CN104504292A - Method for predicting optimum working temperature of circulating fluidized bed boiler based on BP neural network - Google Patents

Abstract

The invention discloses a method for predicting an optimum working temperature of a circulating fluidized bed boiler based on a BP neural network. The method comprises the steps of selecting correlated quantity to serve as input of a BP neural network model according to the actual running conditions of the thermal power plant circulating fluidized bed boiler, and enabling the optimum working temperature of the circulating fluidized bed boiler to serve as output of the BP neural network; recording and storing on-site historical data and performing filtering processing, selecting the data to serve as a training set sample, determining the number of input layer nodes of the BP neural network, the number of hidden layer nodes, weights and threshold parameters; obtaining the predicted optimum working temperature of the thermal power plant fluidized bed boiler through a method of the BP neural network in combination with analytical calculation of input parameters; performing simulation testing, and performing comparative analysis of predicted results and site actual results. The problems that a fluidized bed boiler operator has judging deviation of the optimum working temperature of the boiler, the fluctuation of the optimum working temperature of a fluidized bed is high, and the stability is poor are solved.

Description

Based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature

Technical field

The present invention relates to a kind of method based on BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature.

Background technology

Technology of Circulating Fluidized Bed Boiler is the high-efficiency low-pollution clean-burning technology developed rapidly in recent ten years.This technology has obtained business application widely in fields such as station boiler, Industrial Boiler and offal treatment utilizations in the world, and develops to the large circulating fluidized bed boiler of hundreds of thousands multikilowatt scale; Domestic research in this respect, development and application are also risen gradually, among existing up to a hundred Circulating Fluidized Bed Boiler put into operation or manufacturing.The period that the coming years will be recirculating fluidized bed develop rapidly.

Circulating fluidized bed boiler systems is made up of fluidized bed combustion chamber (burner hearth), circulating ash separation vessel, fly ash loopback device, back-end surfaces and utility appliance etc. usually.Circulating fluidized bed boiler systems is made up of combustion system and boiler circuit usually, and fuel completes combustion process in the combustion system of boiler, and the fuel of recirculating fluidized bed and desulfurizing agent, through repeatedly circulating, carrying out desulphurization reaction repeatedly, have low NO _xdischarge capacity, desulfuration efficiency is high, and has that fuel tolerance is wide, load adjustment ability good, lime-ash is easy to the advantages such as comprehensive utilization, uses comparatively extensive at home and in the world, promotes rapider.

Bed temperature be one directly affect boiler can the important controling parameters that runs of safe and continuous, while also directly affect desulfuration efficiency in boiler operatiopn and NO _xgeneration.Operator is to the setting of bed working temperature with very strong randomness, and the factor of consideration is very few, often causes the bed temperature of fluidized-bed combustion boiler fluctuated, and then fuel cannot Thorough combustion, and desulfuration efficiency is lower, causes unnecessary environmental pollution.

The Combustion System of Circulating Fluidized Bed Boiler is comparatively complicated, be one have strong jamming, non-linear, time become, the process that is associated of multivariate, the on-the-spot variablees such as fuel quantity, lime stone amount, primary air flow, secondary air flow all have impact to combustion process, control accuracy is low, difficult point to the prediction of the optimum working temperature of boiler, in the production run of reality, determine that the optimum working temperature of fluidized-bed combustion boiler improves burning efficiency, improve the key of desulfuration efficiency, the setting of the working temperature of current adjustment boiler, mainly through the experience of operating personnel, has the following disadvantages:

1, the getting sth into one's head property of operator is too strong;

2, the operation of operator has obvious hysteresis quality;

3, comparatively large, the poor stability of vulcanization bed operating. temperature fluctuations;

4, burning of coal rate does not reach desirable mxm., wastes energy.

Summary of the invention

The present invention is in order to solve the problem, propose a kind of method based on BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, this method utilizes BP neural net method to predict the optimum working temperature of cogeneration plant's fluidized-bed combustion boiler, and then this key parameter of optimum working temperature of fluidized-bed combustion boiler can be provided to operating personnel, improve burning efficiency, improve desulfuration efficiency, reduce the discharge capacity of sulfide, reach the object of energy-saving and emission-reduction.

