CN106228256A - A kind of photovoltaic module temperature predicting method - Google Patents

Abstract

The present invention relates to a kind of photovoltaic module temperature predicting method, including data initialization, determine input vector and object vector；Build assembly temperature forecast model based on BP artificial neural network；Execution error backpropagation algorithm is trained, and obtains each node output of hidden layer and output layer；Adjust and connect weights, obtain each node output of the hidden layer after adjusting and output layer；If meeting the condition of convergence, then training terminates；Otherwise return step (3).By the prediction of photovoltaic module temperature, improve the accuracy to photovoltaic power generation power prediction.

Description

A kind of photovoltaic module temperature predicting method

Technical field

The present invention relates to new forms of energy power forecasting method, be specifically related to a kind of photovoltaic module temperature predicting method.

Background technology

Photovoltaic power station power generation power has inseparable relation with photovoltaic module.Photovoltaic module and other semiconductor device All the most sensitive to temperature.Along with the rising of temperature, the energy gap of silicon materials reduces, and impact great majority characterize the performance of material Parameter, and then affect the unit for electrical property parameters of assembly, cause the open-circuit voltage of assembly to reduce, short circuit current is slightly increased, and totally makes Success rate reduces.Assembly temperature is the key factor affecting photovoltaic module conversion efficiency, the Accurate Prediction to assembly temperature, It is favorably improved the degree of accuracy of photovoltaic power generation power prediction.Therefore, photovoltaic plant needs to be predicted photovoltaic module temperature.

During prediction, the temperature of photovoltaic module has ratio large effect to the generated output of photovoltaic plant, along with The rising of temperature of photovoltaic battery pack, open-circuit voltage reduces, at 20-100 degree Celsius range, every about raising 1 degree Celsius, photovoltaic The voltage of battery reduces 2mV；And photoelectric current slightly rises with the rising of temperature, every about the photoelectric current raising 1 degree Celsius of battery Increase one thousandth.On the whole, temperature often raises 1 degree Celsius, then power reduces 0.35%.Therefore, it is necessary to proposition photovoltaic Assembly temperature prediction effective ways, to probe into the temperature impact for photovoltaic power station power generation power of photovoltaic module,

Summary of the invention

In order to realize the demand, the present invention provides a kind of photovoltaic module temperature predicting method, can be photovoltaic electric exactly Stand generated output prediction provide basis.

It is an object of the invention to use following technical proposals to realize:

A kind of photovoltaic module temperature predicting method, inputs assembly temperature based on BP artificial neural network respectively by input vector Degree forecast model, is predicted photovoltaic module temperature；Described method includes:

(1) data initialization, determines input vector and object vector；

(2) assembly temperature forecast model based on BP artificial neural network is built；

(3) perform error backpropagation algorithm training, obtain each node output of hidden layer and output layer；

(4) adjust connection weights, obtain each node output of the hidden layer after adjusting and output layer；

(5) if meeting the condition of convergence, then training terminates；Otherwise return step (3).

Preferably, in described step (1), determine that input vector and object vector include: definition affects photovoltaic module temperature The meteorological effect factor, choose the predictive value of the meteorological effect factor in reference period；

Definition meteorological effect factor predictive value is input vector, and photovoltaic module temperature prediction value is object vector；

Further, the described meteorological effect factor, including ambient temperature, solar irradiance and wind speed；

The acquisition methods of described meteorological effect factor predictive value is: gather in historical data, all rings in reference period Border temperature, solar irradiance and wind speed；Record the predictive value of its correspondence respectively.

Preferably, described step (2) specifically includes, and it is optimum for minimizing minE (w, v, θ, γ) with global error function E Solve, build assembly temperature forecast model based on BP artificial neural network by formula (1)；

$\left\{\begin{array}{c}\min E(w,v,\mathrm{\&theta;},\mathrm{\&gamma;})=\frac{1}{N_1}\underset{k=1}{\overset{N_1}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{N}{\mathrm{\&Sigma;}}}{\&lsqb;y_k\left(t\right)-y_k\left(t\right)\&rsqb;}^2<{\mathrm{\&epsiv;}}_1\\ y_k\left(t\right)=\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{v}_{jk}\&CenterDot;f\&lsqb;\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{xw}}_{ij}+{\mathrm{\&theta;}}_j\&rsqb;+{\mathrm{\&gamma;}}_t\\ f\left(x\right)=\frac{1}{1+e^{-x}}\\ \begin{array}{cc}s.t.& w_{ij}\&Element;R^{m\&times;p},v_{jk}\&Element;R^{p\&times;N_1},{\mathrm{\&theta;}}_j\&Element;R^p,{\mathrm{\&gamma;}}_t\&Element;R^N\end{array}\end{array}\right.---\left(1\right)$

