CN116316591A

CN116316591A - Short-term photovoltaic power prediction method and system based on hybrid bidirectional gating cycle

Info

Publication number: CN116316591A
Application number: CN202310267273.1A
Authority: CN
Inventors: 吉兴全; 张玉敏; 王金玉; 叶平峰; 杨明; 于一潇; 赵国航
Original assignee: Shandong University of Science and Technology
Current assignee: Shandong University of Science and Technology
Priority date: 2023-03-17
Filing date: 2023-03-17
Publication date: 2023-06-23

Abstract

The invention provides a short-term photovoltaic power prediction method and system based on hybrid bi-directional gating circulation, and relates to the field of photovoltaic power generation. The method comprises the steps of obtaining original photovoltaic power data and corresponding meteorological data, and decomposing the original photovoltaic power data to obtain a plurality of photovoltaic power components; preprocessing each photovoltaic power component and corresponding meteorological data to obtain preprocessed photovoltaic power components and corresponding meteorological data; constructing an HBiGRU model comprising a CNN layer, a BiGRU layer and an attention layer, and respectively inputting the preprocessed multiple photovoltaic power components and corresponding meteorological data into the HBiGRU model to obtain a prediction result of each photovoltaic power component; and superposing the prediction results of the photovoltaic power components to obtain a short-term photovoltaic power prediction result. According to the method, uncertainty of photovoltaic power data is reduced, a hybrid bidirectional gating circulation unit model is built, characteristic relations between components and photovoltaic power influence factors are fully excavated, and short-term photovoltaic power prediction accuracy is improved.

Description

Short-term photovoltaic power prediction method and system based on hybrid bidirectional gating cycle

Technical Field

The invention belongs to the technical field of photovoltaic power generation power prediction, and particularly relates to a short-term photovoltaic power prediction method and system based on hybrid bidirectional gating circulation.

Background

The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.

Short-term photovoltaic power generation power prediction is mainly divided into a physical method and a statistical method. Due to the complex modeling process of the physical method and high requirement on the accuracy of input data, more and more researchers tend to adopt a statistical method to predict the short-term photovoltaic power generation power. Common statistical methods include methods such as a support vector machine, a neural network, an extreme learning machine, an echo state network and the like, but the statistical methods have limited capability of processing complex nonlinear problems, and are difficult to comprehensively capture characteristic relations between photovoltaic power data and related influence factors, so that prediction accuracy is difficult to improve.

With the rapid development of artificial intelligence technology, the deep learning method has been widely applied to the fields of fault diagnosis, power distribution network reconstruction, state estimation, load prediction, new energy prediction and the like in a power system because of strong nonlinear fitting capability and generalization capability. In recent years, attention has been paid to students at home and abroad especially in the field of new energy prediction. Researchers respectively adopt a Back Propagation Neural Network (BPNN) and a Convolution Neural Network (CNN) to predict the photovoltaic output, so that the photovoltaic power prediction precision is effectively improved. But the time sequence correlation between historical data is ignored in the BPNN and CNN modeling process, and the photovoltaic power prediction accuracy is limited.

The gating cyclic unit network (GRU) is an improved model based on a cyclic neural network (RNN), overcomes the inherent gradient explosion problem of the RNN, can effectively mine the time sequence characteristics of data, improves the prediction accuracy, and has been well established in the field of photovoltaic prediction. The method comprises the steps that researchers preprocess a photovoltaic power data set by adopting an improved principal component analysis method, then construct a GRU model to predict the processed data set, and obtain a photovoltaic power prediction result; in addition, researchers have proposed a photovoltaic power prediction model based on CNN-GRU, and the result shows that the CNN-GRU model has more excellent prediction performance compared with the GRU model.

However, the inventors found that the GRU model and CNN-GRU do not consider the influence of the time series data bidirectional time information on the prediction result, and that the problem of forgetting important time series information and the like easily occurs when processing a longer time series.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides a short-term photovoltaic power prediction method and a system based on hybrid bidirectional gating circulation, which are based on a combined quadratic modal decomposition method of self-adaptive noise complete set empirical modal decomposition (CEEMDAN), sample Entropy (SE) and Variation Modal Decomposition (VMD), so as to decompose photovoltaic power data to obtain a series of relatively stable eigenmode function components, reduce the uncertainty of the photovoltaic power data, construct a hybrid bidirectional gating circulation unit model (HBiGRU), fully excavate the characteristic relation between each component and the photovoltaic power influencing factors to obtain each component prediction result, and then superimpose each component prediction result to obtain a short-term photovoltaic power prediction result, thereby improving the short-term photovoltaic power prediction precision.

