CN105184678A - Method for constructing photovoltaic power station generation capacity short-term prediction model based on multiple neural network combinational algorithms - Google Patents

Abstract

The invention provides a method for constructing a photovoltaic power station generation capacity short-term prediction model based on multiple neural network combinational algorithms and belongs to the technical field of photovoltaic power generation, power grid connection technology and solar energy photovoltaic forecasting. The method overcomes the problem that a usually-used algorithm for constructing the photovoltaic power station generation capacity short-term prediction model is single and is likely to fall into local optimization, further resulting in big measurement error of the prediction model. The technical construction method of the invention is realized as follows: firstly using four different neural network algorithms to construct sub-models for neural network prediction; secondly screening and classifying weather information and analyzing the suitability of the various sub-models for neural network prediction; giving weighted parameter values of the sub-models in a combined model according to the suitability to further make the combined neural network model for prediction suitable for different weather conditions and then completing the construction of the photovoltaic power station generation capacity short-term prediction model. The method is mainly used for photovoltaic power station grid connection short-term prediction.

Description

Based on the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model of multiple neural network group hop algorithm

Technical field

The invention belongs to photovoltaic generation and interconnection technology, solar energy power generating amount forecast field.

Background technology

Along with the fast development of world economy, the demand of the energy increases day by day, traditional non-regeneration energy (coal, oil, rock gas etc.) day by day reduces and faces exhaustion, the study hotspot that sun power relies on its cleanliness without any pollution, low cost, high-effect, reserves are large, renewable and self the advantage such as to have a very wide distribution has become countries in the world.Become the important channel of Application of Solar Energy in China's solar energy power generating, the scale of grid-connected photovoltaic power station and quantity are all continuing to increase.The output of photovoltaic generating system affects by the factor such as weather and intensity of solar radiation, has undulatory property and intermittence, is uncontrollable factor concerning grid-connected, can affect the safety of electric system and stablize.Therefore setting up the high forecast model of accuracy is the important means improving grid-connected rear bulk power grid safety, is also technological difficulties.What exist in current invention mainly contains following two classes based on photovoltaic power station power generation amount forecasting model, one class is principle predicted method, the method does not consider that weather environment factor affects, for the phenomenon that there is energy loss in solar electrical energy generation process in opto-electronic conversion link and inversion link, set up experimental formula and experience factor, prediction photovoltaic power generation quantity, the advantage of the method is principle and calculates all very simple, but the short error of time scale of prediction is also comparatively large, the short forecast model of time scale requires that electrical network has higher adaptability to changes.Another kind of is artificial intelligent predicting method, be mainly the forecast model based on neural network, the association of neural network, self study and memory capability is utilized to predict photovoltaic power generation quantity, affect greatly because photovoltaic plant exports by weather environment, different meteorologic factors and special environment condition can make neural network produce comparatively big error when training and predict, therefore weather and the impact of environmental factor on forecast model accuracy be can not ignore.

At present, mostly being based on the method for building up of neural network prediction model in invention selects single a kind of neural network or its innovatory algorithm to set up, main as RBF or BP neural network prediction method, wherein fractional prediction model considers the impact of meteorologic factor and environmental baseline in its Individual forecast algorithm adopted, and improves the accuracy of prediction to a certain extent.But still lack at present, for Various Seasonal and meteorologic factor environmental baseline, the Comparative Study of applicability is carried out to different neural network algorithm, not yet analyze the optimization model of different model under similar weather, the adaptability of the forecast model therefore built on this Research foundation to weather and meteorologic factor is more weak.In addition, to stress in the current patent for Neural Network Prediction model or to stay in the research of algorithm predicts model more, and the algorithm predicts model that binding goes out, design and provide a kind of still rarely found for practical grid-connected photovoltaic power station generated energy output power forecasting software platform further.The theoretical research result explored photovoltaic generation is predicted is converted into the utility instructing actual prediction, also has Practical significance to promoting that large-scale photovoltaic power station is grid-connected.

Summary of the invention

The present invention in order to solve existing structure photovoltaic power station power generation amount forecast model selected by algorithm single, be common to the poor for applicability of different weather, be easily absorbed in local optimum, and then cause the problem that forecast model measuring error is larger.The invention provides a kind of construction method of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm.