Wherein, x is training sample, y_kT () is the actual output of network, y_kT () is the desired output of network, w_ijFor input The connection weights of node layer i to hidden layer node j, i=1,2 ..., n, n are input layer number, j=1,2,3..., m； v_jkFor the connection weights of hidden layer node j to output layer node k, k=1,2,3..., m；M is the number of hidden layer node, θ_jFor Threshold value at hidden layer node j, γ_tFor the threshold value at output node t, t=1,2,3..., p；P is output layer node number, N₁ For the neuron number of hidden layer, N is the neuron number of output layer, and f (x) is activation primitive, ε₁For error predetermined threshold value, R^{m ×p}For m row p column matrix, R^p×N1For p row N₁Column matrix, R^pFor p row 1 column matrix, R^NFor N row 1 column matrix.Preferably, described step (3) in, the execution of error backpropagation algorithm training includes: use gradient descent method so that global error function E is by under gradient Fall, its expression formula is:

$\left\{\begin{array}{c}-\frac{\&part;E}{\&part;w_{ij}}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;w_{ij}})\\ -\frac{\&part;E}{\&part;{\mathrm{\&theta;}}_j}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;{\mathrm{\&theta;}}_j})\\ -\frac{\&part;E}{\&part;v_{jk}}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;v_{jk}})\\ -\frac{\&part;E}{\&part;{\mathrm{\&gamma;}}_t}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;{\mathrm{\&gamma;}}_t})\end{array}\right.---\left(2\right)$

Described gradient descent method, the change connecting weights and threshold value of the most each node is directly proportional to downward gradient, then:

$\left\{\begin{array}{c}{\mathrm{\&Delta;v}}_{jk}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;v_{jt}}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;y_t}\&CenterDot;\frac{\&part;y_t}{\&part;v_{jt}}\\ {\mathrm{\&Delta;w}}_{ij}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;w_{ij}}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;b_j}\&CenterDot;\frac{\&part;b_j}{\&part;s_j}\\ {\mathrm{\&Delta;\&gamma;}}_t=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;{\mathrm{\&gamma;}}_t}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;y_t}\&CenterDot;\frac{\&part;y_t}{\&part;{\mathrm{\&gamma;}}_t}\\ {\mathrm{\&Delta;\&theta;}}_j=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;{\mathrm{\&theta;}}_j}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;b_j}\&CenterDot;\frac{\&part;b_j}{\&part;s_j}\&CenterDot;\frac{\&part;s_j}{\&part;{\mathrm{\&theta;}}_j}\\ b_j=f\left(s_j\right)\\ s_j=f(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{w}_{ij}{x}_i+{\mathrm{\&theta;}}_j)\end{array}\right.---\left(3\right)$

Wherein, η is learning rate, and 0 ＜ η ＜ 1；w_ijFor the connection weights of input layer i to hidden layer node j, v_jkFor Hidden layer node j is to the connection weights of output layer node k, Δ ν_jkHidden layer node j is to the connection modified weight of output layer node k Amount, Δ ω_ijConnection weights correction for input layer i to hidden layer node j；θ_jFor the threshold value at hidden layer node j, Δ θ_jFor the threshold value correction at hidden layer node j, γ_tFor the threshold value at output node t, Δ γ_tThreshold value at output node t is repaiied Positive quantity；b_jFor the output of each neuron of hidden layer, x_iFor training sample corresponding at node i, i=1,2 ..., n, n are input layer Node number, s_jFor the intermediate object program of neuron computing, N₁For the neuron number of hidden layer, m is the number of hidden layer node, N For the neuron number of output layer, f (x) is activation primitive；E_kError signal for hidden layer.