To achieve these goals, the present invention adopts following technical scheme:

Based on a method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, comprise the following steps:

(1) according to the practical operation situation of thermal power plant circulating fluidized bed boiler, fuel quantity x is chosen ₁, lime stone amount x ₂, a component of degree n n x ₃, secondary air flow x ₄as the input of BP neural network model, using the output of the optimum working temperature of Circulating Fluidized Bed Boiler as BP neural network;

(2) record and store on-the-spot historical data and do filtering process, choosing these data as training set sample, determine the input layer number of BP neural network, hidden layer node number, weights and threshold parameter;

(3) carry out analytical calculation by the methods combining input parameter of BP neural network, obtain the optimum working temperature of the cogeneration plant's fluidized-bed combustion boiler doped;

(4) carry out emulation testing, the result by prediction is compared with on-the-spot actual result.

In described step (1), concrete grammar is: according to the practical operation situation of thermal power plant circulating fluidized bed boiler, analyze relevant input and output amount, pass through emulation experiment, filter out the variable can analyzed recirculating fluidized bed optimum working temperature, as the input of BP neural network model, using the output of the optimum working temperature of Circulating Fluidized Bed Boiler as BP neural network, finally choose fuel quantity x ₁, lime stone amount x ₂, a component of degree n n x ₃, secondary air flow x ₄as input, fluidized bed optimum working temperature y ₁as output.

In described step (2), concrete grammar comprises:

A () is recorded and is stored on-the-spot historical data and do filtering process, choose these data as training set sample, training set sample has 4 groups, is respectively fuel quantity training set, lime stone amount training set, primary air flow training set, secondary air flow training set;

B the node number of the input layer of () definition BP neural network is n, obtain n=4 by analysis above, the node number of hidden layer is q, and the weights of input layer and hidden layer are ν _ki(k=1,2 ..., q; I=1,2 ..., n), threshold value is θ _i(i=1,2 ..., n), the node number of output layer is m, known m=1, and the weights of hidden layer and output layer are ω _jk(j=1,2 ..., m; K=1,2 ..., q), threshold value is f ₁the transport function that () is hidden layer, f ₂the transport function that () is output layer.

In described step (a), 4 groups of sample sizes are all defined as 165.

In described step (3), concrete analysis calculation method is: calculate the output of hidden layer node and the output of output layer node, the expectation value of definition neural network exports, and the output error of calculating is deployed in hidden layer and input layer, the method adjustment weights be directly proportional according to the weights and negative gradient that make BP neural network, training network.

In described step (3), the output of hidden layer node is:

$z_k=f_1(\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{v}_{\mathrm{ki}}{x}_i-{\mathrm{\θ}}_k)---\left(1\right)$

K=1 in above formula, 2 ..., q;

The output of output layer node is:

J=1 in above formula, 2 ..., m;

If the expectation value of neural network exports as d=(d ₁..., d _j..., d _m), when the output of network output and expectation value is inconsistent, can there is output error, it is as follows that we define output error E:

$E=\frac{1}{2}{(d-y)}^2=\frac{1}{2}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{(d_j-y_j)}^2---\left(3\right)$

Formula (3) is deployed into hidden layer, has:

Formula (4) is deployed into input layer, has:

In described step (3), obtain according to above formula (5), E is ν _ki, ω _jkfunction, if change ν _ki, ω _jkvalue, so the value of error E also can change; The optimal result of BP neural network is exactly make E become as far as possible little of to meet our requirement, and the weights of BP neural network and negative gradient are directly proportional, and in the process of optimum working temperature determining fluidized-bed combustion boiler, is exactly according to such method adjustment weights, that is:

${\mathrm{\Δv}}_{\mathrm{ki}}=-\mathrm{\η}\frac{\∂E}{{\∂v}_{\mathrm{ki}}}---\left(6\right)$

K=1 in above formula, 2 ..., q, i=1,2 ..., n;

${\mathrm{\Δ\ω}}_{\mathrm{jk}}=-\mathrm{\η}\frac{\∂E}{{\∂\mathrm{\ω}}_{\mathrm{jk}}}---\left(7\right)$

J=1 in above formula, 2 ..., m, k=1,2 ..., q;

Negative sign in formula (6) and (7) represents Gradient Descent, and η ∈ (0,1) represents scale-up factor, it reflects the speed of study.