To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

the first aspect of the invention provides a short-term photovoltaic power prediction method based on hybrid bi-directional gating loops.

The short-term photovoltaic power prediction method based on the hybrid bi-directional gating cycle comprises the following steps of:

the method comprises the steps of obtaining original photovoltaic power data and corresponding meteorological data, and decomposing the original photovoltaic power data to obtain a plurality of photovoltaic power components;

preprocessing each photovoltaic power component and corresponding meteorological data to obtain preprocessed photovoltaic power components and corresponding meteorological data;

constructing an HBiGRU model comprising a CNN layer, a BiGRU layer and an attention layer, and respectively inputting the preprocessed multiple photovoltaic power components and corresponding meteorological data into the HBiGRU model to obtain a prediction result of each photovoltaic power component;

superposing the prediction results of the photovoltaic power components to obtain a short-term photovoltaic power prediction result;

the BiGRU layer comprises two layers of GRU models with the same output and opposite information transmission directions, the GRU model considering the past moment information is set to be a positive sequence GRU layer, the GRU model considering the future moment information is set to be a negative sequence GRU layer, the output of the BiGRU layer at the t moment is the sum of the product of the output of the positive sequence GRU layer at the t moment and the output weight matrix of the positive sequence GRU layer, the product of the output of the negative sequence GRU layer at the t moment and the output weight matrix of the negative sequence GRU layer, and the output weight matrix of the positive sequence GRU layer and the output weight matrix of the negative sequence GRU layer are obtained by pre-training the BiGRU layer.

A second aspect of the invention provides a short-term photovoltaic power prediction system based on hybrid bi-directional gating loops.

A hybrid bi-directional gating cycle based short term photovoltaic power prediction system comprising:

a decomposition module configured to: the method comprises the steps of obtaining original photovoltaic power data and corresponding meteorological data, and decomposing the original photovoltaic power data to obtain a plurality of photovoltaic power components;

a preprocessing module configured to: preprocessing each photovoltaic power component and corresponding meteorological data to obtain preprocessed photovoltaic power components and corresponding meteorological data;

a prediction module configured to: constructing an HBiGRU model comprising a CNN layer, a BiGRU layer and an attention layer, and respectively inputting the preprocessed multiple photovoltaic power components and corresponding meteorological data into the HBiGRU model to obtain a prediction result of each photovoltaic power component;

a superposition module configured to: superposing the prediction results of the photovoltaic power components to obtain a short-term photovoltaic power prediction result;

A third aspect of the invention provides a computer readable storage medium having stored thereon a program which when executed by a processor implements the steps in a hybrid bi-directional gating cycle based short term photovoltaic power prediction method according to the first aspect of the invention.

A fourth aspect of the invention provides an electronic device comprising a memory, a processor and a program stored on the memory and executable on the processor, the processor implementing the steps in the hybrid bi-directional gating cycle based short term photovoltaic power prediction method according to the first aspect of the invention when the program is executed.

The one or more of the above technical solutions have the following beneficial effects:

1) The invention constructs an HBiGRU model based on CNN, a bidirectional gating circulating unit (BiGRU) and an Attention Mechanism (AM), and is used as a QMD subsequence prediction model, and CNN is introduced into an input side to fully excavate the characteristic relation between a photovoltaic power generation power subsequence and relevant meteorological data; and a time sequence attention mechanism is introduced at the output side, so that the influence of important time sequence information on a prediction result is further highlighted, and the prediction precision of the BiGRU model is effectively improved.

2) The method for decomposing the photovoltaic power data based on the combined secondary mode decomposition method of the self-adaptive noise complete set empirical mode decomposition (CEEMDAN), sample Entropy (SE) and Variation Mode Decomposition (VMD) is provided, the problem that the EMD mode aliasing and EEMD low-frequency components are too many is solved by utilizing the self-adaptive noise complete set empirical mode decomposition (CEEMDAN), the influence of noise on a decomposition sequence is furthest restrained by utilizing the Variation Mode Decomposition (VMD) method, the decomposition effect is improved, and compared with the single decomposition method, a more accurate prediction result can be obtained.

Additional aspects of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention.

Fig. 1 is a flowchart of decomposing original photovoltaic power data according to a first embodiment.

Fig. 2 is a schematic diagram of the CNN structure of the first embodiment.

Fig. 3 is a schematic view of the first embodiment of the GRU structure.

Fig. 4 is a schematic diagram of a first embodiment biglu.

Fig. 5 is a diagram showing the structure of the attention mechanism of the first embodiment.

Fig. 6 is a diagram showing the structure of the hbiglu model of the first embodiment.