Preferably, hidden layer and each node output of output layer after described step (4) obtains adjustment are shown below:

$\left\{\begin{array}{c}{w}_{ij}(l+1)=w_{ij}\left(l\right)+{\mathrm{\&Delta;w}}_{ij}\\ v_{jk}(l+1)=v_{jk}\left(l\right)+{\mathrm{\&Delta;v}}_{jk}\\ {\mathrm{\&theta;}}_j(l+1)={\mathrm{\&theta;}}_j\left(l\right)+{\mathrm{\&Delta;\&theta;}}_j\\ {\mathrm{\&gamma;}}_t(l+1)={\mathrm{\&gamma;}}_t\left(l\right)+{\mathrm{\&Delta;\&gamma;}}_t\end{array}\right.---\left(4\right)$

Wherein, l represents frequency of training.

Preferably, the condition of convergence in described step (5) is that global error function E is less than error predetermined threshold value ε₁。

Further, described ambient temperature, solar irradiance and wind speed are inputted based on BP people as input vector In the assembly temperature forecast model of artificial neural networks, obtain photovoltaic module temperature prediction value.

Compared with immediate prior art, technical scheme has a following excellent effect:

The Forecasting Methodology of the photovoltaic module temperature that the present invention provides, determines photovoltaic module temperature shadow first with historical data Ring the factor, obtain ambient temperature, solar irradiance and wind speed and be predicted；Secondly, according to ambient temperature, the predictive value of wind speed and Incidence relation between three obtains the predictive value of photovoltaic module temperature.The most intuitively reflect the physics meaning of photovoltaic module temperature Justice, it is possible to reduce the difficulty requirement to forecast model to a certain extent, improves the precision of prediction of model.

Accompanying drawing explanation

A kind of photovoltaic module temperature predicting method flow chart that Fig. 1 provides for the present invention；

Fig. 2 is the assembly temperature forecast model structure chart that the present invention provides；

Fig. 3 is the BP neutral net basic structure schematic diagram that the present invention provides.

Detailed description of the invention

Below in conjunction with the accompanying drawings the detailed description of the invention of the present invention is elaborated.

For making the purpose of the embodiment of the present invention, technical scheme and advantage clearer, below in conjunction with the embodiment of the present invention In accompanying drawing, the technical scheme in the embodiment of the present invention is clearly and completely described, it is clear that described embodiment is The a part of embodiment of the present invention rather than whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art The all other embodiments obtained under not making creative work premise, broadly fall into the scope of protection of the invention.

As it is shown in figure 1, the present invention provide a kind of photovoltaic module temperature predicting method, input vector is inputted respectively based on The assembly temperature forecast model of BP artificial neural network, is predicted photovoltaic module temperature；Described method includes:

(1) data initialization, determines input vector and object vector；In step (1), determine input vector and object vector Including: definition affects the meteorological effect factor of photovoltaic module temperature, chooses the predictive value of the meteorological effect factor in reference period；

Definition meteorological effect factor predictive value is input vector, and photovoltaic module temperature prediction value is object vector；

The meteorological effect factor, including ambient temperature, solar irradiance and wind speed etc.；

The acquisition methods of described meteorological effect factor predictive value is: gather in historical data, all rings in reference period Border temperature, solar irradiance and wind speed；Record the predictive value of its correspondence respectively.

As in figure 2 it is shown, the predictive value of described ambient temperature, solar irradiance and wind speed is inputted as input vector In assembly temperature forecast model based on BP artificial neural network, obtain photovoltaic module temperature prediction value.

(2) assembly temperature forecast model based on BP artificial neural network is built；

BP artificial neural network is a kind of multilayer feedforward neural network based on error backpropagation algorithm training, and it is basic Structure is as it is shown on figure 3, include the error feedback of input layer, hidden layer, output layer and back propagation.

