CN103606006B - Sludge volume index (SVI) soft measuring method based on self-organized T-S fuzzy nerve network - Google Patents

Abstract

The invention discloses a sludge volume index (SVI) soft measuring method based on a self-organized T-S fuzzy nerve network and belongs to both the field of control and the field of sewage treatment. The accurate prediction of an SVI is the guarantee for normal operation of a sewage treatment process. The method comprises: first of all, taking the output quantity of a rule layer, i.e., the space activation intensity of the rule layer as a basis for determining whether a fuzzy rule is increased; secondly, on the basis of generating a new fuzzy rule, taking the output quantity of a membership function layer as a basis for determining whether a fuzzy set is increased; and finally obtaining a self-organized T-S fuzzy recursion nerve network by using a gradient decrease algorithm to adjust the weight value parameter of a model and the center value and width of a Gauss function, and establishing an SVI on-line soft measuring model based on an SOTSFEN such that real-time detection of the SVI is realized, and an effective method is provided for preventing sludge expansion.

Description

Sludge volume index flexible measurement method based on self-organizing T-S fuzzy neural networks

Technical field

The present invention sets up the soft-sensing model of sludge volume index SVI using the fuzzy Recursive Networks of self-organizing T-S, and it is right to realize The real-time estimate of sludge settling index S VI.The Accurate Prediction of sludge volume index SVI is the guarantor that sewage disposal process normally runs Card, the present invention had both belonged to control field, and sewage treatment area is belonged to again.

Background technology

Sewage disposal is the Important Action of Chinese government's water resources comprehensive utilization, is also the weight of China's strategy of sustainable development Want part.At present, national each city, county substantially establish urban wastewater treatment firm, sewage treatment capacity and U.S. etc. National developed country is suitable.But sewage disposal operation conditions allows of no optimist, wherein sludge bulking problem is seriously govern at sewage The development of reason.Once occurring, der Pilz amount reproduction, sludge settling property is deteriorated for sludge bulking, and separation of solid and liquid is difficult, causes Water overproof water quality, sludge overflows and is lost in, in some instances it may even be possible to draws barmy generation, causes sewage disposal system to collapse.Therefore, this The hard measurement research of the bright SVI based on the fuzzy recurrent neural network SOTSRFNN of self-organizing T-S is with a wide range of applications.

Sludge volume index SVI is one of important evaluation index of sludge settling property.At present, for the detection side of SVI Method mainly has two classes：1. manual detection method, is detected using graduated cylinder timing sampling, calculates SVI values, but the method takes and error Greatly, it is difficult to meet the increasingly complicated actual requirement of sewage disposal；There is high equipment cost, longevity in 2. automatic detection method, but the method Life it is short, stability difference the shortcomings of, and by site environment and it is manually-operated affect, accuracy of detection cannot be ensured.Hard measurement Technology is estimated using the relation between system change and parameter, the model set up between input and output by easily surveying water quality variable SVI values.With small investment, the time is short, be swift in response, it is easy to the advantages of maintaining and safeguard.Therefore, the soft survey of SOTSRFNN is studied Amount method measures in real time problem and has important practical significance to solution SVI.

The present invention proposes a kind of online soft sensor method of SVI：First, with the spatial activation intensity of rules layer, that is, advise The foundation that then whether the output of layer increases as judgement fuzzy rule；Secondly, on the basis of new fuzzy rule is generated, to be subordinate to The foundation whether membership fuction layer output quantity increases as judgement fuzzy set；Finally, the power of model is adjusted using gradient descent algorithm The central value and width of value parameter and Gaussian function, obtains a kind of fuzzy recurrent neural networks of self-organizing T-S, and is based on SOTSRFNN establishes the online soft sensor model of SVI, realizes the real-time detection of SVI, to prevent sludge bulking to provide one Plant effective ways.

The content of the invention

The present invention analyzes the Crack cause of sludge bulking for the difficult problem of SVI on-line measurements, summarizes close with SVI Related easy survey water quality parameter, using principle component analysis PCA the input quantity of model is determined；And propose a kind of improved fuzzy Recurrent neural network, based on structural self-organizing algorithm, devises SOTSRFNN, establishes the online soft sensor model of SVI；Most Afterwards, the hard measurement of SVI is carried out using the model set up, the on-line measurement of SVI is realized；

Present invention employs following technical scheme and realize step：

1 a kind of SVI flexible measurement methods, it is characterised in that comprise the following steps：

(1) data prediction and auxiliary variable are selected；

Sample set data zero-mean standardized method is normalized, and by pivot analysis PCA auxiliary change is carried out Selected, final determination mixed liquor concentration of suspension MLSS, acidity-basicity ph, aeration tank water temperature T, the aeration tank ammonia NH of amount₄As mould The input variable of type.

