CN104463356A - Photovoltaic power generation power prediction method based on multi-dimension information artificial neural network algorithm - Google Patents

Abstract

A photovoltaic power generation power prediction method based on a multi-dimension artificial neural network algorithm includes the following steps that (1), according to similarity calculation, day characteristic data provided by the weather bureau and day types described in a relative fuzzy mode serve as vectors for similarity calculation; (2), according to selection of similar days, starting from the latest history day, the similarity value FN of the jth day and the ith day is calculated reversely day by day, and an m day highest in similarity within the recent period or an m day of which the similarity is larger than 0.80 is selected as the similarity day of a prediction day; (3), a prediction model is determined and trained, three layers of BP neural networks are adopted for establishing different training samples, new testing data are continuously injected, and BP neural network training is performed on the history day and a similar photovoltaic power generation power sample; (4), a prediction result is output, m similar days of the prediction day are selected, and samples of the similar days are input into trained BP neural networks; (5), a prediction error is analyzed, and a root-mean-square error is adopted as a model estimation method.

Description

A kind of photovoltaic power generation power prediction method based on multidimensional information artificial neural network algorithm

Technical field

What the present invention relates to is a kind of photovoltaic power generation power prediction method based on multidimensional information artificial neural network algorithm, is mainly used for improving electric system photovoltaic generation precision of prediction.

Background technology

Photovoltaic power generation power prediction is urgent problem in energy management.Photovoltaic generation is the same with wind-power electricity generation, all belong to undulatory property and intermittent power supply, and, photovoltaic cell kind and the installation site randomness thereof of each user or community use are also large, photovoltaic generating system is by intensity of illumination and environment, the impact of the climatic factors such as temperature, the change of output power has uncertainty, the disturbance of output power will likely affect the stable of electrical network, therefore, need the research strengthening photovoltaic power generation power prediction, obtain the daily generation curve of photovoltaic generating system in advance, thus coordinate electric system formulation generation schedule, reduce the Randomization of photovoltaic generation to the impact of electric system.Using accumulator to carry out the output of stable photovoltaic generation power is a kind of feasible method, but needs additional cost, and used and scrapped battery also can cause environmental pollution.Therefore, need to carry out Accurate Prediction to the generated output of photovoltaic system, to understand the generator operation characteristic of large-scale solar photovoltaic grid-connection system and the matching problem with dispatching of power netwoks, electric load etc., contribute to planning and the operation of whole electric system like this, thus reduce photovoltaic generation randomness to the impact of electric system, improve the security and stability of system.

Summary of the invention

The object of the invention is to the deficiency overcoming prior art existence, and a kind of precision and accuracy by improving the online real-time estimate of electric system photovoltaic generation based on multidimensional information artificial neural network algorithm are provided, take into full account the many factors to photovoltaic generation predicted impact, construct suitable neural network model, set up different training samples, carry out the photovoltaic power generation power prediction method based on multidimensional information artificial neural network algorithm of repetition training.

The object of the invention is to have come by following technical solution, a kind of photovoltaic power generation power prediction method based on multidimensional information artificial neural network algorithm, this photovoltaic power generation power prediction method comprises the steps:

1) calculating of similarity: the day characteristic utilizing weather bureau to provide, comprise medial temperature, maximum temperature, minimum temperature etc., and the comparison vague description day type that weather bureau provides, comprise the vector as Similarity Measure such as fine day, cloudy, cloudy, rainy day; Described day, the day such as fine day, cloudy, cloudy, rainy day type information to be mainly mapped as the day index of type as the input variable of forecast model between 0-1 according to the statistical study of historical data and experience by type information; For the raw data used in model may bring unfavorable, data are normalized;

2) selection of similar day: from closing on most history day, the reverse Similarity value F calculating jth day and i-th day day by day _n; Choose m day that in nearest a period of time, similarity is the highest or similarity is greater than the similar day of 0.80m day as prediction day, the value of m is greater than 3, has stronger extrapolation to meet training network;

3) establishment of forecast model and the training of model: adopt three layers of BP neural network, comprise input layer, output layer and hidden layer, set up different training samples, and constantly inject new test data, BP neural metwork training is carried out to history day and its similar photovoltaic generation power sample;

4) output predicted the outcome: m the similar day selecting prediction day, by the BP neural network that the input of similar day sample trains, draws the photovoltaic DC field generated output of prediction day;

5) predicated error analysis: for ensureing the stationarity of model predictive error, adopts root-mean-square error as model evaluation method;

$\mathrm{RMSE}=\sqrt{\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\frac{{(P_f^i-P_a^i)}^2}{N}}$

In formula: N is the number of data; for predicted value; for actual value.