Step (2) specifically includes, and minimizes min E (w, v, θ) γ with global error function E and, for optimal solution, passes through formula (1) assembly temperature forecast model based on BP artificial neural network is built；

$\left\{\begin{array}{c}\min E(w,v,\mathrm{\&theta;},\mathrm{\&gamma;})=\frac{1}{N_1}\underset{k=1}{\overset{N_1}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{N}{\mathrm{\&Sigma;}}}{\&lsqb;y_k\left(t\right)-y_k\left(t\right)\&rsqb;}^2<{\mathrm{\&epsiv;}}_1\\ y_k\left(t\right)=\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{v}_{jk}\&CenterDot;f\&lsqb;\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{xw}}_{ij}+{\mathrm{\&theta;}}_j\&rsqb;+{\mathrm{\&gamma;}}_t\\ f\left(x\right)=\frac{1}{1+e^{-x}}\\ \begin{array}{cc}s.t.& w_{ij}\&Element;R^{m\&times;p},v_{jk}\&Element;R^{p\&times;N_1},{\mathrm{\&theta;}}_j\&Element;R^p,{\mathrm{\&gamma;}}_t\&Element;R^N\end{array}\end{array}\right.---\left(1\right)$

Wherein, x is training sample, y_kT () is the actual output of network, y_kT () is the desired output of network, w_ijFor input The connection weights of node layer i to hidden layer node j, i=1,2 ..., n, n are input layer number, j=1,2,3..., m； v_jkFor the connection weights of hidden layer node j to output layer node k, k=1,2,3..., m；M is the number of hidden layer node, θ_jFor Threshold value at hidden layer node j, γ_tFor the threshold value at output node t, t=1,2,3..., p；P is output layer node number, N₁ For the neuron number of hidden layer, N is the neuron number of output layer, and f (x) is activation primitive, ε₁For error predetermined threshold value,

R^m×pFor m row p column matrix, R^p×N1For p row N₁Column matrix, R^pFor p row 1 column matrix, R^NFor N row 1 column matrix.

(3) perform error backpropagation algorithm training, obtain each node output of hidden layer and output layer；

In step (3), the execution of error backpropagation algorithm training includes: use gradient descent method so that global error Function E is declined by gradient, and its expression formula is:

$\left\{\begin{array}{c}-\frac{\&part;E}{\&part;w_{ij}}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;w_{ij}})\\ -\frac{\&part;E}{\&part;{\mathrm{\&theta;}}_j}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;{\mathrm{\&theta;}}_j})\\ -\frac{\&part;E}{\&part;v_{jk}}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;v_{jk}})\\ -\frac{\&part;E}{\&part;{\mathrm{\&gamma;}}_t}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;{\mathrm{\&gamma;}}_t})\end{array}\right.---\left(2\right)$

Described gradient descent method, the change connecting weights and threshold value of the most each node is directly proportional to downward gradient, then:

$\left\{\begin{array}{c}{\mathrm{\&Delta;v}}_{jk}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;v_{jt}}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;y_t}\&CenterDot;\frac{\&part;y_t}{\&part;v_{jt}}\\ {\mathrm{\&Delta;w}}_{ij}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;w_{ij}}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;b_j}\&CenterDot;\frac{\&part;b_j}{\&part;s_j}\\ {\mathrm{\&Delta;\&gamma;}}_t=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;{\mathrm{\&gamma;}}_t}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;y_t}\&CenterDot;\frac{\&part;y_t}{\&part;{\mathrm{\&gamma;}}_t}\\ {\mathrm{\&Delta;\&theta;}}_j=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;{\mathrm{\&theta;}}_j}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;b_j}\&CenterDot;\frac{\&part;b_j}{\&part;s_j}\&CenterDot;\frac{\&part;s_j}{\&part;{\mathrm{\&theta;}}_j}\\ b_j=f\left(s_j\right)\\ s_j=f(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{w}_{ij}{x}_i+{\mathrm{\&theta;}}_j)\end{array}\right.---\left(3\right)$

Wherein, η is learning rate, and 0 ＜ η ＜ 1；w_ijFor the connection weights of input layer i to hidden layer node j, v_jkFor Hidden layer node j is to the connection weights of output layer node k, Δ ν_jkHidden layer node j is to the connection modified weight of output layer node k Amount, Δ ω_ijConnection weights correction for input layer i to hidden layer node j；θ_jFor the threshold value at hidden layer node j, Δ θ_jFor the threshold value correction at hidden layer node j, γ_tFor the threshold value at output node t, Δ γ_tThreshold value at output node t is repaiied Positive quantity；b_jFor the output of each neuron of hidden layer, x_iFor training sample corresponding at node i, i=1,2 ..., n, n are input layer Node number, s_jFor the intermediate object program of neuron computing, N₁For the neuron number of hidden layer, m is the number of hidden layer node, N For the neuron number of output layer, f (x) is activation primitive；E_kError signal for hidden layer.

