CN102508972B - Modeling method for hydrogen energy reactor - Google Patents

Abstract

The invention discloses a modeling method for a hydrogen energy reactor. The method comprises the following steps of: first determining parameters representing the working state and performance of the hydrogen energy reactor as the input and output of a model, and acquiring the data of the hydrogen energy reactor according to the requirements of the input and the output; then establishing the dynamic neural network model of the hydrogen energy reactor on the basis of the data of the hydrogen energy reactor by utilizing a sensitivity-analysis-based dynamic neural network modeling technology; and finally judging whether to optimize the dynamic neural network model of the hydrogen energy reactor on line or not by evaluating the performance of the hydrogen energy reactor in model simulation. By the method, the dynamic neural network model of the hydrogen energy reactor can be established and automatically optimized on line, mutual influence between the specific parameters of the hydrogen energy reactor can be accurately represented, the working conditions of the hydrogen energy reactor in different performance states can be simulated, and a foundation can be laid for the control and performance optimization of the hydrogen energy reactor.

Description

A kind of modeling method of Hydrogen Energy reactor

Technical field

The present invention relates to a kind of method of automatic control and areas of information technology, particularly, the method that relates to the dynamic neural network of a kind of utilization based on sensitivity analysis and Hydrogen Energy reactor is carried out to modeling, belongs to advanced manufacturing technology field.

Technical background

Hydrogen Energy reactor is a kind of new fuel cell device that utilizes hydrogen power generation, has been applied in recent years multiple industry and the fields such as new-energy automobile, compact power, cogeneration system.Along with the commercial applications of Hydrogen Energy reactor, the requirement of its model is also strengthened gradually, set up rationally Hydrogen Energy reactor model efficiently and there is important theory and using value.

Hydrogen Energy reactor model obtains by certain modeling method, can the each variable of quantitative description Hydrogen Energy reactor between interactional abstract mathematics relation.Hydrogen Energy reactor is a kind of typical multivariate Complex Nonlinear System, relates to mass transfer, heat transfer and electrochemical reaction process, and modelling by mechanism has larger difficulty; In addition, the characteristics such as change while having due to it, strong coupling, nonlinearity, utilize empirical method modeling to be difficult to obtain gratifying effect.Therefore, how to utilize new Modeling to set up Hydrogen Energy reactor model, become one of focus in the model investigation of Hydrogen Energy reactor.

Hydrogen Energy reactor model investigation before is mainly divided into two aspects, and the one, the research of mechanism model, the 2nd, the research of empirical model.The former is generally based upon on more rational theory hypothesis basis, uses basic transmission and electrochemical reaction equation to describe Hydrogen Energy reactor characteristic; The latter is relatively simple, only comes matching input, output data to reflect the performance of Hydrogen Energy reactor by setting some parameter.Although above-mentioned modeling method can solve the modeling problem of Hydrogen Energy reactor, also has the following disadvantages: the foundation of (1) mechanism model is comparatively complicated, its calculated amount is large, consuming time many, real-time is poor, is not suitable for actual industrial rig-site utilization; (2) foundation of empirical model is subject to man's activity more, and model optimization is abundant not; (3) because Hydrogen Energy reactor behavior has time-varying characteristics, and above-mentioned two class formations and the fixing model of parameter all do not possess on-line optimization ability, therefore can not fully meet the modeling demand of Hydrogen Energy reactor.

Summary of the invention

The object of the invention is to the modeling method by a kind of Hydrogen Energy reactor is provided, build Hydrogen Energy Reactor kinetics neural network model, and make it have the ability of on-line optimization, and realize the optimization of Hydrogen Energy reactor model, solve the problem existing in Hydrogen Energy reactor modeling process.