Based on the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model of multiple neural network group hop algorithm, the detailed process of this construction method is:

Step one: choosing of neural network algorithm;

BP, Elman, RBF and GRNN tetra-kinds of neural network algorithms are adopted to build neural network prediction submodel A, B, C and D respectively, with the environment temperature T of power station every day working time section different time points _i, per day intensity of solar radiation per day wind speed as the input data of photovoltaic power generation quantity Short-term Forecasting Model, to predict the photovoltaic output power P of day corresponding time point _ias the output data of each neural network prediction submodel A, B, C and D;

Step 2: choosing of sample data;

By photovoltaic data acquisition platform, environment parameter and the generating parameter historical data of the fine day same period, cloudy weather and rainy weather is chosen from historical data base, the Weather information same period is carried out sifting sort, reject unusual data point, in conjunction with the Weather information same period, arrange out the sample data under three kinds of weather conditions respectively, three kinds of described weather conditions are respectively fine day, cloudy weather and rainy weather;

Step 3: the training of each neural network prediction submodel, the applied research analysis of each neural network prediction submodel and the structure to photovoltaic power generation quantity Short-term Forecasting Model;

By step 2 gained in conjunction with the Weather information same period, arrange out the sample data under three kinds of weather conditions respectively, respectively neural network submodel A, B, C and D of building are trained, wherein, the input quantity of above-mentioned four neural network submodels A, B, C and D includes 25 variablees, and described 25 variablees are respectively the environment temperature T of prediction day 23 time points _i, per day intensity of solar radiation with per day wind speed described prediction day, 23 time points were in the working time section 7:00 to 18:00 of power station every day, with 7:00 point for start time point, be a time point, totally 23 time points every 30 minutes, after training, draw the neural network prediction submodel A under three kinds of weather conditions ¹, B ¹, C ¹and D ¹,

According to the neural network prediction submodel A under three kinds of weather conditions after training ¹, B ¹, C ¹and D ¹show that BP, Elman, RBF and GRNN tetra-kinds of neural network algorithms are to the applicability conclusion of three kinds of weather conditions by the comparative analysis of root-mean-square error curve and percentage error curve, the weight modulus value parameter of each neural network prediction submodel is provided according to DIFFERENT METEOROLOGICAL CONDITIONS, build modulus value weight parameter table, and then build combination neural net forecast model, using combination neural net forecast model as final photovoltaic power generation quantity Short-term Forecasting Model, complete the structure of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm.

Based on the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model of multiple neural network group hop algorithm, the method also comprises step 4: to the correction of photovoltaic power generation quantity Short-term Forecasting Model;

A) first, the sample data in step 2 is normalized,

B) secondly, genetic algorithm and particle cluster algorithm is adopted.

Based on the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model of multiple neural network group hop algorithm, the method also comprises step 5: to the assessment of photovoltaic power generation quantity Short-term Forecasting Model;

Mean absolute percentage error MAPE and root-mean-square error RMSE two kinds of error assessment methods are adopted to carry out error evaluation to photovoltaic power generation quantity Short-term Forecasting Model,

$MAPE=\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}\left|\frac{P_p^i-P_a^i}{P_a^i}\right|\×100\%,$

$RMSE=\sqrt{{\frac{\underset{i-1}{\overset{N}{\Σ}}(P_p^i-P_a^i)}{N}}^2},$

Wherein, N represents data total amount, and i is positive integer, represent the predicted value of i-th data point, represent the actual value of i-th data point.

Neural network prediction submodel A, B, C and D include input layer, output layer and hidden layer,

To predict the environment temperature T of day 23 time points _i, per day intensity of solar radiation with per day wind speed totally 25 variablees are as the input layer input data of each neural network prediction submodel A, B, C and D,

To predict the photovoltaic output power P of day corresponding 23 time points _iall as the output data of each neural network prediction submodel A, B, C and D,

The output layer activation function of neural network prediction submodel A, B, C and D all adopts pureline function to realize,

The hidden layer of neural network prediction submodel A, B, C and D all adopts single layer structure, hidden layer neuron node experimental formula is utilized to obtain neural network start node number, adopt method of trial and error, obtaining hidden layer node number is 15, and hidden layer activation function adopts tansig function to realize.

