CN116227677A - Power prediction correction method considering photovoltaic power climbing characteristics - Google Patents

Abstract

The invention belongs to the field of new energy grid-connected consumption, and discloses a power prediction correction method considering the climbing characteristic of photovoltaic power, which comprises the following steps: acquiring historical photovoltaic power data and meteorological factors, preprocessing, and extracting photovoltaic power climbing characteristics by using a climbing characteristic calculation formula; building an EOF-OPTICS clustering model based on the climbing characteristic to cluster the historical photovoltaic power generation power; carrying out photovoltaic power prediction on the clustering subset by using a RMSprop-QWEGRU photovoltaic prediction model; calculating a prediction error of a climbing section of the photovoltaic power generation power; and predicting future photovoltaic power by using a RMSprop-QWEGRU photovoltaic prediction model, mapping historical photovoltaic prediction errors to future photovoltaic predictions by using a shape DTW similarity measurement method, and carrying out error correction of corresponding time periods. The method can improve the accuracy of photovoltaic power climbing section prediction and provide technical support for stable operation of the power grid.

Description

Power prediction correction method considering photovoltaic power climbing characteristics

Technical Field

The invention belongs to the field of new energy grid-connected consumption, and particularly relates to a power prediction correction method considering the climbing characteristic of photovoltaic power.

Background

With the rapid development of economy, the phenomenon of energy shortage, environmental deterioration and the like is generally presented worldwide. In this background, the development of new energy sources represented by photovoltaics has been coming into the spotlight. The photovoltaic climbing phenomenon means that when the weather changes due to day-night conversion and short-term environment, the photovoltaic output power rises or falls greatly in a short time. Along with the continuous increase of the photovoltaic grid-connected capacity, the influence of the photovoltaic climbing phenomenon on the stability of the power grid and the quality of output electric energy is larger and larger, so that the accurate prediction of the photovoltaic power climbing phenomenon has important significance on the stability and the safe operation of the power grid.

At present, the photovoltaic climbing power prediction method is mainly divided into a direct prediction method and an indirect prediction method. The method for directly predicting the climbing event carries out the climbing prediction through the environmental variable, and is visual and has a certain accuracy. However, as the photovoltaic power climbing event is a small probability event, the number and quality of training samples of the direct prediction method model are often limited, and the prediction effect of the model is also often poor. While the main stream indirect method can improve the prediction accuracy by improving the prediction model and the prediction method, the indirect method has the problems of error accumulation, climbing, missing report and the like due to the introduction of power prediction. Therefore, how to improve the prediction model and the method has important significance for improving the effect of the photovoltaic power generation power climbing prediction.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a power prediction correction method considering the characteristics of photovoltaic climbing, which aims to improve the accuracy of predicting the photovoltaic power climbing section by carrying out characteristic analysis on historical photovoltaic climbing data and feeding back a prediction error to a power prediction model according to the characteristics, and provides technical support for reducing the adverse effect of photovoltaic climbing on a power grid.

In order to solve the technical problems, the invention provides the following technical scheme:

a power prediction correction method taking into account characteristics of a photovoltaic power ramp, the method comprising the steps of:

step 1, acquiring historical photovoltaic power data, and processing the original data into a standard data format;

the processed data comprise historical photovoltaic power and air condition influence factors, power climbing characteristics are extracted according to a climbing characteristic calculation formula optimized by utilizing the historical photovoltaic power, and the climbing characteristics comprise climbing points P _i ' amplitude of power change ΔP _r (t), a power change rate R (t), a climbing occurrence probability O (t);

step 2, an EOF-OPTICS clustering model is established to cluster the climbing characteristics of the photovoltaic power, and a subset under different climbing characteristics is obtained;

step 3, training and predicting the divided subsets by using an RMSprop-QWEGRU photovoltaic prediction model to obtain power prediction results of all the sub-clusters;

step 4, calculating the errors of the predicted value and the true value under each subgroup based on the training result of each cluster, and calculating an error rate time sequence by using the predicted result and the true value;

and 5, carrying out error correction on the photovoltaic power generation power prediction result by using the prediction error rate.