In described step (3), the concrete grammar of training network is: hidden layer node number is set to 1, training network, then progressively increases node number, with identical sample training, and nodes corresponding when determining that error is minimum; When using the method, use some experimental formulas; The nodes that experimental formula draws is "ball-park" estimate value, as the initial value of method of trial and error.

First try initial value, if not all right, nodes adds 1, then tries again, if also not all right, then adds; The larger convergence of numerical value is faster; Until cannot just obtain suitable nodes during convergence speedup speed, conventional experimental formula is as follows:

$q=\sqrt{n+l}+\mathrm{\α}---\left(8\right)$

q＝log ₂n (9)

$q=\sqrt{\mathrm{nl}}---\left(10\right)$

$q=\sqrt{0.43\mathrm{qn}+{0.12q}^2+2.54n+0.77q+0.35}+0.51---\left(11\right)$

In above-mentioned experimental formula, n is input layer vector dimension, and l is output layer vector dimension, and α is the constant between 1 and 10, the m value round number obtained after having calculated.

In described step (3), following two pacing itemss when determining the node number of hidden layer, must be met:

(1) number of training must connect flexible strategy more than model, is 2-10 times; Otherwise, sample needs to be divided into several part, takes the way of " trained in turn ", obtains network model of good performance with this;

(2) the node number of hidden layer is less than N-1 (for N number of training); Otherwise, the error of model and sample properties have nothing to do, and go to zero, the model that is set up is without generalization ability; Identical reason, the number of input layer is also less than N-1.

Beneficial effect of the present invention is:

(1) can find out that the fluidized-bed combustion boiler optimum working temperature and the on-the-spot optimum working temperature error obtained that are obtained by BP neural network prediction are within ± 3% from simulation result, and variation tendency is consistent, illustrate that the present invention well establishes the model of fluidized-bed combustion boiler by BP neural net method, and best working temperature can be doped;

(2) can provide fluidized-bed combustion boiler optimum working temperature this key parameter to operating personnel, improve burning efficiency, improve desulfuration efficiency, reduce the discharge capacity of sulfide, reach the object of energy-saving and emission-reduction;

(3) judgment bias of fluidized-bed combustion boiler operator to the optimum working temperature of boiler is solved, and the problem such as the fluctuation of the optimum working temperature of fluidized bed is large, poor stability.

Accompanying drawing explanation

Fig. 1 is the structural representation of BP neural network of the present invention.

Fig. 2 is program flow diagram of the present invention;

Fuel quantity figure after mean filter process that Fig. 3 (a) chooses for the present invention;

Lime stone amount figure after mean filter process that Fig. 3 (b) chooses for the present invention;

Primary air flow figure after mean filter process that Fig. 3 (c) chooses for the present invention;

Secondary air flow figure after mean filter process that Fig. 3 (d) chooses for the present invention;

Fig. 4 is effect comparison chart of the present invention.

Embodiment:

Below in conjunction with accompanying drawing and embodiment, the invention will be further described.

As shown in Figure 2, utilize BP neural net method to predict the optimum working temperature of cogeneration plant's Fluidized Bed Boiler, specifically comprise the following steps:

Step 1, according to the practical operation situation of thermal power plant circulating fluidized bed boiler, analyze relevant input and output amount, pass through emulation experiment, filter out the variable can analyzed recirculating fluidized bed optimum working temperature, as the input of BP neural network model, using the output of the optimum working temperature of Circulating Fluidized Bed Boiler as BP neural network.Finally choose fuel quantity x ₁, lime stone amount x ₂, a component of degree n n x ₃, secondary air flow x ₄as input, fluidized bed optimum working temperature y ₁as output.