The present invention adopts following technological means to realize:

The method of utilizing the dynamic neural network based on sensitivity analysis to carry out modeling to Hydrogen Energy reactor comprises following steps:

1.1: design Hydrogen Energy Reactor kinetics neural network model, choosing the parameter (reactor temperature, hydrogen inlet pressure, air intake pressure, hydrogen inlet flow, air intake flow, humidification tank temperature, current density) that can characterize Hydrogen Energy reactor duty is mode input; Choosing the parameter (reactor output voltage) that can embody Hydrogen Energy reactor behavior is model output.

1.2: required input, the output totally 8 item number certificates of specified structure Hydrogen Energy Reactor kinetics neural network model in acquisition step 1.1, the specific requirement that gathers every item number certificate is as follows:

1. reactor temperature, considers Hydrogen Energy reactor actual condition, take 10 ℃ as data acquisition step-length;

2. hydrogen inlet pressure, considers Hydrogen Energy reactor actual condition, take 10kPa as data acquisition step-length;

3. air intake pressure, considers Hydrogen Energy reactor actual condition, take 10kPa as data acquisition step-length;

4. hydrogen inlet flow, considers Hydrogen Energy reactor actual condition, take 5L/min as data acquisition step-length;

5. air intake flow, considers Hydrogen Energy reactor actual condition, take 5L/min as data acquisition step-length;

6. humidification tank temperature, considers Hydrogen Energy reactor actual condition, take 10 ℃ as data acquisition step-length;

7. current density, considers Hydrogen Energy reactor actual condition, utilizes variable density method to carry out data acquisition, at low current density (0～100mA/cm ²) time, with 10mA/cm ²for data acquisition step-length, at middle current density (100～500mA/cm ²) time, with 50mA/cm ²for data acquisition step-length, at high current density (> 500mA/cm ²) time, with 100mA/cm ²for data acquisition step-length;

8. reactor output voltage, in the time that above-mentioned 7 image data are analog value, gathers reactor output voltage.

1.3: the Hydrogen Energy reactor data that gather in step 1.2 are screened and processed.Screening comprises two steps successively: 1. reject the data that exceed Hydrogen Energy reactor working range; 2. reject and exceed the data of triple according to sample standard deviation.Process and comprise successively two steps: the eigenwert of 1. extracting data; 2. data are normalized.

First, whether the data that gather in determining step 1.2 meet the real work scope of Hydrogen Energy reactor, directly reject if do not meet these data and with its other data that simultaneously gather.

Secondly, utilize " three times of standard deviation methods " judge and in data, whether have abnormal data, if existence reject these data and with its other data that simultaneously gather.Be located under same state same input (or output quantity) amount x _icollect altogether L data, calculate the sample average of these data

$\stackrel{\‾}{x_i}=\frac{1}{L}\underset{l=1}{\overset{L}{\mathrm{\Σ}}}{x}_i\left(l\right),i=\mathrm{1,2},L7---\left(1\right)$

Data deviation is

l=1,2, L, L, calculates standard deviation according to Bessel formula

${\mathrm{\σ}}_i=\sqrt{\underset{l=1}{\overset{L}{\mathrm{\Σ}}}\frac{{(x_i\left(l\right)-\stackrel{\‾}{x_i})}^2}{L-1}}---\left(2\right)$

If a certain data deviation meets

|D _i(l)|≥3σ _i，l＝1，2，L，L (3)

Think that these data are abnormal data.

Again, recalculate sample average

as the eigenwert of this modeling data.

Finally, utilize " minimax method " to be normalized by class the data after screening, formula is

$x_{\mathrm{in}}=\frac{x_i-x_{i\max }}{x_{i\max }-x_{i\min }}---\left(4\right)$

Wherein, x _infor the data after normalization.

Screening is trained and verification msg as Hydrogen Energy Reactor kinetics neural network model with data after treatment.

1.4: adopt the dynamic neural network based on sensitivity analysis to build above-mentioned Hydrogen Energy reactor model, according to the described definite mode input of step (1) and output, utilize the data obtained in step 1.2,1.3 to train and checking it; This Hydrogen Energy Reactor kinetics neural network model builds based on Back Propagation Algorithm, and network structure is specifically divided into: input layer, hidden layer, output layer.Dynamic neural network model schematic diagram based on sensitivity analysis as shown in Figure 2.