Further, in the step 1, the following climbing features are extracted using a climbing feature calculation formula: amplitude of change in power ΔP _r (t) wherein ΔP _r (t) > 0 is climbing up, ΔP _r (t) < 0 is a downhill climb, a power change rate R (t), a climbing occurrence probability O (t),

ΔP _r (t)＝P _i+Δi '[t,Δt]-P _i '[t,Δt]＞P _ε

P _ε ＝ηP _d

wherein: p (P) _i ' is the climbing point, through

Determining, P _t For the power value of the climbing point, i is a data point in delta t time, delta i is a time interval of the data point, t is the climbing starting time, delta t is the climbing duration, and P _ε To be climbing threshold value, P _d For the photovoltaic installed capacity, eta is the climbing threshold value to occupy the percentage of the installed capacity, and the optimized parameter combination < omega, lambda > is the sigmoid function linearity and the threshold value.

Further, in the step 2, the step of reconstructing the characteristic space of the photovoltaic power climbing characteristic by using the empirical orthogonal function EOF specifically includes: normalizing the photovoltaic power data characteristics, and constructing a matrix X from the processed data _m×n Wherein m is a feature dimension and n is a time series; solving principal components of original photovoltaic data space eigenvectors

Obtaining the main component of the characteristic vector of the original photovoltaic power generation data, wherein +.>

Is the transpose of the eigenvector matrix; solving the variance contribution rate of the kth eigenvector>

Wherein lambda is a characteristic value sequence (lambda ₁ ,…,λ _m ) And (lambda) ₁ ＞λ ₂ ＞…＞λ _m ) By sequentially applying the first n items K _i Summing to obtain a feature vector space which can best embody the original feature relation, and reconstructing a feature matrix to obtain a photovoltaic sample set R _o×n Where o is the most capable of representing the originalFeature numbers of the feature relationships;

the method for establishing the OPTICS cluster model based on the reconstructed feature matrix with higher correlation coefficient obtained by the EOF algorithm comprises the following specific steps: first calculate a set of photovoltaic samples R _o×n Core distance of each sample point:

calculating the reachable distance between the core sample and the neighbor point y of the sample:

wherein y, r.epsilon.R { R ₁ ,r ₂ ,…,r _m -a }; epsilon represents the input parameter neighborhood; minPts represents the number of input judgment core object samples; n (N) _ε (R) is the number of sample sets in the sample set R having a distance from the sample point R of not more than ε;

for representing the set N _ε An ith node of (r) closest to node r; the distances among the samples are Euclidean metrics according to the characteristics;

traversing the sample set R to obtain a core photovoltaic data sample object set omega, processing neighbor points in sequence from small to large according to the reachable distance of the core objects, and if the rest core objects exist in the neighbor point set, preferentially processing the core objects and the neighbor points of the core objects; the result is an ordered list { p } of the photovoltaic sample set ₁ ,p ₂ ,…,p _N P, where _i The node sequence number is the node sequence number processed for the ith time; core distance { c of sample node ₁ ,c ₂ ,…,c _N -a }; reachable distance { rd for sample nodes ₁ ,rd ₂ ,…,rd _N Finally, clustering is completed according to the ordered list set P to obtain a subset { S }, wherein the subset { S } ₁ ,S ₂ ,…,S _m }。

Further, the saidIn step 3, an empirical orthogonal function EOF is used to reconstruct a feature space for features with higher correlation coefficients, and the specific steps include: the original photovoltaic data is randomly divided into a training set and a testing set after normalization, and a RMSprop-QWGRU photovoltaic prediction neural network is trained, wherein the QWGRU is a quantum weighting gating circulating unit, and the QWGRU inputs photovoltaic signs of the testing set

The output layer is photovoltaic power generation power +.>

The hidden layer is

The neuron parameters comprise an input x _m Weights |phi of (2) _m >Neuronal Quantum Activity value->

The relation between the aggregation operator sigma, the activation function F and the excitation function F, and the input photovoltaic sign matrix and the output predicted photovoltaic power generation power is expressed as follows:

wherein, the F activation function is inner product operation; phi _m >＝[cosα _m ,sinα _m ] ^T ，α _m Is of the value of phi _m >Is a phase of (2);

beta is->

Is a phase of (2); alpha _i Is of the value of phi _i >Is a phase of (2); x is x _i Is the input of neurons; m is the neuron input number;

the QWGUR parameters are adjusted through an RMSprop self-adaptive learning rate optimization algorithm, and a loss function is defined:

wherein o is _i As predicted value, y _i Is a true value, and N is a group of data number;

through type son

Updating parameters of neuron, in which Eg ² ] _r ＝0.9E[g ² ] _t-1 +0.1g _t ⊙g _t Meaning a gradient squared weighted average and as a gradient scaling factor; the product of Hadamard is expressed as the multiplication of the corresponding position elements of the matrix; gamma is 1e-6; gradient->

Meaning that the loss function L (f (x; θ) _t ) Y) about theta _t Is a derivative of (2); θ _t To the neuron parameters that need to be optimized; η is the learning rate.