Principle analysis: by photovoltaic data acquisition platform, the data of 1 year are divided by spring, summer, autumn and winter four Various Seasonal, environment parameter and the generating parameter historical data of fine day in same season, cloudy weather and rainy weather is chosen from historical data base, the seasonal weather information same period is carried out sifting sort, reject the data point that error is larger, in conjunction with the Weather information same period, arrange out the data under three kinds of weather conditions respectively, and sample data is normalized, reduce neural network algorithm training error further; Single for the algorithm selected in invention at present, the problem that predicated error is still undesirable, adopts neural network group hop algorithm, sets up Combination neural network model.Built-up pattern selects four kinds of different neural network algorithms to set up neural network prediction submodel for three kinds of typical meteorological conditions respectively, and carry out applied research and analysis, by choosing of weight modulus value parameter, the predictor model adopting applicability higher is realized for different weather situation, give full play to different neural network algorithm applicability advantage separately, improve the overall precision of prediction of forecast model under DIFFERENT METEOROLOGICAL CONDITIONS, and the data verification validity of method by experiment, reduce predicated error;

Achieve the real-time estimate of the photovoltaic output power to one day future, and then instruct photovoltaic plant despatching work, guarantee the operation of the grid-connected rear whole stabilization of power grids, safety.

The beneficial effect that the present invention brings is, the present invention is by environment parameter and the taxonomic revision of the generating parameter historical data same period in season 1 year spring, summer, autumn and winter four, reject the data that error is larger, devise a kind of forecast model based on multiple neural network group hop algorithm, this model can be applied to the meteorological weather of difference of four Various Seasonal of a year respectively, and combinational algorithm forecast model is maximized favourable factors and minimized unfavourable ones and played the different applicability advantages of four kinds of neural network algorithms to three kinds of typical weather.With the actual measurement power generation values comparing result of actual light overhead utility, predicting the outcome of this method shows that this method effectively reduces predicated error, improve precision of prediction and reach about 5%, can predict the photovoltaic power station power generation amount of Various Seasonal different weather environmental baseline.Achieve the real-time estimate of the photovoltaic output power to one day future, and then reach the work instructing photovoltaic plant to dispatch.Make in data bulk, data fitting degree and method practicality all guaranteed by the division same period in four seasons.With the grid-connected photovoltaic power station generated energy output power forecasting software platform corresponding to the inventive method, a kind of computer-aided prognosis instrument that can be practical of can yet be regarded as.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the construction method of the photovoltaic power generation quantity Short-term Forecasting Model of the photovoltaic plant based on neural network of the present invention;

Fig. 2 is BP neural network basic structure schematic diagram;

Fig. 3 is the process flow diagram of BP neural network algorithm;

Fig. 4 is Elman neural network basic structure schematic diagram;

Fig. 5 is the process flow diagram of Elman neural network algorithm;

Fig. 6 is RBF neural network basic structure schematic diagram;

Fig. 7 is RBF neural algorithm flow chart;

Fig. 8 is GRNN neural network basic structure schematic diagram;

Fig. 9 is GRNN neural network algorithm process flow diagram;

Figure 10 is weight parameter table.

Embodiment

Embodiment one: present embodiment is described see Fig. 1, the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm described in present embodiment, the detailed process of this construction method is:

Step one: choosing of neural network algorithm;

BP, Elman, RBF and GRNN tetra-kinds of neural network algorithms are adopted to build neural network prediction submodel A, B, C and D respectively, with the environment temperature T of power station every day working time section different time points _i, per day intensity of solar radiation per day wind speed as the input data of photovoltaic power generation quantity Short-term Forecasting Model, to predict the photovoltaic output power P of day corresponding time point _ias the output data of each neural network prediction submodel A, B, C and D;

Step 2: choosing of sample data;

By photovoltaic data acquisition platform, environment parameter and the generating parameter historical data of the fine day same period, cloudy weather and rainy weather is chosen from historical data base, the Weather information same period is carried out sifting sort, reject unusual data point, in conjunction with the Weather information same period, arrange out the sample data under three kinds of weather conditions respectively, three kinds of described weather conditions are respectively fine day, cloudy weather and rainy weather;

Step 3: the training of each neural network prediction submodel, the applied research analysis of each neural network prediction submodel and the structure to photovoltaic power generation quantity Short-term Forecasting Model;