Initialization Hydrogen Energy Reactor kinetics neural network model: this dynamic neural network model structure is I-J-K, follow according to mode input, output number and hidden layer neuron number experimental formula J=2I+1, known I=7, J=15, K=1, be that input layer has 7 neurons, hidden layer has 15 neurons, and output layer has 1 neuron; To connection weights, threshold values random assignment wherein; The input x of this dynamic neural network model is [x ₁, x ₂, L, x ₇], the actual y that is output as, desired output is y _d; If t moment dynamic neural network model is input as [x ₁(t), x ₂(t), L, x ₇(t)], the computing function of its each several part is:

Input layer, this layer is made up of 7 neurons

$I_i^{\left(1\right)}\left(t\right)=x_i\left(t\right),O_i^{\left(1\right)}\left(t\right)=I_i^{\left(1\right)}\left(t\right),i=\mathrm{1,2},L,7---\left(5\right)$

Wherein,

be respectively input layer i neuronic input, output;

Hidden layer, this layer is initially made up of 15 neurons

Wherein,

be respectively hidden layer j neuronic input, output; w _jirepresent that i neuron of output layer is to hidden layer j neuronic connection weights,

represent a hidden layer j neuronic activation function, b _jrepresent a hidden layer j neuronic threshold values;

Output layer, this layer only has 1 neuron

Wherein,

be respectively output layer k neuronic input, output; w _kjrepresent that j neuron of hidden layer is to output layer k neuronic connection weights,

represent an output layer k neuronic activation function, b _krepresent an output layer k neuronic threshold values.

The actual of dynamic neural network model is output as

$y\left(t\right)=O_k^{\left(3\right)}\left(t\right),k=1---\left(8\right)$

Definition Model prediction error functions is

$e=\frac{1}{M}\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{(y_m\left(t\right)-y_d\left(t\right))}^2---\left(9\right)$

Wherein, M is number of training, y _m(t) the actual output of dynamic neural network model while being the m time training.

By Levenberg-Marquardt algorithm, above-mentioned dynamic neural network model is trained, the object of training is to make the model predictive error functional value of formula (9) definition reach expectation value e _d.The training of dynamic neural network model and verification msg be by step 1.3, training data and account for 70% of whole data total amount, and all the other are 30% as verification msg; In training process, utilize sensitivity analysis to analyze dynamic neural network model hidden layer neuron, delete redundant neurons, the overweight neuron of division load, to realize the dynamic optimization of Artificial Neural Network Structures.The concrete steps of said process are:

1. utilize training data to pass through in Levenberg-Marquardt Algorithm for Training step 1.4 dynamic neural network model building, until training batch deconditioning while reaching N;

2. training data input step 1. in gained dynamic neural network model, is write down to output layer input corresponding to each hidden layer neuron

determine its maximal value minimum value

right

carry out Fourier transform,

$w_{\mathrm{kj}}{O}_j^{\left(2\right)}\left(s\right)=\frac{1}{2}(w_{\mathrm{kj}}{O_j^{\left(2\right)}}_{\max }+w_{\mathrm{kj}}{O_j^{\left(2\right)}}_{\min })+\frac{1}{\mathrm{\π}}(w_{\mathrm{kj}}{O_j^{\left(2\right)}}_{\max }-w_{\mathrm{kj}}{O_j^{\left(2\right)}}_{\min })\arcsin \left(\sin \left({\mathrm{\ω}}_{j}s\right)\right)---\left(10\right)$

Wherein, ω _jfor selected proper frequency.Make dynamic neural network model output

$O_k^{\left(3\right)}\left(t\right)=f\left(w_{\mathrm{kj}}{O}_j^{\left(2\right)}\left(t\right)\right)---\left(11\right)$