Fig. 7 is a short-term photovoltaic power prediction model based on QMD-hbiglu of the first embodiment.

Fig. 8 shows the entropy of each IMF component sample of the photovoltaic power data sample of the first embodiment.

FIG. 9 is a diagram showing the prediction results of the IMF1-3 after reconstruction according to the first embodiment.

FIG. 10 is a comparative diagram showing the predicted results of the IMF1 according to the first embodiment.

Fig. 11 is a schematic diagram showing the photovoltaic power prediction results of the first example of 6 models on 3 months and 8 days.

Fig. 12 is a graph showing the photovoltaic power prediction results of the first example of 6 models on 3 months and 9 days.

Fig. 13 is a schematic diagram showing the predicted results of the photovoltaic power of the first example, 6 models on 3 months and 10 days.

Fig. 14 is a schematic view of a photovoltaic power data sample of the first embodiment.

Fig. 15 is a schematic diagram showing VMD decomposition results of a photovoltaic power data sample of the first embodiment.

Fig. 16 is a flowchart of a method of the first embodiment.

Fig. 17 is a system configuration diagram of the second embodiment.

Detailed Description

It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the invention. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the present invention.

Embodiments of the invention and features of the embodiments may be combined with each other without conflict.

Example 1

The embodiment discloses a short-term photovoltaic power prediction method based on hybrid bi-directional gating cycle.

As shown in fig. 6,7 and 16, the short-term photovoltaic power prediction method based on the hybrid bi-directional gating cycle comprises the following steps:

Further, the method for decomposing the original photovoltaic power data specifically comprises the following steps:

performing preliminary decomposition on the original photovoltaic power data by using an adaptive noise complete set empirical mode decomposition method to obtain a plurality of IMF components;

calculating the time complexity of each IMF component based on a sample entropy method, and overlapping and reconstructing the IMF components with similar complexity to obtain IMF components with reconstructed entropy values;

and performing secondary decomposition on the high-frequency component in the IMF component with the reconstructed entropy value by using a variable-fraction modal decomposition method to obtain a photovoltaic power component.

More specifically:

and (1) original photovoltaic power treatment:

decomposing the original photovoltaic power data by using a QMD method to obtain N relatively stable component sets { DIMF ₁ ,DIMF ₂ ,…,DIMF _N }。

The invention provides a QMD method for decomposing an original photovoltaic power sequence, and decomposing a non-stable photovoltaic power sequence into a series of components with different modes, so that the complexity of the original photovoltaic power sequence is reduced. The QMD flow is shown in fig. 1.

The QMD method consists of CEEMDAN, SE and VMD methods. Firstly, performing preliminary decomposition on photovoltaic power data by using a CEEMADN method to reduce the complexity of the photovoltaic power time sequence data; secondly, calculating the complexity of time sequence data of each component based on an SE method, overlapping components with similar complexity, and resetting the concept of replacing high dimension with low dimension, so as to reduce the calculated amount of the model; finally, components with high complexity are further decomposed using the VMD method to reduce residual noise in the high complexity components.

1.1 adaptive noise complete set empirical mode decomposition (CEEMDAN)

The CEEMDAN method is a self-adaptive data decomposition method, and the data is decomposed into a limited number of eigen-mode function (intrinsic mode function, IMF) components with different time scales by adding white noise with opposite signs in the decomposition process, so that the problems of modal aliasing, excessive residual noise and the like existing in EMD and EEMD are effectively solved.

The basic steps of CEEMADN decomposition are as follows:

1) Adding Gaussian white noise with the average value of 0 for n times into data x (t) to be decomposed to obtain a group of data x to be decomposed _i (t), wherein i=1, 2,3, …, n.

x _i (t)＝x(t)+δv _i (t) (1)

Wherein delta is Gaussian white noise coefficient; v _i (t) is the i-th added white gaussian noise.

2) For x _i (t) performing EMD decomposition to obtain a first IMF component

For the n obtained->

Mean->

As the first IMF component of CEEMDAN decomposition.

in the formula ,

is the residual obtained after the first decomposition.

3) At the residual error

Continuously adding Gaussian white noise to obtain a group of new data to be decomposed, and repeating the step 2 to obtain a second IMF component of CEEMADN decomposition>

And residual->

4) Repeating the steps until the obtained residual error is a monotonic function and can not be decomposed continuously, and stopping iteration.

1.2 Sample Entropy (SE)

The CEEMDAN method is adopted to process the original photovoltaic power sequence, so that the prediction precision of the prediction model can be remarkably improved, but with the increase of IMF components, the calculated amount of the prediction model is remarkably increased, and the model prediction precision is easily reduced due to the occurrence of the over-fitting phenomenon. In order to avoid the over-fitting problem, the invention provides a SE method for calculating the time complexity of each IMF component and carrying out aggregation reconstruction on the components with similar complexity.