By step 2 gained in conjunction with the Weather information same period, arrange out the sample data under three kinds of weather conditions respectively, respectively neural network submodel A, B, C and D of building are trained, wherein, the input quantity of above-mentioned four neural network submodels A, B, C and D includes 25 variablees, and described 25 variablees are respectively the environment temperature T of prediction day 23 time points _i, per day intensity of solar radiation with per day wind speed described prediction day, 23 time points were in the working time section 7:00 to 18:00 of power station every day, with 7:00 point for start time point, be a time point, totally 23 time points every 30 minutes, after training, draw the neural network prediction submodel A under three kinds of weather conditions ¹, B ¹, C ¹and D ¹,

According to the neural network prediction submodel A under three kinds of weather conditions after training ¹, B ¹, C ¹and D ¹show that BP, Elman, RBF and GRNN tetra-kinds of neural network algorithms are to the applicability conclusion of three kinds of weather conditions by the comparative analysis of root-mean-square error curve and percentage error curve, the weight modulus value parameter of each neural network prediction submodel is provided according to DIFFERENT METEOROLOGICAL CONDITIONS, build modulus value weight parameter table, and then build combination neural net forecast model, using combination neural net forecast model as final photovoltaic power generation quantity Short-term Forecasting Model, complete the structure of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm.

In present embodiment, drawn the neural network prediction submodel A under three kinds of weather conditions after training by prediction correlation curve and predicted percentage graph of errors ¹, B ¹, C ¹and D ¹, draw following applicability conclusion: wherein fine day adopts RBF neural algorithm predicts error minimum; Cloudy weather adopts GRNN Neural Network Prediction effect best; Overcast and rainy predicated error is relatively large by adopting a kind of neural network prediction algorithm to the com-parison and analysis balance of historical data, and wherein relative error is less in the ordinary course of things for RBF neural algorithm.Weight parameter table, specifically see Figure 10.

Embodiment two: the difference of the construction method of present embodiment and the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm described in embodiment one is, the method also comprises step 4: to the correction of photovoltaic power generation quantity Short-term Forecasting Model;

A) first, the sample data in step 2 is normalized,

B) secondly, genetic algorithm and particle cluster algorithm is adopted.

In present embodiment, first, the sample data in step 2 is normalized, reduces the predicated error of predictor model further.Secondly, adopt genetic algorithm and particle cluster algorithm, two kinds of methods are optimized all neural network prediction submodels, neural network algorithm can be avoided easily to be absorbed in the problem of local optimum, further reduce the predicated error of neural network prediction submodel.

In present embodiment, to the environment temperature T of the power station every day working time section different time points gathered _i, per day intensity of solar radiation per day wind speed raw data is screened, and rejects the data point that error is larger, carries out taxonomic revision in conjunction with the Weather information same period by Various Seasonal, weather pattern.Because input variable type is different, be normalized sample data to reduce training error, last output layer to be converted back raw data by renormalization.Data normalization formula:

$Y_i=\frac{X_i-X_{min}}{X_{\max }-X_{min}},$

X _i＝(X _max-X _min)Y _i+X _min，

Wherein, Y _irepresent the value after normalization, X _irepresent the value before normalization, X _minrepresent the minimum value of input amendment, X _maxrepresent the maximal value of input amendment.

The present invention is in concrete application process:

By B/S mode light overhead utility data monitoring platform, by Ethernet photovoltaic data collector, the inverter of collection generating parameter and weather station environment parameter data are uploaded to web data storehouse in real time, achieve generating parameter and the real-time display of environment parameter data, statistical summaries and the function such as history data store, inquiry, user can also log in monitor supervision platform whenever and wherever possible and obtain photovoltaic plant in real time and historical data.