2, the photovoltaic power generation power prediction method based on multidimensional information artificial neural network algorithm according to claim 1, is characterized in that:

Described step 1) in, mainly utilize weather bureau to provide day feature weather information to select similar day;

The influence factor of photovoltaic generation power is constructed as follows vector:

$Y=[\stackrel{\‾}{T},T_{\min },T_{\max }X]$

In formula: for medial temperature, T _min, T _maxbe respectively minimum temperature and maximum temperature, X is day type;

Described normalization formula is:

$P_n=\frac{p_n-p_{\min }}{p_{\max }-p_{\min }}$

$N_n=\frac{n_n-n_{\min }}{n_{\max }-n_{\min }}$

In formula: p _n, n _n---original input, target data;

P _min, p _max, n _min, n _max---the minimum value in p and n and maximal value;

P _n, N _n---the input after normalization, target data

Impact vector:

Y _p＝[Y _p(1),Y _p(2),Y _p(3),Y _p(4)] ^T

Y _N＝[Y _N(1),Y _N(2),Y _N(3),Y _N(4)] ^T

Y _pwith Y _nat the correlation coefficient of a kth factor be

${\mathrm{\ϵ}}_N\left(k\right)=\frac{\underset{N}{\min }\underset{k}{\min }|Y_P\left(k\right)-Y_N\left(k\right)|+\mathrm{\ρ}\underset{N}{\max }\underset{k}{\max }|Y_P\left(k\right)-Y_N\left(k\right)|}{|Y_P\left(k\right)-Y_N\left(k\right)|+\mathrm{\ρ}\underset{N}{\max }\underset{k}{\max }|Y_P\left(k\right)-Y_N\left(k\right)|}$

In formula: ρ is resolution ratio, its value generally gets 0.5; The correlation coefficient of comprehensive each point, defines whole Y _pwith Y _nsimilarity be

$F_N=\underset{k=1}{\overset{4}{\mathrm{\Π}}}{\mathrm{\ϵ}}_N\left(k\right)$

Adopt this company to take advantage of mode to define similarity, simply, automatically can identify leading factor, and solve each factor weight setting problem.

Step 3) in: defining of three layers of BP neural network node:

The determination of a, input layer:

Input layer corresponds to the input variable of model, and model adopts 10 input variables, and these 10 input variables are the photovoltaic generation data of similar day;

The determination of b, output layer node:

This is determined by the content that will predict.Output vector (object vector) is herein the generated output of prediction day, and therefore output layer node number is 1;

The determination of c, hidden layer and Hidden nodes:

The empirical formula that hidden layer node is conventional:

$m=\sqrt{n+l}+\mathrm{\α}$

m＝log2 ⁿ

$m=\sqrt{\mathrm{nl}}$

Various middle m is the number of hidden nodes above, and n is input layer number, and l is output node number, and α is the constant between 1-10.Because input layer is 10 points, output layer node is 1 point, considers above various, determines that the number of hidden nodes is 5 eventually.

The invention has the beneficial effects as follows:

By multidimensional information artificial neural network algorithm, take into full account the many factors to photovoltaic generation predicted impact, make the accuracy of input amendment higher, the foundation of forecast model is more suitable for, and reaches the effect improved photovoltaic power generation power prediction precision.Contribute to scheduling and the operation of electric system.

Accompanying drawing illustrates:

Fig. 1 is the process flow diagram of photovoltaic generation prediction.

Fig. 2 is the actual prediction result figure of certain photovoltaic plant.

Embodiment

The present invention will be described in detail below in conjunction with the accompanying drawings and the specific embodiments: shown in Fig. 1, and a kind of photovoltaic power generation power prediction method based on multidimensional information artificial neural network algorithm of the present invention, this photovoltaic power generation power prediction method comprises the steps:

1) calculating of similarity, utilize weather bureau to provide day feature weather information to select similar day.