SE is an improved method for measuring time series data complexity based on approximate entropy, and SE calculation is independent of data length, so that better consistency is achieved. The greater the SE, the greater the complexity of the timing data. The SE detailed calculation procedure is as follows:

1) Assuming that there is a time sequence s= { S (1), S (2), …, S (N) }, the sequence S is sequentially reconstructed into an m-dimensional sequence S (l), expressed as:

S(l)＝[s(l),s(l+1),L,s(l+m-1)] (4)

where l=1, 2, …, N-m+1.

2) Calculating the maximum difference d between corresponding elements of the rest sequence S (k) (k=1, 2, …, N-m+1, and k+.l) _max (S(l),S(k))。

3) Given a threshold r (r>0) Statistical difference d _max (S (l), S (k)) is less than r, and the ratio of the sum of the differences is calculated and recorded as

Wherein num is d _max (S(l),S(k))<r number.

4) For all of

Taking an average value:

5) Increasing the dimension to m+1, and repeating the steps to obtain

Average value of (2):

6) According to D ^m(r) and D^m+1 (r) data calculation to obtain sample entropy sampenn (m, r) of time series s:

1.3 Variation Modal Decomposition (VMD)

After the IMF component is reconstructed according to the entropy value, the high-frequency component still contains partial noise, so that the complexity of the high-frequency component is higher, and the prediction accuracy of the high-frequency component is lower. Therefore, the VMD method is provided for carrying out secondary decomposition on the high-frequency component, so that the complexity of the high-frequency component is further weakened.

The VMD method is a self-adaptive and completely non-recursive modal decomposition method, and in the process of decomposing the sequence, the center frequency and the bandwidth of each component are determined by iteratively searching the optimal solution of the variational model, so that the optimal center frequency and the bandwidth of each modal are self-adaptively matched, and the effective separation of IMF components is realized.

The VMD detailed solution process is as follows:

1) Decomposing the original signal χ into j finite bandwidth modal functions, for each modal function u _j (t) performing Hilbert transform to obtain a single-sided frequency spectrum.

2) For u _j (t) center frequency ζ of frequency spectrum _j (t) modifying to tune to the baseband bandwidth.

3) Estimating u by computing the gradient binary norms of the demodulated signal _j Building a variational model with constraint conditions of the fundamental frequency bandwidth of (t):

in the formula ,{u_j -a set of modal functions; { xi } _j The central frequency set of the spectrum corresponding to the modal function;

is a partial derivative operator; * Representing a convolution calculation; delta (t) is the unit pulse function.

4) In order to simplify the calculation process, the variation model is converted into a variation model of an unconstrained condition, a Lagrange operator lambda (t) and a secondary penalty term theta are introduced, an extended Lagrange expression is constructed by substituting the Lagrange operator lambda (t) and the secondary penalty term theta into the step 3, and the solution is carried out through an alternate direction multiplier method.

And secondly, carrying out normalization processing on each photovoltaic power component and the corresponding meteorological data, and compressing the range of data values to [0,1] to obtain the preprocessed photovoltaic power components and the preprocessed meteorological data.

And (3) carrying out normalization processing on each component and meteorological data to unify dimensions, and compressing the value ranges of each sequence to [0,1] after normalization processing, wherein the formula is as follows:

wherein f and f' are data before and after normalization, f _min and f_max Respectively, the minimum and maximum of the data set.

And thirdly, constructing an HBiGRU prediction model.

And building an HBiGRU prediction model aiming at the decomposed Nth component, training the HBiGRU model by taking historical data and meteorological data of each component as input vectors, and building a corresponding prediction model through iterative optimization model parameters of a loss function.

The hbiglu model is divided into five layers, each layer of network structure and parameters are as follows:

(1) An input layer. And preprocessing the photovoltaic power component with the length of T and meteorological data, and inputting the preprocessed photovoltaic power component and meteorological data into a prediction model.

(2) CNN layer. The CNN layer consists of 1 one-dimensional convolution layer, 1 one-dimensional pooling layer and a full connection layer and is used for extracting key characteristics of input data. The convolution layer selects a tanh function as an activation function, and a pooling mode of the pooling layer selects maximum pooling. The input data is subjected to dimension reduction processing of a convolution layer and a pooling layer, then feature vectors are extracted through a full-connection layer, the feature vectors are used as input of a BiGRU model, and the full-connection layer selects a Sigmoid function as an activation function.