Simultaneously, affect on the basis of the main meteorological environmental factor that photovoltaic plant is exerted oneself in research, in conjunction with 80 groups of valid data samples that the same period, 3 to May of spring in 2014 filtered out by Weather information by fine, cloudy, overcast and rainy three kinds of different weather types are classified, with environment parameter (7 to 18: 23 ambient temperature data points, per day solar irradiance, per day wind speed) as input, 23 time point photovoltaic output powers are as output, adopt BP, Elman, RBF, the neural network algorithm that GRNN tetra-kinds is different, devise the combination neural net photovoltaic power generation quantity short-term forecasting submodel for three kinds of typical weather patterns, and test emulation and predicated error MPAE and RMSE comparative analysis are carried out to predictor model, the weather Adaptability Analysis lacking multiple neural network algorithm compared to Present Domestic in photovoltaic plant exerts oneself short-term forecasting compares with model preferred, predicated error is larger, this model has more raising precision of prediction targetedly for different weather type and reaches about 5%, wherein fine day optimum prediction model M APE is 7.5%, broken sky is 15.7%, overcast and rainy is 27.2%.

Embodiment three: the difference of the construction method of present embodiment and the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm described in embodiment one or two is, the method also comprises step 5: to the assessment of photovoltaic power generation quantity Short-term Forecasting Model;

Mean absolute percentage error MAPE and root-mean-square error RMSE two kinds of error assessment methods are adopted to carry out error evaluation to photovoltaic power generation quantity Short-term Forecasting Model,

$MAPE=\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}\left|\frac{P_p^i-P_a^i}{P_a^i}\right|\×100\%,$

$RMSE=\sqrt{{\frac{\underset{i=1}{\overset{N}{\Σ}}(P_p^i-P_a^i)}{N}}^2},$

Wherein, N represents data total amount, and i is positive integer, represent the predicted value of i-th data point, represent the actual value of i-th data point.

In present embodiment, select effective historical data of Shanghai XIHE5kW exemplary optical overhead utility (comprising different weather type) as sample data, adopt BP, Elman, RBF, GRNN neural network algorithm to build photovoltaic for different weather type combination respectively to exert oneself Short-term Forecasting Model, by repeatedly predictive simulation, obtain optimum prediction comparing result and the predicted percentage error result of three kinds of lower four kinds of neural network algorithms of different weather condition.

Sample raw data is the inverter generating parametric data and weather station environment parameter data every 5min, and sample data is unified selects the time period to be 7:00-18:00, with every day 7:00-18:00 every the environment temperature T of 30min totally 23 time points ₁, T ₂..., T ₂₃, per day intensity of solar radiation per day wind totally 25 variablees input as neural network prediction model, the photovoltaic output power P of corresponding 23 time points ₁, P ₂..., P ₂₃export as neural network prediction submodel.

More than com-parison and analysis three groups of lower four kinds of different Neural Network Prediction situations of different weather condition and corresponding prediction error data:

Fine day four kinds of Neural Network Prediction curves are all comparatively level and smooth, and prediction is fluctuated also smaller with actual measurement percentage error curve.Wherein fine day adopts RBF neural algorithm predicts error minimum; Adopt the MAPE of BP, Elman, GRNN tri-kinds of neural network algorithms to be more or less the same, and BP neural network RMSE is maximum, GRNN is minimum.

For cloudy weather, photovoltaic output power curve has certain fluctuation, and this is relevant with the fluctuation of cloudy weather solar radiation, and all relative fine day of predicated error MAPE and RMSE increases, wherein adopt the prediction effect of GRNN neural network algorithm best, MAPE is 15%; Other three kinds of Neural Network Prediction error MAPE are between 17 to 20%, and BP predicated error RMSE is maximum, Elman and RBF predicated error RMSE is suitable.

For overcast and rainy, overcast and rainy photovoltaic output power compares fine day, cloudy weather fluctuation is comparatively large, and photovoltaic is exerted oneself also less, and this is because overcast and rainy solar radiation intensity is relatively weak; Predicated error is relatively also larger simultaneously, and MAPE is all greater than 25%, and wherein BP Neural Network Prediction error MAPE is greater than 30%; Integrated comparative four kinds of Neural Network Prediction error MAPE and RMSE, RBF neural algorithm predicts error is minimum.