Further, in the step 5, a time series correspondence f is calculated _Tsc The method comprises the following specific steps: s is selected by using shape DTW similarity measurement method _i Representative elements in the subset

Calculation of +.about.using shape DTW method>

Time-series correspondence f with power prediction curve FCST (t) _Tsc The method comprises the steps of carrying out a first treatment on the surface of the The ShapeDTW measures the similarity between photovoltaic power curves, firstly, the euclidean distance is used for measuring the similarity between sequence shape descriptors d, so that the corresponding relation between two photovoltaic power time sequences is obtained, and Ji Guangfu power time sequences are realized; where d= (d ₁ ,d ₂ ,…d _L ) ^T ,d _i ∈R ^m ，d＝F(S)＝(F(s ₁ ),F(s ₂ ),…,F(s _L )) ^T The shape descriptor computation function F takes the form of a derivative, i.e. for the raw power data s _i Taking a first derivative with respect to time;

error rate sigma of prediction of subset _i The sequence is according to the corresponding relation f _Tsc Mapping to a target time-of-day sequence, comprising the following specific steps: determining a climbing time period by comparing the difference value of the power of adjacent climbing points of the power prediction curve FCST (t) in the delta t time with the set threshold delta, and recording the climbing starting time t _begin Climbing end time t _end The method comprises the steps of carrying out a first treatment on the surface of the FCST (t) and representative element prediction by photovoltaic generation power

Time series correspondence f between _Tsc Calculation of the photovoltaic power prediction bias σ (t) =0.15 σ, eventually used to correct the moment t, using a weighted moving average error _i (f _Tsc (t)-1)+0.7σ _i (f _Tsc (t))+0.15σ _i (f _Tsc (t) +1), where t ε (t) _begin ,t _end ) The method comprises the steps of carrying out a first treatment on the surface of the Performing error correction on FCST (t), and finally predicting result +.>

Compared with the prior art, the invention has the beneficial effects that:

1) According to the invention, an optimized climbing characteristic calculation formula is established, the photovoltaic power climbing characteristic calculation accuracy is improved, and the problem of missing detection of the photovoltaic power climbing characteristic is solved;

2) According to the invention, an EOF-OPTICS clustering method is used for clustering historical photovoltaic data based on climbing features and photovoltaic features, an EOF algorithm can extract the features with the maximum correlation coefficient with photovoltaic power climbing, and the OPTICS method is classified based on density, wherein clusters with smaller density can be ignored, uncertainty factors influencing photovoltaic power climbing can be further reduced, and clustering results can better reflect the climbing features of the historical data;

3) According to the method, an RMSprop-QWEGRU prediction model is established, historical photovoltaic data is used for training and prediction, a prediction error rate is fed back to a prediction result according to sequence feature similarity, error correction is carried out on a photovoltaic power climbing section with high instability and difficult predictability, the accuracy of photovoltaic power prediction is further improved, and the problem that photovoltaic power climbing is difficult to predict is solved;

4) The invention uses a shape DTW method to calculate the similarity and time sequence corresponding relation between the sub-cluster representative data and the predicted target, maps the predicted climbing period of the target photovoltaic power to the power prediction error rate of the historical data subset according to the time point of the time sequence corresponding relation, and takes the weighted moving average as the final correction error rate; the method can effectively reduce the influence of noise points of the photovoltaic historical data, and can perform error rate feedback more accurately compared with a direct correspondence method.

Drawings

Fig. 1 is a flowchart of an overall power prediction correction method considering the climbing characteristic of photovoltaic power.

Detailed Description

The invention is further described below with reference to the accompanying drawings.