The CNN model maps the original data to a high-dimensional space in a local connection and weight sharing mode, and key features of the input data are extracted efficiently. The CNN model structure is shown in fig. 2, where the convolution layer extracts the input data features through a convolution kernel that shares weights, and uses an activation function to non-linearly map the neurons. The pooling layer reduces the dimension of the features obtained by convolution operation through a maximum pooling or average pooling method, compresses the number of the features and reduces the risk of overfitting. The fully connected layer is composed of fully connected neural networks and is used for flattening and integrating the characteristics extracted by the convolution layer and the pooling layer.

(3) A biglu layer. The BiGRU layer is used for learning the feature vector extracted by the CNN layer. And building a single-layer BiGRU network by taking the relu function as an activation function, learning the local feature internal relation and time sequence change rule extracted by CNN, and predicting future time data.

The bidirectional gating circulation unit (BiGRU) is a special variant structure of the GRU, and consists of two layers of GRU models with the same output and opposite information transmission directions, and the BiGRU can fully mine the time sequence characteristics of the photovoltaic power by considering the information of the past moment and the future moment simultaneously when the prediction is carried out, so that the data utilization rate and the model prediction precision are improved. The single-layer GRU structure is shown in figure 3.

According to the structure of fig. 3, the parameter calculation of the GRU model can be expressed as:

in the formula ,X_t The input of the GRU model at the moment t is represented; delta and tanh represent sigmoid and hyperbolic tangent activation functions, respectively;

representing candidate hidden states, represented by input X _t And the hidden layer output h at the last moment _t-1 Calculating to obtain; h is a _t Representing hidden layer output; w (W) _c 、W _g And W represents the weight matrix of the reset gate, update gate and candidate hidden state, respectively;

c _t representing GRU reset gate, representing input X at time t _t Output h from hidden layer at last moment _t-1 Hiding state from candidates

Influence of g _t Represents the GRU update gate, represents the hidden layer output h at the last time _t-1 Hidden layer output h for time t _t Is a function of the size of the impact.

According to the GRU model, a positive sequence GRU layer and a negative sequence GRU layer are arranged to construct a BiGRU model, the basic structure of the BiGRU model is shown in figure 4,

represents the positive-order GRU model output at time t, < >>

And (5) indicating the output of the negative sequence GRU model at the time t.

The biglu model calculation formula can be expressed as:

in the formula ,

representing the output quantity of the positive sequence GRU model at the time t-1; />

Expressing the output quantity of the negative sequence GRU model at the time t+1; w (W) _f1 And W is equal to _f2 Respectively inputting a weight matrix for the positive sequence GRU model and outputting the weight matrix at the moment t-1; w (W) _b1 And W is equal to _b2 Respectively inputting a weight matrix for the negative sequence GRU model and outputting the weight matrix at the time t+1; w (W) ₁ And W is equal to ₂ And outputting weight matrixes for the positive sequence GRU model and the negative sequence GRU model respectively.

(4) Attention layer. The attention layer is used for processing the BiGRU hidden layer output vector, calculating weights corresponding to different hidden layer vectors, deciding an optimal weight matrix, and effectively mining deep time sequence correlation among features.

The Attention Mechanism (AM) is a weight distribution strategy provided by simulating a human brain attention resource distribution mechanism, and the core idea is to apply larger weight to key data by changing the weights of different input data, so that important information in the input data is rapidly screened, and the model prediction accuracy is improved. The attention mechanism unit structure is shown in fig. 5.

Hidden layer output vector h obtained by BiGRU model _t As input, a weight matrix is calculated according to the AM weight allocation principle, and the specific calculation steps are as follows:

1) Calculating the attention weight e of the input characteristic by taking the tanh function as an activation function _t ：

e _t ＝μtanh(wh _t +b) (17)

Where μ and w are weight coefficients and b is a bias coefficient.

2) The attention weight is normalized through the Softmax function, the attention weight sum is ensured to be 1, and the normalized weight coefficient alpha _t ：

3) The input information is weighted and summed according to the weight coefficient to obtain the attentionValue h _t ′：

(5) And an output layer. The output layer selects a Sigmoid function as an activation function, and a final prediction result is calculated according to the output of the attention layer.

And superposing the component prediction results to obtain a photovoltaic power prediction result, comparing the photovoltaic power prediction result with other prediction methods through error evaluation indexes, and analyzing the performance of the prediction model.