Therefore, the predicated error based on the photovoltaic power station power generation amount Short-term Forecasting Model of multiple neural network group hop algorithm is: fine day optimum is 7.5%, and broken sky is 15.7%, and overcast and rainy is 27.2%.Affect greatly because photovoltaic plant exports by weather environment, different meteorologic factors and special environment condition can make different neural network algorithms produce different error characters when training and predict, by the error character of different neural network algorithm under different weather environmental impact of analyzing and researching, the relevance grade modulus value of these neural network algorithms to Different climate environmental baseline is found out in contrast, set up relevance grade modulus value table, build multiple neural network joint forecast model.Provide a kind of construction method of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network unified algorithm.Affect greatly because photovoltaic plant exports by weather environment, meteorologic factor and special environment condition can make different neural network algorithms produce different errors when training and predict, therefore weather and such environmental effects is taken into full account, at the different neural network algorithm of comparative analysis on the applied research basis of different weather, set up multiple neural network algorithm built-up pattern, there is higher prediction accuracy, predict very important to grid-connected photovoltaic power station generated energy output power.And the multiple neural network ensemble algorithm model set up on this analysis foundation.

Embodiment four: the difference of the construction method of present embodiment and the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm described in embodiment one or two is, neural network prediction submodel A, B, C and D include input layer, output layer and hidden layer

To predict the environment temperature T of day 23 time points _i, per day intensity of solar radiation with per day wind speed totally 25 variablees are as the input layer input data of each neural network prediction submodel A, B, C and D,

To predict the photovoltaic output power P of day corresponding 23 time points _iall as the output data of each neural network prediction submodel A, B, C and D,

The output layer activation function of neural network prediction submodel A, B, C and D all adopts pureline function to realize,

The hidden layer of neural network prediction submodel A, B, C and D all adopts single layer structure, hidden layer neuron node experimental formula is utilized to obtain neural network start node number, adopt method of trial and error, obtaining hidden layer node number is 15, and hidden layer activation function adopts tansig function to realize.

In present embodiment, hidden layer adopts single layer structure, utilizes hidden neuron node experimental formula determine start node number, method of trial and error is adopted successively to increase or reduce interstitial content again, make network error minimum, finally determine node in hidden layer, hidden layer activation function adopts tansig function to realize, wherein, l is hidden layer neuron number, m is input layer number, and n is output layer neuron number, and a is empirical constant between value 1 to 10.

Present embodiment, Fig. 2 is BP neural network basic structure; Input layer has m neuron node, and hidden layer has p neuron node, and output layer has n neuron node, W _ij(i=1,2 ..., m; J=1,2 ..., p) for input layer is to the weights of hidden layer, W _jk(j=1,2 ..., p; K=1,2 ..., n) for hidden layer is to the weights of output layer, θ _j(j=1,2 ..., p) be the threshold values of hidden layer, α _k(k=1,2 ..., n) be output layer threshold values, (X ₁, X ₂... X _m) be input vector, (Y ₁, Y ₂..., Y _n) be neural network output vector, (Y _h1, Y _h2..., Y _hn) be desired output, (e ₁, e ₂, e ₃..., e _n) be neural network desired output and the actual error exported.

Input layer input vector is X _i, hidden layer activation function is f ₁, output layer activation function is f ₂, then a hidden layer jth neuronic output Z _jwith an output layer kth neuronic output Y _kbe respectively:

$Z_j=f_1(\underset{i=1}{\overset{m}{\Σ}}{W}_{ij}{X}_i+{\mathrm{\θ}}_j)$

$Y_k=f_2(\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{z}_j{w}_{jk}+{\mathrm{\α}}_k)=f_2\left(\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{f}_1\right(\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{W}_{ij}{X}_i+{\mathrm{\θ}}_j)W_{jk}+{\mathrm{\α}}_k)$

If BP neural network exports Y _kwith desired output Y _hkthere is error, error along the backpropagation of BP neural network, and constantly can revise each layer weights of neural network, threshold values, the stopping iteration until error meets the demands.

The idiographic flow of BP neural network algorithm is see Fig. 3; Elman neural network basic structure is see Fig. 4; The flow process of Elman neural network algorithm is specifically see Fig. 5; Elman neural network meets following formula:

x _c(k)＝x(k-1)

x(k)＝f(w ¹x _c(k)+w ²u(k)+a)

y(k)＝g(w ³x(k)+b)

Wherein, u is that r ties up input vector, and x is that n ties up hidden layer node output vector, and y is that m ties up output vector, x _cfor n ties up feedback vector, w ¹for accepting the connection weights of layer to hidden layer, w ²for input layer is to the connection weights of hidden layer, w ³for hidden layer is to the connection weights of output layer, f and g is respectively hidden layer and output layer activation function, x _ck () represents that n ties up feedback vector, x (k) represents that n ties up hidden layer node output vector, and y (k) is expressed as m and ties up output vector.