(2) the Recurrent Fuzzy Neural Network model of the hard measurement of SVI is set up, input quantity is MLSS, pH, T, NH₄, model Output quantity is SVI.Recurrent Fuzzy Neural Network topological structure：Input layer is ground floor, the membership function layer i.e. second layer, rules layer I.e. third layer, parameter layer are the 4th layer, output layer i.e. layer 5, feedback layer.

Neural network structure：Input layer has 4 input nodes, and each input layer connects m membership function node layer, Rules layer has m node, and m represents number of fuzzy rules, and the number of fuzzy rules generated in network structure training process determines m's Value, parameter layer node number, feedback layer nodes are equal with the nodes of rules layer, and output layer has 1 node.x=[x₁,x₂, x₃,x₄] represent neutral net input, y_dRepresent the desired output of neutral net.If kth group sample data is x (k)=[x₁ (k),x₂(k),x₃(k),x₄(k)].When kth group sample data is input into：

The output of i-th node of input layer is expressed as：

Wherein,Represent output of i-th node of input layer when kth group sample is input into；

The node total number of membership function layer is：4m, the node of each input layer is all connected with the node of m membership function layer, Membership function layer is output as：

Wherein, membership function adopts Gaussian function, c_ijAnd σ_ijRepresent that i-th node of input layer is corresponding to be subordinate to letter respectively The central value and width of the Gaussian function of several layers of j-th node, each Gaussian function is a fuzzy set, in structural adjustment Stage assigns initial value to Gaussian function.Represent j-th node of the corresponding membership function layer of i-th node of input layer defeated Enter output during kth group sample；

The nodes of rules layer are m, and j-th node of the corresponding membership function layer of each node of input layer is all connected to J-th node of rules layer.Feedback link is introduced in rules layer, in feedback layer built-in variable is added.Feedback node layer includes two The node of type：Node is accepted, summation operation is weighted to the output quantity and weights of rules layer, calculate built-in variable；Instead Feedback node, using sigmoid functions as membership function, calculates the output quantity of feedback layer.Rules layer each node is all connected with institute Some feedback layers accept node.The undertaking node of feedback layer is corresponded with feedback node, and nodes are equal, feedback node and rule Then node layer is corresponded, and nodes are equal, and change with the change of number of fuzzy rules, consistently equal to number of fuzzy rules. J-th node of rules layer is output as：

(5)

Wherein, ω_jqIt is the connection weight of j-th node with q-th undertaking node of feedback layer of rules layer, initial assignment For the random number between 0 to 1.Output quantity of j-th node of rules layer in the group sample of kth -1 is represented,h_qTable Show the built-in variable of q-th undertaking node of feedback layer.f_qRepresent the output quantity of q-th feedback node of feedback layer.Rules layer it is defeated Go outThat is the intensity of activation of fuzzy rule, whereinRepresentation space intensity of activation, f_qExpression time intensity of activation.

The node of rules layer is one-to-one with the node of parameter layer, and the output that parameter layer has m node, this layer is represented For：

(7)

Wherein, a_ijLinear dimensions is represented, initial value is entered as 0 to 1 random number, a_j=[a_1j,a_2j,a_3j,a_4j]。W_j(k) table Show the value of linear dimensions weighted sum of j-th node of parameter layer when kth group sample is input into,Represent the jth of parameter layer Output of the individual node in kth group sample.

Network model is multiple input single output, and output layer has 1 node, and all of parameter node layer is connected to output section Point.Network output is expressed as：

Wherein, y (k) represents the network output of kth sample.

(3) fuzzy neural network determines network structure by structural self-organizing adjustment first.The structural self-organizing of network is adjusted It is whole：x(k)=[x₁(k),x₂(k),x₃(k),x₄(k)] represent that model is currently input into kth group sample, open from first group of sample data Begin to be input into, complete to total data input, be often input into one group of sample data and judge whether to increase new mould using following steps Paste rule, judges whether to increase new fuzzy set on the basis of fuzzy rule is increased.

1. number of fuzzy rules is 0 in network model initial configuration, first group of data input, and number of fuzzy rules is changed into 1, and increases Plus new fuzzy set.The initialization of fuzzy set, the i.e. initial table of Gaussian function central value and width value is shown as：

c(1)=x(1)=[x₁(1),x₂(1),x₃(1),x₄(1),] (9)

σ(1)=[σ₁₁,σ₂₁,σ₃₁,σ₄₁]=[0.5,0.5,0.5,0.5] (10) its Middle c (1) represents the central value of the first group of fuzzy set for generating, the i.e. central value of membership function.σ (1) represents first for generating The width value of the width value of fuzzy set, i.e. membership function.