The following gives an example analysis in this embodiment:

(1) Experimental platform configuration and data sources

The invention adopts a Jupiter integrated development environment and a tensorsurface deep learning framework in anaconda software to write a short-term photovoltaic power prediction model. And (3) selecting actual measurement data of a photovoltaic station of the solar center of Australian desert knowledge to carry out simulation experiments. Taking the power generation data and the meteorological data of the photovoltaic station 2017 from 1 month 1 day to 3 months 10 days as examples, the effectiveness of the model and the method provided by the invention is verified through simulation. Because the photovoltaic output is zero at night, the invention selects data of 7:00-19:00 a day for simulation, the sampling interval is 5min, and the obtained photovoltaic power sample is shown in fig. 14. A total of 9504 sampling points are selected as training samples for the first 66 days, and 432 sampling points are selected for the last 3 days for test verification.

(2) Evaluation index

The invention selects normalized mean absolute error (normalized mean absolute error, NMAE), normalized root mean square error (normalized root mean square error, NRMSE) and decision coefficient (coefficient of determination, R) ² ) As a main evaluation index, the model prediction accuracy is characterized. NMAE, NRMSE and R ² The calculation formula is as follows:

wherein ω is a predicted time length; y is _Σ Rated installed capacity for the photovoltaic field station; y is _i and y′_i The real value and the predicted value of the photovoltaic power at the moment i are respectively;

is the average value of the photovoltaic power.

(3) Secondary decomposition effectiveness verification

In order to verify the superiority of the QMD method provided by the invention in noise elimination, the CEEMDAN-SE method and the QMD method are adopted to decompose photovoltaic power data, an HBiGRU prediction model is constructed for the decomposed subsequence, and the prediction result is compared and verified.

1) CEEMDAN-SE decomposition effect and predictive analysis

As can be seen from fig. 14, the photovoltaic power has obvious fluctuation, in order to improve the photovoltaic power prediction accuracy, the CEEMDAN method is first used to decompose the photovoltaic power sequence to obtain 13 IMF components, and according to SE theory, the entropy value of each IMF component is calculated, and the calculation result is shown in fig. 8.

As can be seen from fig. 8, the first 4 IMF components have higher entropy values and similar complexity, so the first 4 IMF components are superimposed as one IMF component for prediction. And the other IMF components are processed by the similar method to obtain superposed IMF components, and the superposed IMF components are shown in a table 1.

TABLE 1 IMF component superposition results

Table1 IMF weight stack results

And respectively establishing an HBiGRU prediction model for the overlapped and reconstructed IMF components to predict, wherein the prediction result is shown in figure 9.

As can be seen from fig. 9, the IMF2 and IMF3 components after superposition reconstruction have gentle trend, the fitting effect of the predicted value and the true value is good, and the IMF1 component is disordered, and the prediction accuracy is lower than that of IMF2 and IMF 3. After superposition polymerization, the IMF1 still contains a large amount of noise, which affects the prediction accuracy. Therefore, the invention adopts the VMD to carry out secondary decomposition on the IMF1 component, thereby further reducing the complexity of the IMF1 and improving the prediction precision.

2) QMD effect and predictive analysis

In order to verify the performance of the QMD method for reducing noise and improving prediction precision, the invention further adopts a VMD method to process the IIMF1 component on the basis of the CEEMDAN-SE method, eliminates irrelevant noise in the IMF1 component, reduces the complexity degree of the IMF1 component, and further improves the prediction precision of the IMF1 component.

The VMD decomposition needs to set the number of decomposition modes in advance, and if the number of decomposition is too large, excessive noise is generated by the decomposition, and if the number of decomposition is too small, the mode decomposition is insufficient, so that the prediction accuracy is affected. Whether overdecomposing is performed is mainly based on whether center frequency components of all modes are close or not, so that the VMD decomposition mode number K is determined by adopting a center frequency method. And respectively setting K=3, 4,5,6 and 7 to perform VMD decomposition, calculating the center frequency of each component, finding out the K value of the component with similar center frequency, and selecting K-1 as the decomposition mode number. The center frequencies of the components v at k=3, 4,5,6,7 are shown in table 2.

Table 2 center frequencies corresponding to different mode numbers K

Table2 The center frequency corresponding to the differ-ent mode number K

As can be seen from table 3, as the K value increases, the difference in center frequencies of the components gradually decreases. When k=7, v ₁ and v₂ The center frequencies of the two are similar, and an excessive decomposition phenomenon occurs, so that extra noise is generated. To ensure the VMD denoising effect, the inventionSelecting k=6 to perform VMD decomposition on IIMF1, and the VMD decomposition result is shown in fig. 15.

And (3) establishing an HBiGRU model for each decomposed modal component, carrying out prediction, overlapping and reconstructing the result, obtaining an IMF1 prediction result after QMD decomposition, as shown in fig. 10, and comparing the IMF1 prediction result with a CEEMDAN-SE decomposition IMF1 prediction result, as shown in table 3.