Dynamic neural network model output can be transformed to

$y\left(s\right)=O_k^{\left(3\right)}\left(s\right)=F\left(w_{\mathrm{kj}}{O}_j^{\left(2\right)}\left(s\right)\right)=\underset{{\mathrm{\ω}}_j=-\∞}{\overset{+\∞}{\mathrm{\Σ}}}(A_j\cos \left({\mathrm{\ω}}_{j}s\right)+B_j\sin \left({\mathrm{\ω}}_{j}s\right))---\left(13\right)$

Wherein, Fourier coefficient

$A_j=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}F\left(s\right)\cos \left({\mathrm{\ω}}_{j}s\right)\mathrm{ds}---\left(14\right)$

$B_j=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}F\left(s\right)\sin \left({\mathrm{\ω}}_{j}s\right)\mathrm{ds}---\left(15\right)$

3. j output layer input being independent of other output layer inputs can be expressed as the contribution of dynamic neural network model output

one order expression formula,

$S_j=\frac{{\mathrm{var}}_j\left[E\left(y\right|w_{\mathrm{kj}}{O}_j^{\left(2\right)})\right]}{\mathrm{var}\left(y\right)}---\left(16\right)$

Wherein,

for output layer input

on the impact of output variance, var (y) is the variance of output y;

According to formula (12), the variance of the output y that derives is

$\mathrm{var}\left(y\right)=\frac{1}{2\mathrm{\π}}{\∫}_{-\mathrm{\π}}^{\mathrm{\π}}{F}^2\left(s\right)\mathrm{ds}-{\left[E\left(y\right)\right]}^2\≈\underset{{\mathrm{\ω}}_j=-\∞}{\overset{+\∞}{\mathrm{\Σ}}}(A_{{\mathrm{\ω}}_j}^2+B_{{\mathrm{\ω}}_j}^2)-(A_0^2+B_0^2)\≈2\underset{{\mathrm{\ω}}_j=1}{\overset{+\∞}{\mathrm{\Σ}}}(A_{{\mathrm{\ω}}_j}^2+B_{{\mathrm{\ω}}_j}^2)---\left(17\right)$

Output layer input

can be expressed as the impact of output variance:

$\mathrm{var}\left[E\left(y\right|w_{\mathrm{kj}}{O}_j^{\left(2\right)})\right]=2\underset{K=1}{\overset{+\∞}{\mathrm{\Σ}}}(A_{K{\mathrm{\ω}}_j}^2+B_{K{\mathrm{\ω}}_j}^2)---\left(18\right)$

Wherein, K ω _jrepresent

frequency corresponding to K subharmonic.

Because the outlet chamber of the hidden layer neuron of dynamic neural network model does not interact, it is upper that Fourier's amplitude mainly concentrates on fundamental frequency (K=1), and the susceptibility of j hidden layer neuron is

${\mathrm{ST}}_j=\frac{\underset{K=1}{\overset{+\∞}{\mathrm{\Σ}}}(A_{K{\mathrm{\ω}}_j}^2+B_{K{\mathrm{\ω}}_j}^2)}{\underset{{\mathrm{\ω}}_j=1}{\overset{+\∞}{\mathrm{\Σ}}}(A_{{\mathrm{\ω}}_j}^2+B_{{\mathrm{\ω}}_j}^2)}=\frac{A_{{\mathrm{\ω}}_j}^2+B_{{\mathrm{\ω}}_j}^2}{\underset{{\mathrm{\ω}}_j=1}{\overset{+\∞}{\mathrm{\Σ}}}(A_{{\mathrm{\ω}}_j}^2+B_{{\mathrm{\ω}}_j}^2)}---\left(19\right)$

Total susceptibility sum of j hidden layer neuron is

${\mathrm{SUM}}_{\mathrm{ST}}=\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{\mathrm{ST}}_j---\left(20\right)$

The susceptibility of j hidden layer neuron of normalization is

${\mathrm{ST}}_{\mathrm{jn}}=\frac{{\mathrm{ST}}_j}{{\mathrm{SUM}}_{\mathrm{ST}}}---\left(21\right)$