2. input data sequentially inputs, and is often input into one group of data, judges whether to increase new fuzzy rule, judgment formula It is expressed as：

WhereinRepresent the spatial activation intensity of j-th node of rules layer.J is represented and worked asThe value of j when taking maximum.N is represented Present Fuzzy rule number.For threshold value set in advance, value is 0.24.

If meeting formula (13), increase a new fuzzy rule, 3. N'=N+1 performs.

If being unsatisfactory for formula (13), repeat 2., to be input into next group of data；

3. on the basis of a new fuzzy rule is increased, judge whether to increase new fuzzy set, judgment formula is represented For：

(15)

Wherein I is represented and worked asThe value of j when taking maximum, h represents the fuzzy set number of "current" model, h=N.I_thTo set in advance Fixed threshold value, value is 0.92.

If meeting formula (15), increase a new fuzzy set, h'=h+1, h'=N'.Fuzzy set to increasing newly assigns initial value Newly-increased Gaussian function central value and width value i.e. to membership function layer assigns initial value.Initial table is shown as：

c_N+1=x(k) (16)

Wherein, N represents the existing number of fuzzy rules of "current" model, p=1,2 ..., N, c_N+1Represent newly-increased membership function The initial value of central value, r represents overlap coefficient, and value is 0.6.c_pIt is expressed as the center of p-th Gaussian function in "current" model Value.c⁺Represent and work as x (k) and c_pC when space length is minimum_pValue.σ_i,N+1Represent the width initial value of newly-increased membership function, input Amount x (k) and c⁺Difference absolute value and overlap coefficient product.Increase new fuzzy set, i.e., in each input layer pair The membership function layer answered increases a new node, and rules layer, feedback layer, parameter layer respectively increase accordingly corresponding node.

If being unsatisfactory for (15), new fuzzy set h'=h, and N''=N'-1 are not increased, the fuzzy rule that will be newly increased Then delete.

4. the structure for adjusting neutral net is continued, input data sequentially inputs, repeats 2. 3., and all input datas have been input into Cheng Hou, neural network structure adjusting training is completed.

(4) structure of network model is determined, then network parameter is just adjusted, with the data after correction god is trained Jing networks, wherein totally 150 groups of sample datas, 90 groups of training sample data, 60 groups of test sample data, train epochs are 1000 Step, in the training process of each step, 90 groups of training sample data are fully entered, and every group of training sample data input adopts gradient Descent algorithm is adjusted to network parameter, wherein the parameter for adjusting includes：The linear dimensions a of parameter layer, rules layer is to feedback Connection weight ω of layer, central value c and width value σ of membership function layer Gaussian function.

The object function of definition training is that systematic error is defined as：

Wherein y (k) for network reality output amount, y_dK () is the desired throughput of network, E (k) represents systematic error.

Gradient descent algorithm：Partial derivative is asked using output quantity of the object function to each layer, the error of per layer of output is calculated Propagate item.Partial derivative, partial derivative of the calculating target function to parameter value, partial derivative is asked to be the tune of each parameter by compound function Whole amount.

The error propagation item δ of parameter layer⁽⁴⁾, i.e., the partial derivative that object function is exported to parameter layer：

To linear dimensions a_jAdjustment amount be：

Rules layer error propagation itemIt is as follows：

Recurrence layer is to rules layer connection weight ω_jqRegulation rule it is as follows：

The error propagation item of membership function layerIt is as follows：

Central value c of membership function layer Gaussian function_ijRegulation rule it is as follows：

The width cs of membership function layer Gaussian function_ijRegulation rule it is as follows：

Wherein η is the learning rate of parameter, and value is 0.15.

(5) test sample is predicted, using test sample data as the neutral net for training input, nerve net The output of network is predicting the outcome for SVI.

The creativeness of the present invention is mainly reflected in：

(1). the present invention is difficult to the problem of on-line checking for current sludge volume index SVI, it is proposed that based on one kind certainly The flexible measurement method of the fuzzy Recursive Networks of tissue T-S, realizes the online of mapping relations between auxiliary variable and SVI and SVI Detection；The model have high precision, good stability, it is real-time the features such as, it is a kind of effective to prevent sludge bulking to provide Detection method, promotes the raising of sewage disposal automatization level；

(2). the present invention is improved the feedback layer of fuzzy recurrent neural network for the characteristic of sludge bulking.It is based on Rule produces criterion, is automatically generated fuzzy rule, by effective fuzzy set generating algorithm, dynamic adjustment network structure.Solution Network structure of having determined is difficult to the problem for determining, while model accuracy is ensured, effectively simplifies network structure.Using gradient Drawdown parameter learning algorithm, improves the learning ability of network.