As shown in fig. 1, a power prediction correction method considering the climbing characteristic of photovoltaic power includes the following steps:

step 1, acquiring historical photovoltaic power generation power data and environmental impact factors, wherein a data set comprises photovoltaic power generation power, solar irradiation intensity, environmental temperature, environmental humidity, wind speed, weather conditions and other weather impact factors at corresponding moments, carrying out quantization treatment on the weather conditions (0 for rainy days, 0.5 for cloudy days and 1 for sunny days), deleting abnormal data and non-power output time period data, and finally obtaining the data set: historical photovoltaic Power { P (t) } _{(t＝1,2,3…T)} P (t) is the photovoltaic power at time t; n=5 weather influencing factors { F _n (t)} _{(t＝1,2,3…T)(n＝1,2,3…N)} ，F _n And (t) the nth type environmental impact factor data at the moment t.

Extracting the following photovoltaic climbing characteristics by using an optimized climbing characteristic calculation formula: amplitude of change in power ΔP _r (t) whereinΔP _r (t) > 0 is climbing up, ΔP _r (t) < 0 is downhill climbing; a power change rate R (t); the probability of occurrence of climbing O (t),

ΔP _r (t)＝P _i+Δi '[t,Δt]-P _i '[t,Δt]＞P _ε

P _ε ＝ηP _d

wherein: p (P) _i ' is the climbing point, i is the data point in delta t time, delta i is the time interval of the data point, t is the climbing starting time, delta t is the climbing duration, the value is the difference of the time greater than the threshold power change period, and P _ε To be climbing threshold value, P _d For the photovoltaic installed capacity, eta is the percentage of the installed capacity occupied by the climbing threshold value, and the parameter combination is optimized<ω,λ>The linearity of the sigmoid function and the threshold installed capacity coefficient are respectively determined by an Adam parameter optimizing method, and in the embodiment, delta i is 1h and P _d 5kW, 10% lambda and 0.01% omega;

and 2, establishing an EFO-OPTICS clustering model according to the extracted climbing characteristics to cluster the photovoltaic power generation of each time node, so as to obtain subsets under different climbing characteristics. The EFO-OPTICS clustering model is a model based on density clustering of a data set based on data characteristics processed by an empirical orthogonal function EOF algorithm. Firstly, reconstructing a component feature space for features with higher correlation coefficients by using an empirical orthogonal function EOF: normalizing the photovoltaic power data characteristics, and constructing a matrix X from the processed data _m×n Wherein m is a feature dimension, n is a time sequence, and X is calculated _m×n Covariance of (2)Matrix array

Eigenvalue (lambda) ₁ ,…,λ _m ) And (lambda) ₁ ＞λ ₂ ＞…＞λ _m ) And eigenvector matrix V _m×n . Solving->

Obtaining the main component of the characteristic vector of the original photovoltaic power generation data, wherein +.>

Is the transpose of the eigenvector matrix; solving the variance contribution rate of the kth eigenvector>

For the first n items K in turn _i Summing, when K _i Finishing accumulation when the accumulated value is more than 90%, extracting the characteristics participating in accumulation, and reconstructing a characteristic matrix R _o×n Where o is the feature number that best embodies the original feature relationship.

Based on the reconstructed feature matrix R _o×n And establishing an OPTICS clustering model, and dividing the original data set into n classes based on the characteristics and m outliers. Calculating a sample set R _o×n Core distance of each sample point:

calculating the reachable distance of the core sample:

wherein y, r.epsilon.R { R ₁ ,r ₂ ,…,r _m -a }; epsilon represents the neighborhood of the input parameters, the range is (0, o); minPts represents the number of input judgment core object samples, and is set according to the epsilon value, and the range is (0, n-1); n (N) _ε (R) is the distance from the sample point R in the sample set RA number of sample sets no greater than ε apart;

for representing the set N _ε An ith node of (r) closest to node r; d () is the euclidean metric distance between samples according to the features.

Firstly traversing a sample set R to obtain a core photovoltaic data sample object set omega, processing neighbor points in sequence from small to large according to the reachable distance of the core objects, and if the rest core objects exist in the neighbor point set, preferentially processing the core objects and the neighbor points of the core objects. Obtaining an ordered list set P= { P of the sample set ₁ ,p ₂ ,…,p _N P, where _i The node sequence number is the node sequence number processed for the ith time; core distance { c of sample node ₁ ,c ₂ ,…,c _N -a }; reachable distance { rd for sample nodes ₁ ,rd ₂ ,…,rd _N Finally, clustering is completed according to the ordered list set P to obtain a subset { S }, wherein the subset { S } ₁ ,S ₂ ,…,S _m }。