4. setting reasonable susceptibility interval is [ε _l, ε _h], susceptibility is less than to ε _lhidden layer neuron delete, by maximum susceptibility and be greater than ε _hneuron be split into d neuron; After division, new neuronic connection weights, initial value of threshold are

$w_{\mathrm{jiNew}}=\frac{1}{d}{w}_{\mathrm{ji}},$ $w_{\mathrm{kjNew}}=\frac{1}{d}{w}_{\mathrm{kj}},$ $b_{\mathrm{jNew}}=\frac{1}{d}{b}_j---\left(22\right)$

If there is n _dindividual hidden layer neuron is deleted, has n _iindividual neuron is split into d new neuron, the hidden layer neuron number after optimizing

J _new＝J ₀-n _d+n _i(d-1) (23)

Wherein, J ₀for hidden layer neuron number before neural network model dynamic optimization;

5. to step 4. gained dynamic neural network model train, then every N training batch repeating step 2.-4., until the susceptibility of all hidden layer neuron is all at [ε _l, ε _h] upper, hidden layer neuron number no longer changes;

6. continue to utilize Levenberg-Marquardt algorithm to train Hydrogen Energy Reactor kinetics neural network model, expect predicated error e until error e is less than model _d, complete the structure of Hydrogen Energy Reactor kinetics neural network model.

1.5: in Hydrogen Energy reactor model emulation process, continue the input described in step 1.2, real-time data acquisition is carried out in output, modeling data before image data is no less than, the modeling data before even new data is no less than, then according to method described in step 1.3, it is screened and is processed, if there are 4 to be greater than (or being less than) 150% (or 50%) of data character pair value before in the eigenwert of new data, before and after thinking there is notable difference in modeling data, be that Hydrogen Energy reactor is under different performance state, need to utilize new data to carry out on-line optimization to Hydrogen Energy Reactor kinetics neural network model, in on-line optimization process to the concrete optimization method of dynamic neural network model with described in step 1.4, otherwise, without Hydrogen Energy Reactor kinetics neural network model is implemented to on-line optimization.

The modeling method of a kind of Hydrogen Energy reactor of the present invention, compared with prior art, has following obvious advantage and beneficial effect:

The present invention has realized the object of utilizing the dynamic neural network based on sensitivity analysis Hydrogen Energy reactor to be carried out to modeling, and design object is clear, and process is clear and definite, and reality is feasible.The invention solves in the model investigation of Hydrogen Energy reactor, mechanism model complex structure, poor practicability, the problems such as empirical model is too coarse, optimization deficiency, have realized the innovation to the model investigation of Hydrogen Energy reactor.

Accompanying drawing explanation

Fig. 1 is Hydrogen Energy Reactor kinetics neural net model establishing and the model on-line optimization system schematic diagram in the present invention;

Fig. 2 is the dynamic neural network model schematic diagram based on sensitivity analysis in the present invention;

Fig. 3 is data acquisition result (part) schematic diagram in the present invention;

Fig. 4 is data screening and result (part) schematic diagram in the present invention;

The change procedure schematic diagram of hidden layer neuron number when Fig. 5 is the dynamic neural network model training in the present invention;

The change procedure schematic diagram of hidden layer neuron number when Fig. 6 is the dynamic neural network model on-line optimization in the present invention.

Embodiment

In order to understand better technical scheme of the present invention, below embodiments of the present invention are described further.

According to technical scheme provided by the invention, specifically implement the present invention for certain Hydrogen Energy reactor test platform, step is as follows:

1.1: take the Hydrogen Energy reactor in certain Hydrogen Energy reactor test platform as dynamic neural network modeling object, take reactor temperature, hydrogen inlet pressure, air intake pressure, hydrogen inlet flow, air intake flow, humidification tank temperature, current density as mode input, the output take reactor output voltage as model.