Step 2, determine the BP neural network parameters such as training set sample, input layer number, hidden layer node number, weights, threshold value, and carry out analytical calculation by the method for BP neural network, finally obtain the optimum working temperature of the cogeneration plant's fluidized-bed combustion boiler doped.

Step 3, carries out emulation testing, draws the effect of this patent of invention.

Determine the BP neural network parameters such as training set sample, hidden layer node number, weights in described step 2, and carry out analytical calculation by the method for BP neural network, its concrete steps are:

A () also stores on-the-spot historical data by software records and does filtering process, choose these data as training set sample, training set sample has 4 groups, be respectively fuel quantity training set, lime stone amount training set, primary air flow training set, secondary air flow training set, simulation analysis result shows, 4 groups of sample sizes are all defined as 165.

B the node number of the input layer of () definition BP neural network is n, obtain n=4 by analysis above, the node number of hidden layer is q, and the weights of input layer and hidden layer are ν _ki(k=1,2 ..., q; I=1,2 ..., n), threshold value is θ _i(i=1,2 ..., n), the node number of output layer is m, known m=1, and the weights of hidden layer and output layer are ω _jk(j=1,2 ..., m; K=1,2 ..., q), threshold value is f ₁the transport function that () is hidden layer, f ₂the transport function that () is output layer, as shown in Figure 1.

Here is concrete analytical calculation process:

1) output of hidden layer node is:

$z_k=f_1(\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{v}_{\mathrm{ki}}{x}_i-{\mathrm{\θ}}_k)---\left(1\right)$

K=1 in above formula, 2 ..., q.

The output of output layer node is:

J=1 in above formula, 2 ..., m.

If the expectation value of neural network exports as d=(d ₁..., d _j..., d _m), when the output of network output and expectation value is inconsistent, can there is output error, it is as follows that we define output error E:

$E=\frac{1}{2}{(d-y)}^2=\frac{1}{2}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{(d_j-y_j)}^2---\left(3\right)$

Formula (3) is deployed into hidden layer, has:

Formula (4) is deployed into input layer, has:

2) according to above formula (5), we can obtain, and E is ν _ki, ω _jkfunction, if change ν _ki, ω _jkvalue, so the value of error E also can change.The optimal result of BP neural network is exactly make E become as far as possible little of to meet our requirement, and the weights of BP neural network and negative gradient are directly proportional, and in the process of optimum working temperature determining fluidized-bed combustion boiler, is exactly according to such method adjustment weights, that is:

${\mathrm{\Δv}}_{\mathrm{ki}}=-\mathrm{\η}\frac{\∂E}{{\∂v}_{\mathrm{ki}}}---\left(6\right)$

K=1 in above formula, 2 ..., q, i=1,2 ..., n.

${\mathrm{\Δ\ω}}_{\mathrm{jk}}=-\mathrm{\η}\frac{\∂E}{{\∂\mathrm{\ω}}_{\mathrm{jk}}}---\left(7\right)$

J=1 in above formula, 2 ..., m, k=1,2 ..., q.

Negative sign in formula (6) and (7) represents Gradient Descent, and η ∈ (0,1) represents scale-up factor, it reflects the speed of study.

3) hidden layer node number is set to very little, training network, sees result, then progressively increases node number, with identical sample training, and nodes corresponding when determining that error is minimum.When using the method, we can use some experimental formulas.The nodes that experimental formula draws is "ball-park" estimate value, and this just can as the initial value of method of trial and error.First try initial value, if not all right, nodes adds 1, then tries again, if also not all right, then adds.The larger convergence of numerical value is faster.Until cannot just obtain suitable nodes during convergence speedup speed, conventional experimental formula is as follows:

$q=\sqrt{n+l}+\mathrm{\α}---\left(8\right)$

q＝log ₂n (9)

$q=\sqrt{\mathrm{nl}}---\left(10\right)$

$q=\sqrt{0.43\mathrm{qn}+{0.12q}^2+2.54n+0.77q+0.35}+0.51---\left(11\right)$

In above-mentioned experimental formula, n is input layer vector dimension, and l is output layer vector dimension, and α is the constant between 1 and 10, the m value round number obtained after having calculated.