The influence factor of photovoltaic generation power is constructed as follows vector:

$Y=[\stackrel{\‾}{T},T_{\min },T_{\max }X]$

In formula: for medial temperature, T _min, T _maxbe respectively minimum temperature and maximum temperature, X is day type.

The day type that weather bureau provides is fuzzyyer description: as fine day, fine day to cloudy, cloudy, cloudy day have light rain, light rain turns heavy rain etc., first according to the statistical study of history generated energy and experience the day such as fine day, cloudy, cloudy, rainy day type information is mapped as between 0-1 a day index of type as the input variable of forecast model.For eliminating the difference of the different unit variance of raw data, varying number level, unfavorable with what eliminate that raw data form difference brings, data are normalized.

Normalization formula:

$P_n=\frac{p_n-p_{\min }}{p_{\max }-p_{\min }}$

$N_n=\frac{n_n-n_{\min }}{n_{\max }-n_{\min }}$

In formula: p _n, n _n---original input, target data;

P _min, p _max, n _min, n _max---the minimum value in p and n and maximal value;

P _n, N _n---the input after normalization, target data

Impact vector:

Y _p＝[Y _p(1),Y _p(2),Y _p(3),Y _p(4)] ^T

Y _N＝[Y _N(1),Y _N(2),Y _N(3),Y _N(4)] ^T

Y _pwith Y _nat the correlation coefficient of a kth factor be

${\mathrm{\ϵ}}_N\left(k\right)=\frac{\underset{N}{\min }\underset{k}{\min }|Y_P\left(k\right)-Y_N\left(k\right)|+\mathrm{\ρ}\underset{N}{\max }\underset{k}{\max }|Y_P\left(k\right)-Y_N\left(k\right)|}{|Y_P\left(k\right)-Y_N\left(k\right)|+\mathrm{\ρ}\underset{N}{\max }\underset{k}{\max }|Y_P\left(k\right)-Y_N\left(k\right)|}$

In formula: ρ is resolution ratio, its value generally gets 0.5.The correlation coefficient of comprehensive each point, defines whole Y _pwith Y _nsimilarity be

$F_N=\underset{k=1}{\overset{4}{\mathrm{\Π}}}{\mathrm{\ϵ}}_N\left(k\right)$

Adopt this company to take advantage of mode to define similarity, simply, automatically can identify leading factor, and solve each factor weight setting problem.

2) selection of similar day, from closing on most history day, the reverse Similarity value F calculating jth day and i-th day day by day _n.Choose m day or similarity F that in nearest a period of time, similarity is the highest _nm day of>=r (r is certain numerical value, gets r=0.80 herein) is as the similar day of N day.Usually, similar day number m is greater than 3, has stronger extrapolation to meet training network.

3) determination of forecast model and the training of model; Adopt three layers of BP neural network, comprise input layer, output layer and hidden layer.

The determination of a, input layer

Input layer corresponds to the input variable of model, and model adopts 10 input variables, and these 10 input variables are the photovoltaic generation data of similar day;

The determination of b, output layer node:

This is determined by the content that will predict; Output vector of the present invention (object vector) is the generated output of prediction day, and therefore output layer node number is 1;

The determination of c, hidden layer and Hidden nodes:

The empirical formula that hidden layer node is conventional:

$m=\sqrt{n+l}+\mathrm{\α}$

m＝log2 ⁿ

$m=\sqrt{\mathrm{nl}}$

Various middle m is the number of hidden nodes above, and n is input layer number, and l is output node number, and α is the constant between 1-10.Because input layer is 10 points, output layer node is 1 point, considers above various, determines that the number of hidden nodes is 5 eventually;

The training of model: using history day real data as output, using this history day m similar day same time data as input, BP neural metwork training is carried out to history day and its similar photovoltaic generation power sample.For the uncertainty of the following a period of time generated energy of photovoltaic generating system, sample data under different weather, season is divided, considers the correlativity of photovoltaic generation and meteorological department's weather forecast, set up different training samples, and constantly inject new test data, to improve precision;

4) predict the outcome output, selects m the similar day of prediction day, by the BP neural network that the input of similar day sample trains, draws the photovoltaic DC field generated output of prediction day, as shown in Figure 2.

5) predicated error analysis, for ensureing the stationarity of model predictive error, adopts root-mean-square error as model evaluation method;

$\mathrm{RMSE}=\sqrt{\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\frac{{(P_f^i-P_a^i)}^2}{N}}$

In formula: N is the number of data; for predicted value; for actual value.