1.2: required input, the output totally 8 item number certificates of specified structure Hydrogen Energy Reactor kinetics neural network model in acquisition step 1.1, consider Hydrogen Energy reactor working range and characteristic, the specific requirement that gathers every item number certificate is as follows:

1. reactor temperature (20～80 ℃), considers that reactor actual condition, take 10 ℃ as data acquisition step-length, gathers the data (20 ℃ do not gather) of 6 kinds of operating modes altogether;

2. hydrogen inlet pressure (gauge pressure 0～50kPa), considers reactor actual condition, take 10kPa as data acquisition step-length, gathers altogether the data (0 and 50kPa do not gather) of 4 kinds of operating modes;

3. air intake pressure (gauge pressure 0～30kPa), considers reactor actual condition, take 10kPa as data acquisition step-length, gathers altogether the data (0 and 30kPa do not gather) of 2 kinds of operating modes;

4. hydrogen gas flow (10～20L/min) only gathers the data of 20L/min operating mode in the present embodiment;

5. air inlet flow (10～20L/min) only gathers the data of 20L/min operating mode in the present embodiment;

6. humidification tank temperature (20～40 ℃) only gathers the data of 40 ℃ of operating modes in the present embodiment;

7. current density (0～800mA/cm ²), consider that reactor actual working state utilizes variable density method to carry out data acquisition, at low current density (0～100mA/cm ²) under, with 10mA/cm ²for data acquisition step-length, at middle current density (100～500mA/cm ²) under, with 50mA/cm ²for data acquisition step-length, at high current density (500～800mA/cm ²) under, with 100mA/cm ²for data acquisition step-length, gather altogether the data of 22 kinds of operating modes;

8. reactor output voltage, in the time that above-mentioned 7 class collection capacities are analog value, gathers reactor output voltage (unit, V).

According to above-mentioned specific requirement, gather altogether 1000 groups of Hydrogen Energy reactor datamation data, acquired results (part) is as shown in Figure 3.

1.3: according to method definite in step 3, step 1.2 the data obtained is screened and processed, acquired results (part) as shown in Figure 4.In visible step 1.2 the data obtained, exist and exceed Hydrogen Energy reactor working range and exceed the data of triple according to sample standard deviation, but all disallowable in screening link, the data after screening have 939 groups.

1.4: build Hydrogen Energy Reactor kinetics neural network model according to the described input of step (1), output, and utilize step 1.3 the data obtained to carry out training and the checking of dynamic neural network model, wherein reasonable susceptibility interval [ε _l, ε _h] being made as [5%, 50%], N is made as 200, d and is made as 4.In training process, as shown in Figure 5,, in dynamic neural network model training process, hidden layer neuron is by 15 of initial setting for the number of variations of hidden layer neuron, finally stabilize to 4 by deleting with division, reached the object of neural network model dynamic optimization.

1.5: will exist in the new data input step 1.4 gained dynamic neural network model of significant difference with modeling data before, Hydrogen Energy Reactor kinetics neural network model starts on-line optimization process.As shown in Figure 6, hidden layer neuron finally stabilizes to 5 by deleting with division, has realized the on-line optimization of Hydrogen Energy Reactor kinetics neural network model.

Claims (7)

1. a modeling method for Hydrogen Energy reactor, is characterized in that, comprises following steps:

1.1: design Hydrogen Energy Reactor kinetics neural network model, choosing the parameter that can characterize Hydrogen Energy reactor duty is mode input, i.e. reactor temperature, hydrogen inlet pressure, air intake pressure, hydrogen inlet flow, air intake flow, humidification tank temperature, current density; Choosing the parameter that can embody Hydrogen Energy reactor behavior is model output, i.e. reactor output voltage;

1.2: for the data acquisition link in Hydrogen Energy Reactor kinetics neural net model establishing, utilize definite mode input in variable density sampling method acquisition step 1.1, output data;