And 3, respectively training and predicting the divided subsets by using an RMSprop-QWEGRU photovoltaic prediction model to obtain power prediction results of all the sub-clusters. The RMSprop-QWGRU photovoltaic prediction model is a quantum weighted gating circulating unit neural network, an input signal quantum weighting processing mechanism is introduced into the neural network on the basis of a GUR gating circulating unit, and a RMSprop optimization algorithm is used for optimizing quantum neural network parameters. The original photovoltaic data is randomly divided into a training set and a testing set according to subsets after normalization, and the QWERU photovoltaic prediction neural network inputs the photovoltaic signs of the testing set

The output layer is photovoltaic power generation power +.>

The hidden layer is

Neuron parameters include inputsx _m Weights |phi of (2) _m >Neuronal Quantum Activity value->

The relation between the aggregation operator sigma, the activation function F and the excitation function F, and the input photovoltaic sign matrix and the output predicted photovoltaic power generation power is expressed as follows:

wherein, the F activation function is inner product operation; phi _m >＝[cosα _m ,sinα _m ] ^T ，α _m Is of the value of phi _m >Is a phase of (2);

beta is->

Is a phase of (a) of (b).

RMSprop defines a loss function for the adaptive learning rate parameter optimization algorithm:

wherein o is _i As predicted value, y _i For a true value, N is a set of data numbers, which may be set to 30, 50, 100, etc. according to the original data.

Through type son

Updating parameters of neuron, in which Eg ² ] _t ＝0.9E[g ² ] _t-1 +0.1g _t ⊙g _t Meaning a gradient squared weighted average and as a gradient scaling factor; gamma is 1e-6; gradient of

Meaning loss ofLoss function L (f (x; θ) _t ) Y) about theta _t Is a derivative of (2); θ _t To the neuron parameters that need to be optimized; η is the learning rate.

And 4, calculating errors of the predicted value and the true value of the subgroup based on training results of the clusters. The subset of partitions { S } is divided into subsets using a trained RMSprop-QWEGRU photovoltaic prediction model ₁ ,S ₂ ,…,S _m Power prediction for test set in subset i prediction result

n is the number of samples contained in subset i. The predicted value and the true value

Comparing and calculating error rate time series +.>

Sequentially solving for sigma for the subset to obtain { sigma ] ₁ (t),σ ₂ (t),…,σ _m (t)}。

And step 5, carrying out error correction on the photovoltaic power generation power prediction result by using the prediction error rate. Adopting a RMSprop-QWEGRU photovoltaic power prediction model to predict future power value as { FCST (t) } based on original historical data _{(t＝1,2,3…T)} Extracting climbing features by utilizing an optimized climbing feature formula, and classifying a prediction target into a subset sequence { S } by using an EFO-DBSCAN cluster model based on the extracted climbing features ₁ ,S ₂ ,…,S _m In the }, classify the subset S _i S is selected by using shape DTW similarity measurement method _i Representative elements in the subset

Calculation of +.about.using shape DTW method>

A time-series correspondence f between FCST (t), and the prediction error rate sigma of the subset _i The sequence is mapped to a target time sequence of day according to the corresponding relation f, and the climbing time period is calculatedError correction is carried out on the predicted value, and finally a photovoltaic power predicted result { R (t) } is obtained _{(t＝1,2,3…T)} 。

ShapeDTW measures the similarity between the sequence shape descriptors d by using euclidean distances, thereby deriving the correspondence between the two photovoltaic power time series, and realizing the Ji Guangfu power time series. Where d= (d ₁ ,d ₂ ,…d _L ) ^T ,d _i ∈R ^m ，d＝F(S)＝(F(s ₁ ),F(s ₂ ),…,F(s _L )) ^T The shape descriptor computation function F takes the form of a derivative, i.e. for the raw power data s _i The first derivative is derived with respect to time.

Determining a climbing time period by comparing the difference value of the power of adjacent climbing points of the power prediction curve FCST (t) in the delta t time with the set threshold delta, and recording the climbing starting time t _begin Climbing end time t _end . FCST (t) and representative element by power prediction

The time series correspondence relation f is finally used for correcting the photovoltaic power prediction deviation sigma (t) =0.15 sigma at the moment t by using weighted moving average error calculation _i (f(t)-1)+0.7σ _i (f(t))+0.15σ _i (f (t) +1), where t.epsilon.t _begin ,t _end ). For FCST _(t) Error correction is carried out to obtain a final prediction result +.>

Finally, it should be noted that this patent is a power prediction correction method taking into account the characteristics of the photovoltaic climbing, and the above list is only one specific embodiment of the present invention, and obviously, the present invention is not limited to the above examples, but many variations are possible. All modifications directly derived or suggested to one skilled in the art from the present disclosure should be considered as being within the scope of the present invention.