(4) adjust connection weights, obtain each node output of the hidden layer after adjusting and output layer；

Step (4) obtains each node output of the hidden layer after adjusting and output layer and is shown below:

Wherein, l represents frequency of training.

(5) if meeting the condition of convergence, then training terminates；Otherwise return step (3)；

The condition of convergence in step (5) is that global error function E is less than self-defining error predetermined threshold value ε₁。

Finally should be noted that: above example is only in order to illustrate that technical scheme is not intended to limit, to the greatest extent The present invention has been described in detail by pipe with reference to above-described embodiment, and those of ordinary skill in the field are it is understood that still The detailed description of the invention of the present invention can be modified or equivalent, and any without departing from spirit and scope of the invention Amendment or equivalent, it all should be contained within the claims that application is awaited the reply.

Claims (8)

1. a photovoltaic module temperature predicting method, it is characterised in that input vector is inputted respectively based on BP ANN The assembly temperature forecast model of network, is predicted photovoltaic module temperature；Described method includes:

(1) data initialization, determines input vector and object vector；

(2) assembly temperature forecast model based on BP artificial neural network is built；

(3) perform error backpropagation algorithm training, obtain each node output of hidden layer and output layer；

(4) adjust connection weights, obtain each node output of the hidden layer after adjusting and output layer；

(5) if meeting the condition of convergence, then training terminates；Otherwise return step (3).

2. the method for claim 1, it is characterised in that in described step (1), determines input vector and object vector bag Include: definition affects the meteorological effect factor of photovoltaic module temperature, choose the predictive value of the meteorological effect factor in reference period；

Definition meteorological effect factor predictive value is input vector, and photovoltaic module temperature prediction value is object vector；

3. method as claimed in claim 2, it is characterised in that the described meteorological effect factor, including ambient temperature, solar irradiation Degree and wind speed；

The acquisition methods of described meteorological effect factor predictive value is: gather in historical data, all environment temperature in reference period Degree, solar irradiance and wind speed；Record the predictive value of its correspondence respectively.

4. the method for claim 1, it is characterised in that described step (2) specifically includes, with global error function E's Minimizing minE (w, v, θ, γ) is optimal solution, builds assembly temperature based on BP artificial neural network prediction mould by formula (1) Type；

$\left\{\begin{array}{c}\min E(w,v,\mathrm{\&theta;},\mathrm{\&gamma;})=\frac{1}{N_1}\underset{k=1}{\overset{N_1}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{N}{\mathrm{\&Sigma;}}}{\&lsqb;y_k\left(t\right)-y_k\left(t\right)\&rsqb;}^2<{\mathrm{\&epsiv;}}_1\\ y_k\left(t\right)=\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{v}_{jk}\&CenterDot;f\&lsqb;\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{xw}}_{ij}+{\mathrm{\&theta;}}_j\&rsqb;+{\mathrm{\&gamma;}}_t\\ f\left(x\right)=\frac{1}{1+e^{-x}}\\ \begin{array}{cc}s.t.& w_{ij}\&Element;R^{m\&times;p},v_{jk}\&Element;R^{p\&times;N_1},{\mathrm{\&theta;}}_j\&Element;R^p,{\mathrm{\&gamma;}}_t\&Element;R^N\end{array}\end{array}\right.---\left(1\right)$

Wherein, x is training sample, y_kT () is the actual output of network, y_kT () is the desired output of network, w_ijSave for input layer The connection weights of some i to hidden layer node j, i=1,2 ..., n, n are input layer number, j=1,2,3..., m；v_jkFor Hidden layer node j is to the connection weights of output layer node k, k=1,2,3..., m；M is the number of hidden layer node, θ_jIt is implicit Threshold value at node layer j, γ_tFor the threshold value at output node t, t=1,2,3..., p；P is output layer node number, N₁For hidden Neuron number containing layer, N is the neuron number of output layer, and f (x) is activation primitive, ε₁For error predetermined threshold value, R^m×pFor m Row p column matrix, R^p×N1For p row N₁Column matrix, R^pFor p row 1 column matrix, R^NFor N row 1 column matrix.