According to analysis above, following two pacing itemss when can obtain the node number determining hidden layer, must be met.

(1) number of training must connect flexible strategy more than model, and generalized case is 2-10 times.Otherwise, sample needs to be divided into several part, takes the way of " trained in turn ", obtains network model of good performance with this.

(2) the node number of hidden layer is less than N-1 (for N number of training).Otherwise, the error of model and sample properties have nothing to do, and go to zero, the model that is set up is without generalization ability.Identical reason, the number of input layer is also less than N-1.

So far, the parameter completing BP neural network is determined and the process of analytical calculation.

Step 1, according to the practical operation situation of thermal power plant circulating fluidized bed boiler, analyzes relevant input and output amount, pass through emulation experiment, filter out the variable can analyzed recirculating fluidized bed optimum working temperature, as the input of BP neural network model, finally choose fuel quantity x ₁, lime stone amount x ₂, a component of degree n n x ₃, secondary air flow x ₄as input, data amount check is all respectively 165, the historical data curve of collection in worksite after mean filter process as shown in Fig. 3 (a)-Fig. 3 (d).

Step 2, determine the BP neural network parameters such as training set sample, input layer number, hidden layer node number, weights, threshold value, and carry out analytical calculation by the method for BP neural network, finally obtain the optimum working temperature of the cogeneration plant's fluidized-bed combustion boiler doped.Show that the scope of hidden layer node number is 4 to 14 according to formula (8), show that hidden layer node number is 6 according to formula (9), show that hidden layer node number is 6 according to formula (10), show that hidden layer node number is 9 according to formula (11).Therefore following chapters and sections we when studying, first select hidden layer node number to be 6 tests, then increase gradually and then reach optimum performance.

According to above chapters and sections, we learn that training set number of samples is 165, test set number of samples is 50, select tanh S type function as hidden layer transport function, rule of thumb initially connect weights generally interval (-1,1) between, here the random number between our selection (-0.3,0.3) is as initially connecting weights, and we limit maximum network training number of times is 10000.

In order to verify simulated effect better, we define following performance index.

(1) value defining error mean square root RMSE is:

$\mathrm{RMSE}=\sqrt{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\frac{{(y_i-{\hat{y}}_i)}^2}{N}}---\left(12\right)$

(2) relative error average meanA is defined _ecomputing formula be:

${\mathrm{meanA}}_E=\frac{1}{N}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}|y_i-{\hat{y}}_i|---\left(13\right)$

Above-mentioned formula (12) and (13) middle N are sample data number, y _ifor the actual best design temperature of fluidized-bed combustion boiler, for the best design temperature of the prediction using the analysis of BP Simulation of Neural Network to obtain.

When hidden layer node number gets different value, the property indices of BP neural network is as shown in table 1.

Table 1 node in hidden layer is on the impact of model performance

As can be seen from Table 1, along with the increase of the hidden layer node number of BP neural network, the training error root mean square of model reduces gradually, and training relative error average also reduces gradually, and the training time presents the trend increased gradually.But when the node number of hidden layer is 9, the test error root mean square of this model reaches minimum with test relative error average, and now training error root mean square is 0.8575, and training relative error average is 0.0013, test error root mean square is 1.9624, and test relative error average is 0.0028.But from now on, along with the continuation of hidden layer node number increases, training error root mean square reduces thereupon, training relative error average is also along with minimizing, but test error root mean square is but along with increasing, and test relative error average is also along with increasing.This illustrates that generalization ability starts to decline, and namely occurs " over-fitting " phenomenon.