Elman neural network adopts BP neural network error function to carry out reverse modified weight, and namely adopt sum of squared errors function oppositely to revise each layer of Elman neural network as judgment basis and connect weights, threshold values, sum of squared errors function is as follows:

$E=\underset{k=1}{\overset{N}{\Σ}}{\left(y_k-{\hat{y}}_k\right)}^2,$

Wherein, E represents sum of squared errors function, y _krepresent neural network real output value, represent desired output.

Be similar to BP neural network algorithm, first to be normalized input amendment data, weights and threshold values is connected between each layer of initialization network, then each layer of neural network is carried out, difference is that Elman neural network hidden layer output valve needs to feed back to hidden layer input end, then according to the output connecting weights and corresponding threshold values and activation function computational grid between each layer through accepting layer one step time delay operator.

RBF neural network basic structure; Adopt three-decker, specifically see Fig. 6, wherein input layer is made up of m neuron node, and hidden layer is made up of h neuron node, and output layer is made up of n neuron node.X=(x ₁, x ₂..., x _m) ^tfor input vector, y=(y ₁, y ₂..., y _n) ^tfor output vector, hidden layer activation function φ _iadopt Gaussian function:

${\mathrm{\φ}}_i\left(x\right)=\exp \[-\frac{{(x-c_i)}^2}{2{\mathrm{\σ}}^2}\]$

In formula, c _i=[c ₁, c ₂..., c _h] ^tfor Gaussian function center vector, σ is corresponding variance.|| * || represent European norm, ∑ represents that output layer neuron adopts linear weighted function, W _ijrepresent the connection weights of input layer to hidden layer, W _jkrepresent the connection weights of hidden layer to output layer.The output of output layer kth can be expressed as:

$y_k=\underset{j=1}{\overset{h}{\Σ}}{W}_{jk}{\mathrm{\φ}}_j\left(\right||x-c_j|\left|\right)$

Fig. 7 is RBF neural algorithm flow chart.

GRNN neural network basic structure, specifically see Fig. 8, wherein input layer number is m, X=(x ₁, x ₂..., x _m) ^tfor input vector, mode layer neuron number equals input layer number, and wherein mode layer i-th neuron exports and be:

$P_i=\exp \[-\frac{{(X-x_i)}^T(X-x_i)}{2{\mathrm{\σ}}^2}\],i=1,2,\mathrm{...},m,$

Wherein, X=(x ₁, x ₂..., x _m) ^tfor input vector, x _ibe i-th neuron input amendment value, σ represents variance.

Summation layer uses two kinds of computing formula to carry out summation operation, and a class is arithmetic summation, meets following expression:

$S_D=\underset{i=1}{\overset{n}{\Σ}}{P}_i=\underset{i=1}{\overset{n}{\Σ}}\exp \[-\frac{{(X-x_i)}^T(X-x_i)}{2{\mathrm{\σ}}^2}\],$

Another kind of is weighted sum, and namely in mode layer, i-th neuron is i-th output neuron y with being connected weights between a summation layer jth neuron _ia middle jth element, meets following formula:

$S_{Nj}=\underset{i=1}{\overset{m}{\Σ}}{y}_{ij}{P}_i,j=1,2,\mathrm{...},n,$

S _njrepresent weighted sum, S _drepresent arithmetic summation, output layer has n neuron node, and two class summation layer neuron node export to be divided by and can obtain the output of output layer neuron node, and an output layer jth neuron node exports y _jfor:

$y_j=\frac{S_{Nj}}{S_D},j=1,2,\mathrm{...},n.$

GRNN neural network algorithm process flow diagram, specifically see Fig. 9.

Embodiment five: the difference of the construction method of present embodiment and the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm described in embodiment one or two is, the method also comprises step 6: the foundation of software platform;

Design a kind of for practical grid-connected photovoltaic power station generated energy output power forecasting software platform, as a kind of utility of actual prediction.

