_{k+1|k}^{\left(i\right)}-{\hat{Y}}_{k+1|k})}^T+R_{k+1}---\left(19\right)$

Wherein,

N_k+1=V_k+1-βR_k+1 (21)

Wherein, β is the reduction factor, β >=1；

$M_{k+1}=\underset{i=0}{\overset{2n}{\mathrm{\&Sigma;}}}{W}_c^{\left(i\right)}\&CenterDot;(Y_{k+1|k}^{\left(i\right)}-{\hat{Y}}_{k+1|k}){(Y_{k+1|k}^{\left(i\right)}-{\hat{Y}}_{k+1|k})}^T+R_{k+1}---\left(22\right)$

$V_{k+1}=\left\{\begin{array}{cc}{e}_{k+1}{e}_{k+1}^T& k=0\\ \frac{{\mathrm{\&rho;V}}_k+e_{k+1}{e}_{k+1}^T}{1+\mathrm{\&rho;}}& k\&GreaterEqual;1\end{array}\right.,\mathrm{\&rho;}\&Element;(0,1)---\left(23\right)$

$e_{k+1}=Y_{k+1}-\underset{i=0}{\overset{2n}{\mathrm{\&Sigma;}}}{W}_m^{\left(i\right)}\&CenterDot;h({\hat{X}}_{k+1|k}^{\left(i\right)},I_k)---\left(24\right)$

Step S526：Calculate the state prior estimate of the state variable at (k+1) momentPriori observation with (k+1) moment is pre- SurveyBetween covarianceThe covarianceFor：

$P_{x_{k+1}{y}_{k+1}}={\mathrm{\&lambda;}}_{k+1}\underset{i=0}{\overset{2n}{\mathrm{\&Sigma;}}}{W}_c^{\left(i\right)}\&CenterDot;(X_{k+1|k}^{\left(i\right)}-{\hat{X}}_{k+1|k}){(Y_{k+1|k}^{\left(i\right)}-{\hat{Y}}_{k+1|k})}^T---\left(25\right)$

Step S527：By setting up covarianceAnd covarianceRelation, update the state variable at (k+1) moment State estimation and covariance, obtain the optimal State Estimation value at (k+1) moment respectivelyAnd covariance

4. the high sulfur-containing natural gas based on ST-UPFNN algorithms as claimed in claim 3 purify the dynamic evolutionary modeling side of technique Method, wherein, the covariance that S527 sets upAnd covarianceRelation be：

$K_{k+1}=\frac{P_{x_{k+1}{y}_{k+1}}}{P_{y_{k+1}}}---\left(26\right)$

Wherein, K_k+1It is gain matrix；And,

The optimal State Estimation of the state variable at (k+1) moment after renewalFor：

${\hat{X}}_{k+1}={\hat{X}}_{k+1|k}+K_{k+1}(Y_{k+1}-{\hat{Y}}_{k+1|k})---\left(27\right)$

The covariance P of the state variable at (k+1) moment after renewal_k+1For：

$P_{k+1}=P_{k+1|k}-K_{k+1}{\left(P_{x_{k+1}{y}_{k+1}}\right)}^T---\left(28\right)$

The state estimation of the optimum state variable at (k+1) moment after by renewalWith covariance P_k+1During respectively as (k+1) The optimal State Estimation value at quarterAnd covariance