5. the method as described in claim 1 or 4, in described step (3), the execution of error backpropagation algorithm training includes: Use gradient descent method so that global error function E is declined by gradient, and its expression formula is:

$\left\{\begin{array}{c}-\frac{\&part;E}{\&part;w_{ij}}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;w_{ij}})\\ -\frac{\&part;E}{\&part;{\mathrm{\&theta;}}_j}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;{\mathrm{\&theta;}}_j})\\ -\frac{\&part;E}{\&part;v_{jk}}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;v_{jk}})\\ -\frac{\&part;E}{\&part;{\mathrm{\&gamma;}}_t}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(-\frac{\&part;E_k}{\&part;{\mathrm{\&gamma;}}_t})\end{array}\right.---\left(2\right)$

Described gradient descent method, the change connecting weights and threshold value of the most each node is directly proportional to downward gradient, then:

$\left\{\begin{array}{c}{\mathrm{\&Delta;v}}_{jk}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;v_{jt}}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;y_t}\&CenterDot;\frac{\&part;y_t}{\&part;v_{jt}}\\ {\mathrm{\&Delta;w}}_{ij}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;w_{ij}}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;b_j}\&CenterDot;\frac{\&part;b_j}{\&part;s_j}\\ {\mathrm{\&Delta;\&gamma;}}_t=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;{\mathrm{\&gamma;}}_t}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;y_t}\&CenterDot;\frac{\&part;y_t}{\&part;{\mathrm{\&gamma;}}_t}\\ {\mathrm{\&Delta;\&theta;}}_j=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;{\mathrm{\&theta;}}_j}=-\mathrm{\&eta;}\&CenterDot;\frac{\&part;E_k}{\&part;b_j}\&CenterDot;\frac{\&part;b_j}{\&part;s_j}\&CenterDot;\frac{\&part;s_j}{\&part;{\mathrm{\&theta;}}_j}\\ b_j=f\left(s_j\right)\\ s_j=f(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{w}_{ij}{x}_i+{\mathrm{\&theta;}}_j)\end{array}\right.---\left(3\right)$

Wherein, η is learning rate, and 0 ＜ η ＜ 1；w_ijFor the connection weights of input layer i to hidden layer node j, v_jkIt is implicit Node layer j is to the connection weights of output layer node k, Δ ν_jkThe connection weights correction of hidden layer node j to output layer node k, Δω_ijConnection weights correction for input layer i to hidden layer node j；θ_jFor the threshold value at hidden layer node j, Δ θ_j For the threshold value correction at hidden layer node j, γ_tFor the threshold value at output node t, Δ γ_tThreshold value correction at output node t Amount；b_jFor the output of each neuron of hidden layer, x_iFor training sample corresponding at node i, i=1,2 ..., n, n are input layer joint Point number, s_jFor the intermediate object program of neuron computing, N₁For the neuron number of hidden layer, m is the number of hidden layer node, and N is The neuron number of output layer, f (x) is activation primitive；E_kError signal for hidden layer.

6. the method as described in claim 1 or 4, it is characterised in that described step (4) obtains the hidden layer after adjusting and output Each node output of layer is shown below:

$\left\{\begin{array}{c}{w}_{ij}(l+1)=w_{ij}\left(l\right)+{\mathrm{\&Delta;w}}_{ij}\\ v_{jk}(l+1)=v_{jk}\left(l\right)+{\mathrm{\&Delta;v}}_{jk}\\ {\mathrm{\&theta;}}_j(l+1)={\mathrm{\&theta;}}_j\left(l\right)+{\mathrm{\&Delta;\&theta;}}_j\\ {\mathrm{\&gamma;}}_t(l+1)={\mathrm{\&gamma;}}_t\left(l\right)+{\mathrm{\&Delta;\&gamma;}}_t\end{array}\right.---\left(4\right)$

Wherein, l represents frequency of training.

7. the method for claim 1, it is characterised in that the condition of convergence in described step (5) is global error function E Less than error predetermined threshold value ε₁。

8. method as claimed in claim 3, it is characterised in that by the prediction of described ambient temperature, solar irradiance and wind speed It is worth and inputs in assembly temperature forecast model based on BP artificial neural network respectively as input vector, obtain photovoltaic module temperature Predictive value.