Description of the drawings

Fig. 1. the topology diagram of the fuzzy recurrent neural networks of self-organizing T-S；

Fig. 2. inventive network fuzzy rule generates figure；

Fig. 3. inventive network trains fitting result figure；

Fig. 4. hard measurement result figure of the present invention；

Specific embodiment

Experimental data derives from the actual daily sheet of Beijing small sewage treatment plant.The neutral net that Fig. 1 gives SVI is pre- Model is surveyed, its input is respectively mixed liquor concentration of suspension MLSS, acidity-basicity ph, aeration tank water temperature T, aeration tank ammonia NH₄, model It is output as sludge volume index SVI.Wherein MLSS refers to the weight of dewatered sludge contained by unit volume biochemistry pool mixed liquor；PH reacts The soda acid degree of influent quality;T is the current sewage temperature in aeration tank；NH₄The ammonia content of aeration tank water inlet is represented, SVI is represented After aeration tank mixed liquor was precipitated Jing 30 minutes, the volume shared by corresponding 1 gram of dewatered sludge.In input quantity in addition to pH and T, other lists Position is mg/litre.Output quantity unit ml/g.Totally 150 groups of data, wherein 90 groups of data are used for training network, another 60 groups of works For test sample, dynamic change is carried out to neutral net using structural self-organizing algorithm.

The soft-sensing model of SVI is set up using the fuzzy recurrent neural networks of self-organizing T-S, real-time detection is carried out to SVI；Tool Body step is as follows：

(1) neutral net is initialized, input node is 4, and output node is 1, and number of fuzzy rules is 0, to neutral net Weights assign initial value, and the initial weight of this experiment is the random number for 0 to 1.

(2) sample data is corrected, then does normalized.

(3) self-organizing adjustment, formula (13) are carried out to the structure of neutral netFormula (15) I_th=0.92, input is The input data of correction：

1. first group of data x (1)=[4.72 23.1 43.6 250.4], normalized x (1)=[0.1092 are input into 0.8867 0.7572 0.3915].Number of fuzzy rules is changed into 1, according to formula (9) (10) to first group of fuzzy set Gaussian function center Value and width value initialization：Central value c₁(1)=[0.1092 0.8867 0.7572 0.3915], width value σ₁(1)=[0.5 0.5 0.5 0.5]；

2. second group of data x (2)=[5.63 23.4 37.2 250.1], normalized x (2)=[0.3707 are input into 0.9434 0.5527 0.33907] output of membership function layer is calculated according to formula (1) (2) Formula (11) calculates spatial activation intensityValue,Because only that a fuzzy rule,For maximum,Formula (13) is unsatisfactory for, does not increase new fuzzy rule；

3. data are sequentially input, until being input into the 23rd group of data, normalized x (23)=[0.1954 0.1886 0.5495 0.3968].The output of membership function layer is calculated according to formula (1) (2)Formula (11) Calculate spatial activation intensityValue, because only that a fuzzy rule,For maximum, formula (13) is met：Then increase a new fuzzy rule；

Calculate the output of membership function node layerFor maximum Value, meets formula (15)Then the fuzzy rule to newly increasing generates corresponding fuzzy set, according to formula (16)~(18) The parameter initialization of the fuzzy set to increasing newly：Central value initial value value c₂=x (23), r values are 0.6, width initial value σ_i2=r*|x_i (23)-c_i1|, i=1,2,3,4, σ_i2=[0.0517 0.0.4189 0.1246 0.0032]。

4. when kth group data are input into, according to (1) (2) the node output of membership function layer is calculatedComputation rule layer The spatial activation intensity of nodeValue, according to formula (12), find outMaximum；Judging the value of spatial activation intensity is It is no less than predetermined thresholdIf meeting formula (13), increase a new fuzzy rule, perform the and 5. walk；If be unsatisfactory for Formula (13), then be input into the group data of kth+1, performs the and 6. walks.