Embodiment:

The present invention chooses most suitable similar day from history Japan and China, makes input amendment bad data rate lower, is conducive to prediction convergency value closer to actual value.

The influence factor of photovoltaic generation power of the present invention is constructed as follows vector:

$Y=[\stackrel{\‾}{T},T_{\min },T_{\max }X]$

In formula: for medial temperature, T _min, T _maxbe respectively minimum temperature and maximum temperature, X is day type.

By Application of Neural Network when the forecasting problem of systems generate electricity power, in the raw data of training network, different variablees is usually with different unit variance, the difference of the order of magnitude is also larger, if the variation range of generated energy is between 0 to 100, but the common variation range of temperature is then between-10 to 40.Can be known by the characteristic of neuron activation functions, neuronic output is limited in certain scope usually, the nonlinear activation function used in the application of most people artificial neural networks is S function, its output is limited at (0,1) or (-1,1) between, directly carrying out training with raw data to network can cause neuron saturated, therefore pre-service must be carried out to data before network is trained, with eliminate raw data form difference bring unfavorable, common way is normalized.Research shows: being normalized data in a proper manner can the convergence of accelerans network.

Normalization independently can be carried out on the single input variable passage of model, also can carry out together all input channels.The method for normalizing that input variable is commonly used has following several: (1) simple normalization; (2) linear transformation is interval to [0,1]; (3) linear transformation is on interval [a, b].When needs input and target data fall into [0,1] interval, normalization formula is

$P_n=\frac{p_n-p_{\min }}{p_{\max }-p_{\min }}$

$N_n=\frac{n_n-n_{\min }}{n_{\max }-n_{\min }}$

In formula: p _n, n _n---original input, target data;

P _min, p _max, n _min, n _max---the minimum value in p and n and maximal value;

P _n, N _n---the input after normalization, target data

Impact vector:

Y _p＝[Y _p(1),Y _p(2),Y _p(3),Y _p(4)] ^T

Y _N＝[Y _N(1),Y _N(2),Y _N(3),Y _N(4)] ^T

Y _pwith Y _nat the correlation coefficient of a kth factor be

${\mathrm{\ϵ}}_N\left(k\right)=\frac{\underset{N}{\min }\underset{k}{\min }|Y_P\left(k\right)-Y_N\left(k\right)|+\mathrm{\ρ}\underset{N}{\max }\underset{k}{\max }|Y_P\left(k\right)-Y_N\left(k\right)|}{|Y_P\left(k\right)-Y_N\left(k\right)|+\mathrm{\ρ}\underset{N}{\max }\underset{k}{\max }|Y_P\left(k\right)-Y_N\left(k\right)|}$

In formula: ρ is resolution ratio, its value generally gets 0.5.The correlation coefficient of comprehensive each point, defines whole Y _pwith Y _nsimilarity be

$F_N=\underset{k=1}{\overset{4}{\mathrm{\Π}}}{\mathrm{\ϵ}}_N\left(k\right)$

Adopt this company to take advantage of mode to define similarity, simply, automatically can identify leading factor, and solve each factor weight setting problem.

The concrete steps of N day similar day are selected to be:

1) from closing on most history day, the reverse Similarity value F calculating jth day and i-th day day by day _n.

2) the m day that in nearest a period of time, similarity is the highest or similarity F is chosen _nm day of>=r (r is certain numerical value, gets r=0.80 herein) is as the similar day of N day.

Adopt three layers of BP neural network, consider the correlativity of photovoltaic generation and meteorological department's weather forecast, set up different training samples, and constantly inject new test data, repetition training is carried out to forecast model.

Select m the similar day of prediction day, by the BP neural network that the input of similar day sample trains, draw the photovoltaic DC field generated output of prediction day.

Adopt root-mean-square error as model evaluation method.