1.3: for data screening and processing links in Hydrogen Energy Reactor kinetics neural net model establishing, first, whether the data that determining step 1.2 gathers meet the real work scope of Hydrogen Energy reactor, directly reject if do not meet these data and with its other data that simultaneously gather; Secondly, judge in data, whether there is abnormal data, even data deviation is greater than 3 haplotype data sample standard deviations and thinks that these data are abnormal data, should reject these data and with its other data that simultaneously gather; Again, the average of every input data after calculating rejecting abnormalities data, as the eigenwert of this modeling data; Finally, Hydrogen Energy reactor data are normalized by item;

1.4: adopt the dynamic neural network based on sensitivity analysis to build Hydrogen Energy reactor model, according to step 1.1 definite dynamic neural network model input and output parameter, utilize the data obtained in step 1.3 to carry out training and the checking of dynamic neural network model; This Hydrogen Energy Reactor kinetics neural network model builds based on Back Propagation Algorithm, after training certain batch, calculate the susceptibility of each hidden layer neuron, then judge that it is whether on predefined reasonable susceptibility interval, if lower than interval lower limit or higher than the interval upper limit corresponding selection delete or divide corresponding hidden layer neuron, continue neural network training until reach estimated performance requirement, so far complete the foundation of Hydrogen Energy Reactor kinetics neural network model;

1.5: the input of Hydrogen Energy reactor, output are carried out to real-time data acquisition, the modeling data before the new data of collection is no less than; Then according to described in step 1.3, it being screened and is processed, if there are the several data character pair values before that are greater than or less than in the eigenwert of new data, before and after thinking there is notable difference in modeling data, be that Hydrogen Energy reactor is under different performance state, need to utilize new data to carry out on-line optimization to Hydrogen Energy Reactor kinetics neural network model, the specific implementation method of on-line optimization is with step 1.4; Otherwise, Hydrogen Energy Reactor kinetics neural network model is not implemented to on-line optimization.

2. the modeling method of a kind of Hydrogen Energy reactor according to claim 1, is characterized in that, described reasonable susceptibility interval is 5%～50%, 5% to be interval lower limit, and 50% is the interval upper limit.

3. the modeling method of a kind of Hydrogen Energy reactor according to claim 1, is characterized in that, 120% of the modeling data before the new data of described collection is no less than.

4. the modeling method of a kind of Hydrogen Energy reactor according to claim 1, is characterized in that, several in described step 1.5 in the eigenwert of new data are 4, and before being greater than or less than, data character pair value is: be greater than 150% or be less than 50%.

5. the modeling method of a kind of Hydrogen Energy reactor according to claim 1, is characterized in that: in step 1.2, adopt variable density sampling method to carry out data acquisition to the Hydrogen Energy reactor mode input of specifying in step 1.1, output; The data screening stage in step 1.3, whether abnormal by considering the Hydrogen Energy reactor real work scope data that directly judgement gathers; In step 1.3 by calculating the eigenwert specified data feature of Hydrogen Energy reactor modeling data, as the basis for estimation of whether Hydrogen Energy Reactor kinetics neural network model being carried out on-line optimization in model emulation process.

6. the modeling method of a kind of Hydrogen Energy reactor according to claim 1, it is characterized in that: the foundation of the Reactor kinetics of Hydrogen Energy described in described step 1.4 neural network model, by being carried out to sensitivity analysis, hidden layer neuron judges the rationality of its existence, delete or divide irrational hidden layer neuron to realize neural network structure dynamic optimization, then continuing training Hydrogen Energy Reactor kinetics neural network model until network structure and parameter all meet the requirement of estimated performance index.

7. the modeling method of a kind of Hydrogen Energy reactor according to claim 1, it is characterized in that: described step 1.5 is on the basis of step 1.4, by analyzing new data eigenwert and the difference of modeling data eigenwert before, judge that Hydrogen Energy reactor is whether under different performance state, and then select whether to utilize new data to carry out on-line optimization to the Hydrogen Energy Reactor kinetics neural network model of having set up.

Images