5. judge that membership function node is exported according to formula (15)Maximum whether be more than predetermined threshold I_th, meet Formula (15), then for the corresponding fuzzy set of fuzzy rule generation for newly increasing, according to formula (the 16)~parameter of (18) to fuzzy set Initialization；If being unsatisfactory for formula (15), new fuzzy set is not generated, and delete the fuzzy rule for newly increasing；

6. judge whether input data fully enters to complete, complete then to enter (4th) step, otherwise go to the and 4. walk, all Data training completes symbiosis into 4 fuzzy rules.The initial value of Gaussian function is：

The central value initial value of Gaussian function is

The width value initial value of Gaussian function is

Rules layer to the initial weight of feedback layer is

The initial value of linear dimensions is

(4) with the training sample data training neutral net after correction, train epochs s=1000, if train epochs choosing Too small, then the information content for gathering is inadequate；If train epochs choosing is excessive, it may appear that Expired Drugs.

(5) to the parameter adjustment of neutral net, train epochs s=1, input data is calculated each from first group of data input The network inputs output quantity of layer, according to formula (19) calculating target function value, using gradient descent method to 4 parameter adjustments：Parameter The linear dimensions a of layer, connection weight ω of rules layer to feedback layer, central value c and width value of membership function layer Gaussian function σ。

1. it is input into first group of data x (1)=[0.1092 0.8867 0.7572 0.3915]

The output of input layer：o⁽¹⁾(1)=x(1)=[0.1092 0.8867 0.7572 0.3915]

The output of membership function layer：

Feedback layer and rules layer are exported：o⁽³⁾(0)=[0 00 0], so h=[0 00 0]

So o⁽³⁾(1)=[0.0044 0 0 0.0281]

The output of parameter layer：o⁽⁴⁾(1)=o⁽³⁾(1)W(1)=[0.0047 0 0 0.0301]

The output of model：

②y_d(1)=0.5960, computing system error

3. using gradient descent method to parameter adjustment：According to formula (20)~(22), calculating parameter layer linear dimensions a adjustment Amount, the parameter value after being adjusted；According to formula (23)~(25), the connection weight ω adjustment amount of computation rule layer to feedback layer, Parameter value after being adjusted；According to formula (26)~(30), central value c and width value σ adjustment amounts of Gaussian function are calculated, obtained Parameter value after adjustment.

4. next group of data are input into, the output quantity that each layer of computation model adjusts each parameter value using gradient descent algorithm, Complete until all training datas are fully entered.Train epochs add 1.

(6) continue to train neutral net, repeat (5th) step, until train epochs reach 1000 steps, complete network training.

(7) test sample is predicted：Using test sample data as the neutral net for training input, nerve net Network is output as sludge volume index SVI.

Claims (1)

1. a kind of SVI flexible measurement methods, it is characterised in that comprise the following steps：

(1) data prediction and auxiliary variable are selected；

Sample set data zero-mean standardized method is normalized, and by pivot analysis PCA auxiliary variable is carried out It is selected, it is final to determine mixed liquor concentration of suspension MLSS, acidity-basicity ph, aeration tank water temperature T, aeration tank ammonia NH₄As model Input variable；

(2) the Recurrent Fuzzy Neural Network model of the hard measurement of SVI is set up, input quantity is MLSS, pH, T, NH₄, the output of model Measure as sludge volume index SVI；Recurrent Fuzzy Neural Network topological structure：Input layer is ground floor, membership function layer i.e. second Layer, rules layer are third layer, parameter layer i.e. the 4th layer, output layer i.e. layer 5, feedback layer；

Neural network structure：Input layer has 4 input nodes, and each input layer connects m membership function node layer, rule Layer has m node, and m represents number of fuzzy rules, and the number of fuzzy rules generated in network structure training process determines the value of m, ginseng Number node layer number, feedback layer nodes are equal with the nodes of rules layer, and output layer has 1 node；X=[x₁,x₂,x₃,x₄] Represent the input of neutral net, y_dRepresent the desired output of neutral net；If kth group sample data is x (k)=[x₁(k),x₂ (k),x₃(k),x₄(k)]；When kth group sample data is input into：

The output of i-th node of input layer is expressed as：

$o_i^{\left(1\right)}\left(k\right)=x_i\left(k\right),i=1,2,3,4---\left(1\right)$

Wherein,Represent output of i-th node of input layer when kth group sample is input into；

The node total number of membership function layer is：4m, the node of each input layer is all connected with the node of m membership function layer, is subordinate to Function layer is output as：

$o_{ij}^{\left(2\right)}\left(k\right)=\exp (-\frac{{\left(o_i^{\left(1\right)}\right(k)-c_{ij}(k\left)\right)}^2}{{\left({\mathrm{\σ}}_{ij}\right(k\left)\right)}^2}),j=1,2...m---\left(2\right)$