Claims (5)

1. A power prediction correction method taking into account the characteristics of a photovoltaic power ramp, the method comprising the steps of:

step 1, acquiring historical photovoltaic power data, and processing the original data into a standard data format;

the processed data comprise historical photovoltaic power and air condition influence factors, power climbing characteristics are extracted according to a climbing characteristic calculation formula optimized by utilizing the historical photovoltaic power, and the climbing characteristics comprise climbing points P _i ' amplitude of power change ΔP _r (t), a power change rate R (t), a climbing occurrence probability O (t);

step 2, an EOF-OPTICS clustering model is established to cluster the climbing characteristics of the photovoltaic power, and a subset under different climbing characteristics is obtained;

step 3, training and predicting the divided subsets by using an RMSprop-QWEGRU photovoltaic prediction model to obtain power prediction results of all the sub-clusters;

step 4, calculating the errors of the predicted value and the true value under each subgroup based on the training result of each cluster, and calculating an error rate time sequence by using the predicted result and the true value;

and 5, carrying out error correction on the photovoltaic power generation power prediction result by using the prediction error rate.

2. The power prediction correction method considering the climbing characteristics of the photovoltaic power according to claim 1, wherein in the step 1, the following climbing characteristics are extracted using a climbing characteristic calculation formula: amplitude of change in power ΔP _r (t) wherein ΔP _r (t) > 0 is climbing up, ΔP _r (t) < 0 is a downhill climb, a power change rate R (t), a climbing occurrence probability O (t),

ΔP _r (t)＝P _i+Δi '[t,Δt]-P _i '[t,Δt]＞P _ε

P _ε ＝ηP _d

wherein: p (P) _i ' is the climbing point, pass P _i ':

Determining, P _t For the power value of the climbing point, i is a data point in delta t time, delta i is a time interval of the data point, t is the climbing starting time, delta t is the climbing duration, and P _ε To be climbing threshold value, P _d For the photovoltaic installed capacity, eta is the percentage of the installed capacity occupied by the climbing threshold value, and the parameter combination is optimized<ω,λ>The sigmoid function linearity and the threshold degree, respectively.

3. The power prediction correction method considering the climbing characteristic of the photovoltaic power according to claim 1, wherein in the step 2, the characteristic space is reconstructed by using the empirical orthogonal function EOF to reconstruct the climbing characteristic of the photovoltaic power, and the specific steps include: normalizing the photovoltaic power data characteristics, and constructing a matrix X from the processed data _m×n Wherein m is a feature dimension and n is a time series; solving principal components of original photovoltaic data space eigenvectors

Obtaining the main component of the characteristic vector of the original photovoltaic power generation data, wherein +.>

Is the transpose of the eigenvector matrix; solving the variance contribution rate of the kth eigenvector>

Wherein lambda is a characteristic value sequence (lambda ₁ ,…,λ _m ) And (lambda) ₁ ＞λ ₂ ＞…＞λ _m ) By sequentially applying the first n items K _i Summing to obtain a feature vector space which can best embody the original feature relation, and reconstructing a feature matrix to obtain a photovoltaic sample set R _o×n Wherein o is a feature number which can best embody the original feature relation;

the method for establishing the OPTICS clustering model based on the reconstructed feature matrix obtained by the EOF algorithm comprises the following specific steps: first calculate a set of photovoltaic samples R _o×n Core distance of each sample point:

calculating the reachable distance between the core sample and the neighbor point y of the sample:

wherein y, r.epsilon.R { R ₁ ,r ₂ ,…,r _m -a }; epsilon represents the input parameter neighborhood; minPts represents the number of input judgment core object samples; n (N) _ε (R) is the number of sample sets in the sample set R having a distance from the sample point R of not more than ε;

for representing the set N _ε An ith node of (r) closest to node r; the distances among the samples are Euclidean metrics according to the characteristics;

traversing the sample set R to obtain a core photovoltaic data sample object set omega, processing neighbor points in sequence from small to large according to the reachable distance of the core objects, and if the rest core objects exist in the neighbor point set, preferentially processing the core objects and the neighbor points of the core objects; the result is an ordered list { p } of the photovoltaic sample set ₁ ,p ₂ ,…,p _N P, where _i For the ith timeA processed node sequence number; core distance { c of sample node ₁ ,c ₂ ,…,c _N -a }; reachable distance { rd for sample nodes ₁ ,rd ₂ ,…,rd _N Finally, clustering is completed according to the ordered list set P to obtain a subset { S }, wherein the subset { S } ₁ ,S ₂ ,…,S _m }。