Claims (2)

1., based on a photovoltaic power generation power prediction method for multidimensional information artificial neural network algorithm, it is characterized in that this photovoltaic power generation power prediction method comprises the steps:

1) calculating of similarity: the day characteristic utilizing weather bureau to provide, comprise medial temperature, maximum temperature, minimum temperature etc., and the comparison vague description day type that weather bureau provides, comprise the vector as Similarity Measure such as fine day, cloudy, cloudy, rainy day; Described day, the day such as fine day, cloudy, cloudy, rainy day type information to be mainly mapped as the day index of type as the input variable of forecast model between 0-1 according to the statistical study of historical data and experience by type information; For the raw data used in model may bring unfavorable, data are normalized;

2) selection of similar day: from closing on most history day, the reverse Similarity value F calculating jth day and i-th day day by day _n; Choose m day that in nearest a period of time, similarity is the highest or similarity is greater than the similar day of 0.80m day as prediction day, the value of m is greater than 3, has stronger extrapolation to meet training network;

3) establishment of forecast model and the training of model: adopt three layers of BP neural network, comprise input layer, output layer and hidden layer, set up different training samples, and constantly inject new test data, BP neural metwork training is carried out to history day and its similar photovoltaic generation power sample;

4) output predicted the outcome: m the similar day selecting prediction day, by the BP neural network that the input of similar day sample trains, draws the photovoltaic DC field generated output of prediction day;

5) predicated error analysis: for ensureing the stationarity of model predictive error, adopts root-mean-square error as model evaluation method;

$\mathrm{RMSE}=\sqrt{\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\frac{{(P_f^i-P_a^i)}^2}{N}}$

In formula: N is the number of data; for predicted value; for actual value.

2. the photovoltaic power generation power prediction method based on multidimensional information artificial neural network algorithm according to claim 1, is characterized in that:

Described step 1) in, mainly utilize weather bureau to provide day feature weather information to select similar day;

The influence factor of photovoltaic generation power is constructed as follows vector:

$Y=[\stackrel{\‾}{T},T_{\min },T_{\max },X]$

In formula: for medial temperature, T _min, T _maxbe respectively minimum temperature and maximum temperature, X is day type;

Described normalization formula is:

$P_n=\frac{p_n-p_{\min }}{p_{\max }-p_{\min }}$

$N_n=\frac{n_n-n_{\min }}{n_{\max }-n_{\min }}$

In formula: p _n, n _n---original input, target data;

P _min, p _max, n _min, n _max---the minimum value in p and n and maximal value;

P _n, N _n---the input after normalization, target data

Impact vector:

$Y_p={[Y_{p\left(1\right)},Y_{p\left(2\right)},Y_{p\left(3\right)},Y_{p\left(4\right)}]}^T$

$Y_N={[Y_{N\left(1\right)},Y_{N\left(2\right)},Y_{N\left(3\right)},Y_{N\left(4\right)}]}^T$

Y _pwith Y _nat the correlation coefficient of a kth factor be

${\mathrm{\ϵ}}_N\left(k\right)=\frac{\underset{N}{\min }\underset{k}{\min }|Y_P\left(k\right)-Y_N\left(k\right)|+\mathrm{\ρ}\underset{N}{\max }\underset{k}{\max }|Y_P\left(k\right)-Y_N\left(k\right)|}{|Y_P\left(k\right)-Y_N\left(k\right)|+\mathrm{\ρ}\underset{N}{\max }\underset{k}{\max }|Y_P\left(k\right)-Y_N\left(k\right)|}$

In formula: ρ is resolution ratio, its value generally gets 0.5; The correlation coefficient of comprehensive each point, defines whole Y _pwith Y _nsimilarity be

$F_N=\underset{k=1}{\overset{4}{\mathrm{\Π}}}{\mathrm{\ϵ}}_N\left(k\right)$

Adopt this company to take advantage of mode to define similarity, simply, automatically can identify leading factor, and solve each factor weight setting problem.

Step 3) in: defining of three layers of BP neural network node:

The determination of a, input layer:

Input layer corresponds to the input variable of model, and model adopts 10 input variables, and these 10 input variables are the photovoltaic generation data of similar day;

The determination of b, output layer node:

This is determined by the content that will predict.Output vector (object vector) is herein the generated output of prediction day, and therefore output layer node number is 1;

The determination of c, hidden layer and Hidden nodes:

The empirical formula that hidden layer node is conventional:

$m=\sqrt{n+l}+\mathrm{\α}$

m＝log2 ⁿ

$m=\sqrt{\mathrm{nl}}$

Various middle m is the number of hidden nodes above, and n is input layer number, and l is output node number, and α is the constant between 1-10.Because input layer is 10 points, output layer node is 1 point, considers above various, determines that the number of hidden nodes is 5 eventually.