Wherein, membership function adopts Gaussian function, c_ijAnd σ_ijThe corresponding membership function layer of i-th node of input layer is represented respectively The central value and width of the Gaussian function of j node, each Gaussian function is a fuzzy set, in the structural adjustment stage pair Gaussian function assigns initial value；Represent j-th node of the corresponding membership function layer of i-th node of input layer in input kth Output during group sample；

The nodes of rules layer are m, and j-th node of the corresponding membership function layer of each node of input layer is all connected to rule J-th node of layer；Feedback link is introduced in rules layer, in feedback layer built-in variable is added；Feedback node layer includes two species The node of type：Node is accepted, summation operation is weighted to the output quantity and weights of rules layer, calculate built-in variable；Feedback section Point, using sigmoid functions as membership function, calculates the output quantity of feedback layer；Rules layer each node is all connected with all of Feedback layer accepts node；The undertaking node of feedback layer is corresponded with feedback node, and nodes are equal, feedback node and rules layer Node is corresponded, and nodes are equal, and change with the change of number of fuzzy rules, consistently equal to number of fuzzy rules；Rule J-th node of layer is output as：

$h_q=\underset{j=1}{\overset{m}{\Σ}}{o}_j^{\left(3\right)}(k-1){\mathrm{\ω}}_{jq},j=1,2,...,m;q=1,2,...,m---\left(3\right)$

$f_q=\frac{1}{1+\exp (-h_q)}---\left(4\right)$

$o_j^{\left(3\right)}\left(k\right)=f_q\underset{i=1}{\overset{4}{\Π}}{o}_{ij}^{\left(2\right)}\left(k\right)---\left(5\right)$

Wherein, ω_jqIt is the connection weight of j-th node with q-th undertaking node of feedback layer of rules layer, initial assignment is 0 Random number to 1 between；Output quantity of j-th node of rules layer in the group sample of kth -1 is represented,h_qRepresent anti- The built-in variable of q-th undertaking node of feedback layer；f_qRepresent the output quantity of q-th feedback node of feedback layer；The output of rules layerThat is the intensity of activation of fuzzy rule, whereinRepresentation space intensity of activation；

The node of rules layer is one-to-one with the node of parameter layer, and the output that parameter layer has m node, this layer is expressed as：

$W_j\left(k\right)=\underset{i=1}{\overset{4}{\Σ}}{a}_{ij}{x}_i\left(k\right)---\left(6\right)$

$o_j^{\left(4\right)}\left(k\right)=o_j^{\left(3\right)}\left(k\right)W_j\left(k\right)---\left(7\right)$

Wherein, a_ijLinear dimensions is represented, initial value is entered as 0 to 1 random number, a_j=[a_1j,a_2j,a_3j,a_4j]；W_jK () represents ginseng The value of linear dimensions weighted sum of several layers of j-th node when kth group sample is input into,Represent j-th section of parameter layer Output of the point in kth group sample；

Network model is multiple input single output, and output layer has 1 node, and all of parameter node layer is connected to output node；Net Network output is expressed as：

$y\left(k\right)=\frac{\underset{j=1}{\overset{m}{\Σ}}{o}_j^{\left(4\right)}\left(k\right)}{\underset{j=1}{\overset{m}{\Σ}}{o}_j^{\left(3\right)}\left(k\right)}---\left(8\right)$

Wherein, y (k) represents the network output of kth group sample；

(3) fuzzy neural network determines network structure by structural self-organizing adjustment first；The structural self-organizing adjustment of network：x (k)=[x₁(k),x₂(k),x₃(k),x₄(k)] represent that model is currently input into kth group sample, start from first group of sample data defeated Enter, complete to total data input, be often input into one group of sample data and judge whether to increase new fuzzy rule using following steps Then, judge whether to increase new fuzzy set on the basis of fuzzy rule is increased；

1. number of fuzzy rules is 0 in network model initial configuration, first group of data input, and number of fuzzy rules is changed into 1, and increases new Fuzzy set；The initialization of fuzzy set, the i.e. initial table of Gaussian function central value and width value is shown as：

C (1)=x (1)=[x₁(1),x₂(1),x₃(1),x₄(1)] (9)

σ (1)=[σ₁₁,σ₂₁,σ₃₁,σ₄₁]=[0.5,0.5,0.5,0.5] (10)

Wherein c (1) represents the central value of the first group of fuzzy set for generating, the i.e. central value of membership function；σ (1) represents what is generated The width value of the width value of first fuzzy set, i.e. membership function；

2. input data sequentially inputs, and is often input into one group of data, judges whether to increase new fuzzy rule, and judgment formula is represented For：