TABLE 3 QMD Pre-post IMF1 prediction error comparison

Table3 Comparison of IMF1 prediction error before and after quadratic decomposition

As can be seen from fig. 10 and table 4, prediction accuracy of IMF1 components after VMD treatment is significantly improved by using hbiglu model, compared with NMAE and NRMSE which are directly predicted after CEEMDAN-SE decomposition, which are reduced by 1.58% and 2.78%, respectively. R is R ² From 0.47 to 0.88. The QMD method provided by the invention can greatly reduce the decomposition noise and effectively improve the prediction precision of the HBiGRU prediction model on complex sequences.

3) Short term photovoltaic power prediction result analysis

In order to verify the superior performance of the QMD-HBiGRU (H6) model in terms of photovoltaic power prediction accuracy improvement, biGRU (H1), HBiGRU (H2), CEEMDAN-SE-BiGRU (H3), CEEMDAN-SE-HBiGRU (H4) and QMD-BiGRU (H5) 5 models are respectively constructed as comparison models, photovoltaic power of 3 days from 3 months 8 days to 3 months 10 days is predicted, the prediction accuracy results of the 6 models are respectively shown in fig. 11 to 13, and the evaluation indexes of the 6 models are shown in table 4.

Table 4 comparison of prediction errors for different models

Table4 Compalison of prediction errors of different mod-els

As can be seen from fig. 11 to 13, the photovoltaic output has volatility and randomness, and if the photovoltaic power is directly predicted, the complex nonlinear relationship is difficult to capture, and the prediction error is large. As can be seen from Table 4, the average NMAE and NRMSE values for 3 days of the H1 predictive model were 4.67% and 7.03%, respectively, with an average R ² The value was 0.79. Further, the fact that the single prediction model H1 is adopted to directly predict photovoltaic power errors is larger is explained, and the requirement on photovoltaic prediction accuracy in engineering practice is difficult to achieve. Average NMAE and NRMSE values for the H2 predictive model were reduced by 0.85% and 1.45%, respectively, compared to the H1 model, R ² The value is increased to 0.87, which shows that after the CNN module and the attention module are added on the basis of the BiGRU model, the capability of capturing nonlinear characteristics of the BiGRU model can be improved by means of the characteristic extraction capability of CNN and an attention mechanism, so that the prediction accuracy is improved. The average NMAE and NRMSE of the H3 predictive model were reduced by 0.92% and 2.04%, respectively, compared to the H1 predictive model. The average NMAE and NRMSE of the H4 predictive model were reduced by 1.09% and 1.37%, respectively, compared to the H2 predictive model. Proved by CEEMDAN-SE method, the influence of fluctuation and randomness of input data on prediction accuracy can be reduced to a certain extent, and the prediction accuracy of the H5 prediction model for obtaining a more accurate prediction result is greatly improved compared with that of H1 and H3 models. The average NMAE value and NRMSE value of the model H6 provided by the invention are smaller than those of other 5 prediction models, namely 1.46% and 1.99%, and the average R of the H6 prediction model is the same as the average R of the other 5 prediction models ² The value is 0.97, the fitting degree of the predicted value and the true value of the H6 prediction model is good, the secondary decomposition by the VMD method is verified, the redundant noise in the high-frequency component is removed, and the photovoltaic power prediction precision can be further improved.

In conclusion, under the condition of larger photovoltaic power fluctuation, the QMD-HBiGRU prediction model provided by the invention can keep higher prediction precision, and the effectiveness of the method provided by the invention is verified.

Example two

The embodiment discloses a short-term photovoltaic power prediction system based on hybrid bi-directional gating cycle.

As shown in fig. 17, a short-term photovoltaic power prediction system based on hybrid bi-directional gating cycles, comprising:

Example III

An object of the present embodiment is to provide a computer-readable storage medium.

A computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps in a hybrid bi-directional gating cycle based short term photovoltaic power prediction method as described in embodiment 1 of the present disclosure.

Example IV

An object of the present embodiment is to provide an electronic apparatus.

An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, which when executed, implements the steps in a hybrid bi-directional gating cycle based short term photovoltaic power prediction method as described in embodiment 1 of the present disclosure.

The steps involved in the devices of the second, third and fourth embodiments correspond to those of the first embodiment of the method, and the detailed description of the embodiments can be found in the related description section of the first embodiment. The term "computer-readable storage medium" should be taken to include a single medium or multiple media including one or more sets of instructions; it should also be understood to include any medium capable of storing, encoding or carrying a set of instructions for execution by a processor and that cause the processor to perform any one of the methods of the present invention.