Can obtain single hidden layer according to performance index just can meet the demands, therefore, the node in hidden layer based on BP neural network is defined as 9, and network structure is:

(1) hidden layer weights are input to:

$v=\left[\begin{array}{cccccccc}-6.0404& 20.5921& -12.8676& 5.6208& 1.6882& 20.9227& 21.2285& -15.4897\\ 11.2562& 10.1158& -4.3327& -9.2332& 17.5343& 3.0811& -9.2746& 5.7843\\ -12.7889& -13.3140& -23.8600& 11.0918& -22.2564& 6.6677& 14.0212& 3.1021\\ 0.0587& 1.5509& -10.1367& -16.4317& -13.0337& -26.1119& -16.1807& 1.2229\\ 6.9535& -16.3618& 12.0347& 12.6808& 5.6723& 24.6739& 21.8484& -13.0471\\ -0.6129& -16.1455& 5.6687& -4.7238& 33.4284& -12.6362& -21.8464& 10.7048\\ -11.6364& -38.4894& -23.4389& -14.9062& 38.8148& -0.2016& 21.2143& 12.7493\\ -21.6638& -7.0040& -6.7117& 4.5349& -18.4027& -14.4674& -1.9573& -6.4926\\ 28.4760& -4.1843& -8.3949& 5.3214& -9.6978& 10.8687& -7.6755& -2.9484\end{array}\right]$

(2) hidden layer threshold value:

θ＝[7.4546 11.6187 -14.4297 -8.1823 -5.4403 -6.0824 -31.6805 0.2332 9.3391] ^T

(3) hidden layer is to output layer weights:

ω＝[0.6122 -0.5591 -0.6537 1.3151 0.8652 0.6033 -0.6286 -0.7779 -0.5384]

(4) output layer threshold value:

Simulation result as shown in Figure 4, as can be seen from Figure 4 the fluidized-bed combustion boiler optimum working temperature obtained by BP neural network prediction and the on-the-spot optimum working temperature error obtained are within ± 3%, and variation tendency is consistent, illustrate that the present invention well establishes the model of fluidized-bed combustion boiler by BP neural net method, and best working temperature can be doped, and then can provide fluidized-bed combustion boiler optimum working temperature this key parameter to operating personnel, improve burning efficiency, improve desulfuration efficiency, reduce the discharge capacity of sulfide, reach the object of energy-saving and emission-reduction.

By reference to the accompanying drawings the specific embodiment of the present invention is described although above-mentioned; but not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various amendment or distortion that creative work can make still within protection scope of the present invention.

Claims (9)

1., based on a method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: comprise the following steps:

(1) according to the practical operation situation of thermal power plant circulating fluidized bed boiler, fuel quantity x is chosen ₁, lime stone amount x ₂, a component of degree n n x ₃, secondary air flow x ₄as the input of BP neural network model, using the output of the optimum working temperature of Circulating Fluidized Bed Boiler as BP neural network;

(2) record and store on-the-spot historical data and do filtering process, choosing these data as training set sample, determine the input layer number of BP neural network, hidden layer node number, weights and threshold parameter;

(3) carry out analytical calculation by the methods combining input parameter of BP neural network, obtain the optimum working temperature of the cogeneration plant's fluidized-bed combustion boiler doped;

(4) carry out emulation testing, the result by prediction is compared with on-the-spot actual result.

2. as claimed in claim 1 based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: in described step (1), concrete grammar is: according to the practical operation situation of thermal power plant circulating fluidized bed boiler, analyze relevant input and output amount, pass through emulation experiment, filter out the variable can analyzed recirculating fluidized bed optimum working temperature, as the input of BP neural network model, using the output of the optimum working temperature of Circulating Fluidized Bed Boiler as BP neural network, finally choose fuel quantity x ₁, lime stone amount x ₂, a component of degree n n x ₃, secondary air flow x ₄as input, fluidized bed optimum working temperature y ₁as output.