WhereinRepresent the spatial activation intensity of j-th node of rules layer；J is represented and worked asThe value of j when taking maximum；N represents current Number of fuzzy rules；For threshold value set in advance, value is 0.24；

If meeting formula (13), increase a new fuzzy rule, 3. N'=N+1 performs；

If being unsatisfactory for formula (13), repeat 2., to be input into next group of data；

3. on the basis of a new fuzzy rule is increased, judge whether to increase new fuzzy set, judgment formula is expressed as：

$I=\arg \underset{1<=j<=h}{max}\left(o_{ij}^{\left(2\right)}\right(k\left)\right)---\left(14\right)$

$o_{iI}^{\left(2\right)}\left(k\right)>I_{th}---\left(15\right)$

Wherein I is represented and worked asThe value of j when taking maximum, h represents the fuzzy set number of "current" model, h=N；I_thFor set in advance Threshold value, value is 0.92；

If meeting formula (15), increase a new fuzzy set, h'=h+1, h'=N'；Fuzzy set to increasing newly assigns initial value Newly-increased Gaussian function central value and width value to membership function layer assigns initial value；Initial table is shown as：

c_N+1=x (k) (16)

$c^{+}=\arg \underset{1<=p<=N}{min}\left|\right|x\left(k\right)-c_p\left|\right|---\left(17\right)$

${\mathrm{\&sigma;}}_{i,N+1}=r|x_i\left(k\right)-c_i^{+}|,i=1,2,3,4---\left(18\right)$

Wherein, N represents the existing number of fuzzy rules of "current" model, p=1,2 ..., N, c_N+1Represent the center of newly-increased membership function The initial value of value, r represents overlap coefficient, and value is 0.6；c_pIt is expressed as the central value of p-th Gaussian function in "current" model；c⁺ Represent and work as x (k) and c_pC when space length is minimum_pValue；σ_i,N+1Represent the width initial value of newly-increased membership function, input quantity x (k) and c⁺Difference absolute value and overlap coefficient product；Increase new fuzzy set, i.e., it is corresponding in each input layer Membership function layer increases a new node, and rules layer, feedback layer, parameter layer respectively increase accordingly corresponding node；

If being unsatisfactory for (15), do not increase new fuzzy set h'=h, and N "=N'-1, the fuzzy rule that will be newly increased Delete；

4. the structure for adjusting neutral net is continued, input data sequentially inputs, repeats 2. 3., and all input data inputs are completed Afterwards, neural network structure adjusting training is completed；

(4) structure of network model is determined, then network parameter is just adjusted, with the data after correction nerve net is trained Network, wherein totally 150 groups of sample datas, 90 groups of training sample data, 60 groups of test sample data, train epochs are 1000 steps, often In the training process of one step, 90 groups of training sample data are fully entered, and every group of training sample data input is declined using gradient Algorithm is adjusted to network parameter, wherein the parameter for adjusting includes：The linear dimensions a of parameter layer, rules layer arrives feedback layer Connection weight ω, central value c and width value σ of membership function layer Gaussian function；

The object function of definition training is that systematic error is defined as：

$E\left(k\right)=\frac{1}{2}{\left(y\right(k)-y_d(k\left)\right)}^2---\left(19\right)$

Wherein y (k) for network reality output amount, y_dK () is the desired throughput of network, E (k) represents systematic error；

Gradient descent algorithm：Partial derivative is asked using output quantity of the object function to each layer, the error propagation of per layer of output is calculated ；Partial derivative, partial derivative of the calculating target function to parameter value, partial derivative is asked to be the adjustment of each parameter by compound function Amount；

The error propagation item of parameter layerThe partial derivative that i.e. object function is exported to parameter layer：

${\mathrm{\&delta;}}_j^{\left(4\right)}\left(k\right)=-\frac{\&part;E\left(k\right)}{\&part;y\left(k\right)}\frac{\&part;y\left(k\right)}{\&part;o_j^{\left(4\right)}\left(k\right)}---\left(20\right)$

To linear dimensions a_jAdjustment amount be：

$\frac{\&part;E\left(k\right)}{\&part;a_j\left(k\right)}=\frac{\&part;E\left(k\right)}{\&part;o_j^{\left(4\right)}\left(k\right)}\frac{\&part;o_j^{\left(4\right)}\left(k\right)}{\&part;a_j\left(k\right)}=-{\mathrm{\&delta;}}_j^{\left(4\right)}\left(k\right)\frac{o_j^{\left(3\right)}\left(k\right)}{\underset{j=1}{\overset{m}{\&Sigma;}}{o}_j^{\left(3\right)}\left(k\right)}---\left(21\right)$