It will be appreciated by those skilled in the art that the modules or steps of the invention described above may be implemented by general-purpose computer means, alternatively they may be implemented by program code executable by computing means, whereby they may be stored in storage means for execution by computing means, or they may be made into individual integrated circuit modules separately, or a plurality of modules or steps in them may be made into a single integrated circuit module. The present invention is not limited to any specific combination of hardware and software.

While the foregoing description of the embodiments of the present invention has been presented in conjunction with the drawings, it should be understood that it is not intended to limit the scope of the invention, but rather, it is intended to cover all modifications or variations within the scope of the invention as defined by the claims of the present invention.

Claims

1. The short-term photovoltaic power prediction method based on the hybrid bi-directional gating cycle is characterized by comprising the following steps of:

2. The short-term photovoltaic power prediction method based on hybrid bi-directional gating cycle of claim 1, wherein the decomposing of the raw photovoltaic power data specifically comprises:

calculating the time complexity of each IMF component based on a sample entropy method, and carrying out polymerization superposition reconstruction on the I MF components with similar complexity to obtain I MF components with the reconstructed entropy value;

and performing secondary decomposition on the high-frequency component in the I MF component with the reconstructed entropy value by using a variable-fraction modal decomposition method to obtain a photovoltaic power component.

3. The hybrid bi-directional gating cycle-based short term photovoltaic power generation method of claim 1 wherein each photovoltaic power component and corresponding meteorological data is normalized and the range of data values is compressed to 0,1 to yield preprocessed photovoltaic power components and meteorological data.

4. The short-term photovoltaic power prediction method based on hybrid bi-directional gating cycle according to claim 1, wherein the preprocessed photovoltaic power components and the corresponding meteorological data are respectively input into an hbiglu model to obtain a prediction result of each photovoltaic power component, specifically:

inputting each preprocessed photovoltaic power component and corresponding meteorological data into a CNN layer, and extracting feature vectors;

inputting the feature vectors into the BiGRU layer, and learning the internal relation and time sequence change rule among the feature vectors to obtain a plurality of hidden layer output vectors;

inputting a plurality of hidden layer output vectors into an attention layer, and calculating weights corresponding to different hidden layer output vectors according to an AM weight distribution principle to obtain an optimal weight matrix;

and obtaining a prediction result of each photovoltaic power component based on the optimal weight matrix output by the attention layer and the output vectors of the plurality of hidden layers.

5. The hybrid bi-directional gating cycle-based short term photovoltaic power prediction method of claim 4, wherein the biglu model is calculated according to the formula:

wherein ,

the positive sequence GRU model output at the moment t is shown; />

The output of a negative sequence GRU model at the moment t is represented; />

Expressing the output quantity of the negative sequence GRU model at the time t+1; w (W) _f1 And W is equal to _f2 Respectively inputting a weight matrix for the positive sequence GRU model and outputting the weight matrix at the moment t-1; w (W) _b1 And W is equal to _b2 Respectively inputting a weight matrix for the negative sequence GRU model and outputting the weight matrix at the time t+1; w (W) ₁ And W is equal to ₂ Respectively outputting weight matrixes for the positive sequence GRU model and the negative sequence GRU model; x is X _t The input of a model at the time t is represented; delta represents a sigmoid activation function.

6. The short-term photovoltaic power prediction method based on hybrid bi-directional gating cycle as claimed in claim 4, wherein the CNN layer is composed of 1 one-dimensional convolution layer, 1 one-dimensional pooling layer and full-connection layer, wherein the convolution layer selects a tanh function as an activation function, the pooling mode of the pooling layer selects maximum pooling, the feature vector is extracted through the full-connection layer after the input data is subjected to the dimension reduction treatment of the convolution layer and the pooling layer, and the full-connection layer selects a Sigmoid function as the activation function.

7. The hybrid bi-directional gating cycle-based short-term photovoltaic power prediction method of claim 4 wherein the hidden layer output vectors are weighted and summed according to the corresponding weight coefficients in the optimal weight matrix to obtain the prediction result for each photovoltaic power component.

8. Short-term photovoltaic power prediction system based on hybrid bi-directional gating cycle, its characterized in that: comprising the following steps:

9. A computer readable storage medium having stored thereon a program, which when executed by a processor implements the steps of the hybrid bi-directional gating cycle based short term photovoltaic power prediction method according to any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor and a program stored on the memory and executable on the processor, wherein the processor implements the steps of the hybrid bi-directional gating cycle-based short term photovoltaic power prediction method of any one of claims 1-7 when the program is executed.