3., as claimed in claim 1 based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: in described step (2), concrete grammar comprises:

A () is recorded and is stored on-the-spot historical data and do filtering process, choose these data as training set sample, training set sample has 4 groups, is respectively fuel quantity training set, lime stone amount training set, primary air flow training set, secondary air flow training set;

B the node number of the input layer of () definition BP neural network is n, obtain n=4 by analysis above, the node number of hidden layer is q, and the weights of input layer and hidden layer are ν _ki(k=1,2 ..., q; I=1,2 ..., n), threshold value is θ _i(i=1,2 ..., n), the node number of output layer is m, known m=1, and the weights of hidden layer and output layer are ω _jk(j=1,2 ..., m; K=1,2 ..., q), threshold value is for the transport function of hidden layer, f ₂the transport function that () is output layer.

4., as claimed in claim 3 based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: in described step (a), 4 groups of sample sizes are all defined as 165.

5. as claimed in claim 1 based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: in described step (3), concrete analysis calculation method is: calculate the output of hidden layer node and the output of output layer node, the expectation value of definition neural network exports, and the output error of calculating is deployed in hidden layer and input layer, the method adjustment weights be directly proportional according to the weights and negative gradient that make BP neural network, training network.

6., as claimed in claim 5 based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: in described step (3), the output of hidden layer node is:

$z_k=f_1(\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{v}_{\mathrm{ki}}{x}_i-{\mathrm{\θ}}_k)---\left(1\right)$

K=1 in above formula, 2 ..., q;

The output of output layer node is:

J=1 in above formula, 2 ..., m;

If the expectation value of neural network exports as d=(d ₁..., d _j, d _m), when the output of network output and expectation value is inconsistent, can there is output error, it is as follows that we define output error E:

$E=\frac{1}{2}{(d-y)}^2=\frac{1}{2}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{(d_j-y_j)}^2---\left(3\right)$

Formula (3) is deployed into hidden layer, has:

Formula (4) is deployed into input layer, has:

7., as claimed in claim 6 based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: in described step (3), obtain according to above formula (5), E is ν _ki, ω _jkfunction, if change ν _ki, ω _jkvalue, so the value of error E also can change; The optimal result of BP neural network is exactly make E become as far as possible little of to meet our requirement, and the weights of BP neural network and negative gradient are directly proportional, and in the process of optimum working temperature determining fluidized-bed combustion boiler, is exactly according to such method adjustment weights, that is:

${\mathrm{\Δv}}_{\mathrm{ki}}=-\mathrm{\η}\frac{\∂E}{\∂v_{\mathrm{ki}}}---\left(6\right)$

K=1 in above formula, 2 ..., q, i=1,2 ..., n;

${\mathrm{\Δ\ω}}_{\mathrm{jk}}=-\mathrm{\η}\frac{\∂E}{{\∂\mathrm{\ω}}_{\mathrm{jk}}}---\left(7\right)$

J=1 in above formula, 2 ..., m, k=1,2 ..., q;

Negative sign in formula (6) and (7) represents Gradient Descent, and η ∈ (0,1) represents scale-up factor, it reflects the speed of study.

8. as claimed in claim 6 based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: in described step (3), the concrete grammar of training network is: hidden layer node number is set to 1, training network, then progressively node number is increased, with identical sample training, nodes corresponding when determining that error is minimum; When using the method, use some experimental formulas; The nodes that experimental formula draws is "ball-park" estimate value, as the initial value of method of trial and error.

9., as claimed in claim 1 based on the method for BP neural network prediction Circulating Fluidized Bed Boiler optimum working temperature, it is characterized in that: in described step (3), following two pacing itemss when determining the node number of hidden layer, must be met:

(1) number of training must connect flexible strategy more than model, is 2-10 times; Otherwise, sample needs to be divided into several part, takes the way of " trained in turn ", obtains network model of good performance with this;

(2) the node number of hidden layer is less than N-1, is N number of training; Otherwise, the error of model and sample properties have nothing to do, and go to zero, the model that is set up is without generalization ability; Identical reason, the number of input layer is also less than N-1.