4. The power prediction correction method considering the climbing characteristic of the photovoltaic power according to claim 1, wherein in the step 3, the feature space is formed by reconstructing the feature by using an empirical orthogonal function EOF, and the specific steps include: the original photovoltaic data is randomly divided into a training set and a testing set after normalization, and a RMSprop-QWGRU photovoltaic prediction neural network is trained, wherein the QWGRU is a quantum weighting gating circulating unit, and the QWGRU inputs photovoltaic signs of the testing set

The output layer is photovoltaic power generation power +.>

The hidden layer is->

The neuron parameters comprise an input x _m Weights |phi of (2) _m >Neuronal Quantum Activity value->

The relation between the aggregation operator sigma, the activation function F and the excitation function F, and the input photovoltaic sign matrix and the output predicted photovoltaic power generation power is expressed as follows:

wherein, the F activation function is inner product operation; phi _m >＝[cosα _m ,sinα _m ] ^T ，α _m Is of the value of phi _m >Is a phase of (2);

beta is->

Is a phase of (2); alpha _i Is of the value of phi _i >Is a phase of (2); x is x _i Is the input of neurons; m is the neuron input number;

the QWGUR parameters are adjusted through an RMSprop self-adaptive learning rate optimization algorithm, and a loss function is defined:

wherein o is _i As predicted value, y _i Is a true value, and N is a group of data number;

through type son

Updating parameters of neuron, in which Eg ² ] _t ＝0.9E[g ² ] _t-1 +0.1g _t ⊙g _t Meaning a gradient squared weighted average and as a gradient scaling factor; the product of Hadamard is expressed as the multiplication of the corresponding position elements of the matrix; gamma is 1e-6; gradient->

Meaning that the loss function L (f (x; θ) _t ) Y) about theta _t Is a derivative of (2); θ _t To the neuron parameters that need to be optimized; η is the learning rate.

5. The power prediction correction method considering the climbing characteristic of the photovoltaic power according to claim 1, wherein in the step 5, a time series correspondence f is calculated _Tsc The method comprises the following specific steps: s is selected by using shape DTW similarity measurement method _i Representative elements in the subset

Calculation of +.about.using shape DTW method>

Time-series correspondence f with power prediction curve FCST (t) _Tsc The method comprises the steps of carrying out a first treatment on the surface of the The ShapeDTW measures the similarity between photovoltaic power curves, firstly, the euclidean distance is used for measuring the similarity between sequence shape descriptors d, so that the corresponding relation between two photovoltaic power time sequences is obtained, and Ji Guangfu power time sequences are realized; where d= (d ₁ ,d ₂ ,…d _L ) ^T ,d _i ∈R ^m ，d＝F(S)＝(F(s ₁ ),F(s ₂ ),…,F(s _L )) ^T The shape descriptor computation function F takes the form of a derivative, i.e. for the raw power data s _i Taking a first derivative with respect to time;

error rate sigma of prediction of subset _i The sequence is according to the corresponding relation f _Tsc Mapping to a target time-of-day sequence, comprising the following specific steps: determining a climbing time period by comparing the difference value of the power of adjacent climbing points of the power prediction curve FCST (t) in the delta t time with the set threshold delta, and recording the climbing starting time t _begin Climbing end time t _end The method comprises the steps of carrying out a first treatment on the surface of the FCST (t) and representative element prediction by photovoltaic generation power

Time series correspondence f between _Tsc Calculation of the photovoltaic power prediction bias σ (t) =0.15 σ, eventually used to correct the moment t, using a weighted moving average error _i (f _Tsc (t)-1)+0.7σ _i (f _Tsc (t))+0.15σ _i (f _Tsc (t) +1), where t ε (t) _begin ,t _end ) The method comprises the steps of carrying out a first treatment on the surface of the Performing error correction on FCST (t), and finally predicting result +.>

/>

Images