$a_j(k+1)=a_j\left(k\right)-\mathrm{\&eta;}\frac{\&part;E\left(k\right)}{\&part;a_j\left(k\right)}---\left(22\right)$

Rules layer jth layer error propagation itemIt is as follows：

${\mathrm{\&delta;}}_j^{\left(3\right)}\left(k\right)=-\frac{\&part;E\left(k\right)}{\&part;o_j^{\left(3\right)}\left(k\right)}={\mathrm{\&delta;}}_j^{\left(4\right)}\left(k\right)\frac{\&part;o_j^{\left(4\right)}\left(k\right)}{\&part;o_j^{\left(3\right)}\left(k\right)}---\left(23\right)$

Wherein,For parameter layer error propagation item,For parameter layer output；

Recurrence layer is to rules layer connection weight ω_jqRegulation rule it is as follows：

$\frac{\&part;E\left(k\right)}{\&part;{\mathrm{\&omega;}}_{jq}\left(k\right)}=\frac{\&part;E\left(k\right)}{\&part;o_j^{\left(3\right)}\left(k\right)}\frac{\&part;o_j^{\left(3\right)}\left(k\right)}{\&part;{\mathrm{\&omega;}}_{jq}\left(k\right)}=-{\mathrm{\&delta;}}_j^{\left(3\right)}\left(k\right)\frac{\&part;o_j^{\left(3\right)}\left(k\right)}{\&part;{\mathrm{\&omega;}}_{jq}\left(k\right)}---\left(24\right)$

${\mathrm{\&omega;}}_{jq}(k+1)={\mathrm{\&omega;}}_{jq}\left(k\right)-\mathrm{\&eta;}\frac{\&part;E}{\&part;{\mathrm{\&omega;}}_{jq}}\left(k\right)---\left(25\right)$

The error propagation item of membership function layerIt is as follows：

${\mathrm{\&delta;}}_{ij}^{\left(2\right)}\left(k\right)=-\frac{\&part;E\left(k\right)}{\&part;o_{ij}^{\left(2\right)}\left(k\right)}={\mathrm{\&delta;}}_j^{\left(3\right)}\left(k\right)\frac{\&part;o_j^{\left(3\right)}\left(k\right)}{\&part;o_{ij}^{\left(2\right)}\left(k\right)}---\left(26\right)$

Central value c of membership function layer Gaussian function_ijRegulation rule it is as follows：

$\frac{\&part;E\left(k\right)}{\&part;c_{ij}\left(k\right)}=\frac{\&part;E\left(k\right)}{\&part;o_{ij}^{\left(2\right)}\left(k\right)}\frac{\&part;o_{ij}^{\left(2\right)}\left(k\right)}{\&part;c_{ij}\left(k\right)}=-{\mathrm{\&delta;}}_{ij}^{\left(2\right)}\left(k\right)\frac{\&part;o_{ij}^{\left(2\right)}\left(k\right)}{\&part;c_{ij}\left(k\right)}---\left(27\right)$

$c_{ij}(k+1)=c_{ij}\left(k\right)-\mathrm{\&eta;}\frac{\&part;E\left(k\right)}{\&part;c_{ij}\left(k\right)}---\left(28\right)$

Wherein,For the error propagation item of j-th node of membership function layer,For j-th node of membership function layer Output；

The width cs of membership function layer Gaussian function_ijRegulation rule it is as follows：

$\frac{\&part;E\left(k\right)}{\&part;{\mathrm{\&sigma;}}_{ij}\left(k\right)}=\frac{\&part;E\left(k\right)}{\&part;o_j^{\left(3\right)}\left(k\right)}\frac{\&part;o_j^{\left(3\right)}\left(k\right)}{\&part;{\mathrm{\&sigma;}}_{ij}\left(k\right)}=-{\mathrm{\&delta;}}_j^{\left(3\right)}\left(k\right)\frac{\&part;o_j^{\left(3\right)}\left(k\right)}{\&part;{\mathrm{\&sigma;}}_{ij}\left(k\right)}---\left(29\right)$

${\mathrm{\&sigma;}}_{ij}(k+1)={\mathrm{\&sigma;}}_{ij}\left(k\right)-\mathrm{\&eta;}\frac{\&part;E\left(k\right)}{\&part;{\mathrm{\&sigma;}}_{ij}\left(k\right)}---\left(30\right)$

Wherein η is the learning rate of parameter, and value is 0.15；

(5) test sample is predicted, using test sample data as the neutral net for training input, neutral net Output is predicting the outcome for SVI.