Claims (4)

1. based on the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model of multiple neural network group hop algorithm, it is characterized in that, the detailed process of this construction method is:

Step one: choosing of neural network algorithm;

BP, Elman, RBF and GRNN tetra-kinds of neural network algorithms are adopted to build neural network prediction submodel A, B, C and D respectively, with the environment temperature T of power station every day working time section different time points _i, per day intensity of solar radiation per day wind speed as the input data of photovoltaic power generation quantity Short-term Forecasting Model, to predict the photovoltaic output power P of day corresponding time point _ias the output data of each neural network prediction submodel A, B, C and D;

Step 2: choosing of sample data;

By photovoltaic data acquisition platform, environment parameter and the generating parameter historical data of the fine day same period, cloudy weather and rainy weather is chosen from historical data base, the Weather information same period is carried out sifting sort, reject unusual data point, in conjunction with the Weather information same period, arrange out the sample data under three kinds of weather conditions respectively, three kinds of described weather conditions are respectively fine day, cloudy weather and rainy weather;

Step 3: the training of each neural network prediction submodel, the applied research analysis of each neural network prediction submodel and the structure to photovoltaic power generation quantity Short-term Forecasting Model;

By step 2 gained in conjunction with the Weather information same period, arrange out the sample data under three kinds of weather conditions respectively, respectively neural network submodel A, B, C and D of building are trained, wherein, the input quantity of above-mentioned four neural network submodels A, B, C and D includes 25 variablees, and described 25 variablees are respectively the environment temperature T of prediction day 23 time points _i, per day intensity of solar radiation with per day wind speed described prediction day, 23 time points were in the working time section 7:00 to 18:00 of power station every day, with 7:00 point for start time point, be a time point, totally 23 time points every 30 minutes, after training, draw the neural network prediction submodel A under three kinds of weather conditions ¹, B ¹, C ¹and D ¹,

According to the neural network prediction submodel A under three kinds of weather conditions after training ¹, B ¹, C ¹and D ¹show that BP, Elman, RBF and GRNN tetra-kinds of neural network algorithms are to the applicability conclusion of three kinds of weather conditions by the comparative analysis of root-mean-square error curve and percentage error curve, the weight modulus value parameter of each neural network prediction submodel is provided according to DIFFERENT METEOROLOGICAL CONDITIONS, build modulus value weight parameter table, and then build combination neural net forecast model, using combination neural net forecast model as final photovoltaic power generation quantity Short-term Forecasting Model, complete the structure of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm.

2. the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm according to claim 1, it is characterized in that, the method also comprises step 4: to the correction of photovoltaic power generation quantity Short-term Forecasting Model;

A) first, the sample data in step 2 is normalized,

B) secondly, genetic algorithm and particle cluster algorithm is adopted.

3. the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm according to claim 1 and 2, it is characterized in that, the method also comprises step 5: to the assessment of photovoltaic power generation quantity Short-term Forecasting Model;

Mean absolute percentage error MAPE and root-mean-square error RMSE two kinds of error assessment methods are adopted to carry out error evaluation to photovoltaic power generation quantity Short-term Forecasting Model,

$MAPE=\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}\left|\frac{P_p^i-P_a^i}{P_a^i}\right|\×100\%,$

$RMSE=\sqrt{\frac{\underset{i=1}{\overset{N}{\Σ}}{(P_p^i-P_a^i)}^2}{N}},$

Wherein, N represents data total amount, and i is positive integer, represent the predicted value of i-th data point, represent the actual value of i-th data point.

4. the construction method of the photovoltaic power station power generation amount Short-term Forecasting Model based on multiple neural network group hop algorithm according to claims 1 or 2, it is characterized in that, neural network prediction submodel A, B, C and D include input layer, output layer and hidden layer,

To predict the environment temperature T of day 23 time points _i, per day intensity of solar radiation with per day wind speed totally 25 variablees are as the input layer input data of each neural network prediction submodel A, B, C and D,

To predict the photovoltaic output power P of day corresponding 23 time points _iall as the output data of each neural network prediction submodel A, B, C and D,

The output layer activation function of neural network prediction submodel A, B, C and D all adopts pureline function to realize,

The hidden layer of neural network prediction submodel A, B, C and D all adopts single layer structure, hidden layer neuron node experimental formula is utilized to obtain neural network start node number, adopt method of trial and error, obtaining hidden layer node number is 15, and hidden layer activation function adopts tansig function to realize.