Photovoltaic Power Prediction Method and System Based on Frequency Domain Decoupling and Multi-Period Fusion

Abstract

The present invention provides a photovoltaic power prediction method and system based on frequency-domain decoupling and multi-period fusion. First, historical photovoltaic power data is obtained and preprocessed to obtain a photovoltaic power sequence; through frequency-domain decoupling, a decoupled two-dimensional tensor of multiple redundant-free periods characterizing the multi-period coupling characteristics in the photovoltaic power sequence is obtained; based on the period scale of the decoupled two-dimensional tensor, an adaptive graph convolutional neural network is used to capture the intra-period trend and cross-period dependence relationships in the photovoltaic power sequence, generating a period feature set; through multi-period fusion of the period feature set, a period fusion convolutional matrix is obtained; finally, based on the period fusion convolutional matrix, a photovoltaic power prediction model is established using the Transformer Encoder architecture to predict the photovoltaic power. The present invention fully considers the multi-period coupling characteristics of the photovoltaic power sequence, and based on frequency-domain decoupling and multi-period fusion, can adapt to the distribution changes of non-stationary photovoltaic time series, greatly improving the prediction accuracy and stability of photovoltaic power.

Description

Technical Field

[0001] The present invention belongs to the technical field of power engineering and relates to a power prediction method and system. Background Art

[0002] To achieve energy conservation and emission reduction, the proportion of renewable energy in the power system is gradually increasing. As a large-scale renewable energy, solar energy has an increasingly growing influence on the power system in solar power grid-connected power generation. However, solar energy is an intermittent energy source. Changes in meteorological factors will cause fluctuations in photovoltaic output power, increasing instability. At the same time, considering that some power systems cannot connect to the Internet to obtain weather forecast information due to network security protection requirements, it further increases the difficulty of power system dispatching and management. Accurate prediction of photovoltaic power can greatly improve the accommodation capacity of photovoltaic power and is of great significance for the power system dispatching in a short period of time.

[0003] Currently, photovoltaic power prediction methods mainly rely on mathematical statistical models or machine learning algorithms. The current methods are simple to operate and respond quickly, but their performance in the prediction regression field is strongly positively correlated with the quantity and quality of data. However, photovoltaic power sequence data usually exhibits different intra-sequence and inter-sequence correlations. The current methods mainly focus on modeling the time variation of one-dimensional photovoltaic power sequences, ignoring the complex and intertwined dependencies inherent in photovoltaic data. At the same time, the short-term fluctuations, rises, and falls in photovoltaic data are often coupled with each other. The algorithms in the current methods are limited by their network characteristics and cannot fully understand the frequency-domain correlations at different time scales between multiple time series, resulting in low accuracy and poor adaptability of the current photovoltaic power prediction methods. Summary of the Invention

[0004] To solve the problems of low accuracy and poor adaptability of the current photovoltaic power prediction methods described in the background art, the present invention provides a photovoltaic power prediction method and system based on frequency-domain decoupling and multi-period fusion.

[0005] The method of the present invention includes:

[0006] Obtain historical photovoltaic power data, and perform preprocessing on the historical photovoltaic power data including missing value processing, outlier interpolation processing, and normalization processing to obtain a photovoltaic power sequence;

[0007] Convert the photovoltaic power sequence from the time domain to the frequency domain through fast Fourier transform to obtain multiple cycle lengths of the photovoltaic power sequence, and make all the cycle lengths non-repetitive to obtain selected cycle lengths. Decouple the one-dimensional input sequence based on the selected cycle lengths to obtain a decoupled two-dimensional tensor of multiple de-redundant cycles characterizing the multi-period coupling characteristics in the photovoltaic power sequence;

[0008] Identify the periodic scales of the decoupled two-dimensional tensor, and capture the intra-period trends and cross-period dependencies in the photovoltaic power sequence based on the periodic scales using an adaptive graph convolutional neural network to generate a periodic feature set;

[0009] Construct a periodic fusion matrix based on the long-period scale tensor and short-period scale tensor in the periodic feature set, calculate the granular tensor of the periodic fusion matrix, and use a one-dimensional convolutional layer to compress the granularity information belonging to different time series to obtain a periodic fusion convolutional matrix;

[0010] Based on the periodic fusion convolutional matrix, establish a photovoltaic power prediction model using the Transformer Encoder architecture to generate long-term time series feature encodings, and map the long-term time series feature encodings to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain the photovoltaic power prediction value.

[0011] Furthermore, the preprocessing of the historical photovoltaic power data includes:

[0012] For missing values, use the mean of the hour before and after the missing moment and the same moment of the previous day to replace them;

[0013] Use the 3 σ sigma principle to detect outliers, and regard the values with errors exceeding the interval of ( μ -3 σ , μ +3 σ ) as outliers, and treat the outliers as missing values;

[0014] Use min-max normalization to scale the photovoltaic power sequence after completing missing value processing and outlier interpolation processing to between [0,1], and its calculation formula is as follows:

[0015] (1),

[0016] where, is the photovoltaic power sequence after normalization processing, X min is the minimum value of the photovoltaic power sequence, X max is the maximum value of the photovoltaic power sequence.

[0017] Even further, the preprocessing obtains a photovoltaic power sequence with a certain time resolution. For the photovoltaic power sequence obtained by preprocessing use a sliding window to intercept sample strips, and the window length is i + j , i represents the length of the input sequence, jLet the label value length be [length], the sliding step be 1, and inside the sliding window, determine the photovoltaic power sequence input for subsequent steps according to the actual prediction requirements. .

[0018] Furthermore, the method for obtaining the decoupled two-dimensional tensor is as follows:

[0019] For the photovoltaic power sequence , use the fast Fourier transform to detect prominent periodicity as the period length { p 1,…, p k} according to the following formula:

[0020] (2),

[0021] where F represents the intensity of each frequency component in L , FFT(·) and Amp(·) respectively represent the fast Fourier transform and amplitude calculation, Avg(·) is the average function, Top k f f f 1,…, f k}; the most prominent frequencies { f 1,…, f k} correspond to k p p 1,…, p k};

[0022] When only considering the positive frequency range, formula (2) is summarized as:

[0023] (3),

[0024] Use Unique (·) function to ensure the uniqueness of the period length. Unique (·) function performs a repeatability check on the period lengths of the first k selected frequencies. If there are repeated period lengths after rounding, remove the repeated period lengths with lower amplitudes and reselect new frequencies from the frequency list sorted by amplitude until all period lengths are unique, as defined below:

[0025] (4),

[0026] Based on the selected period lengths { p’1 ,…, p’ k} Decouple the one-dimensional input sequence using the formula:

[0027] (5),

[0028] where Padding(·) means padding 0 at the end of to make the sequence length divisible by p’ i ;

[0029] Pass the function to decouple the one-dimensional input sequence to obtain a decoupled two-dimensional tensor k with multiple-period coupling characteristics in the photovoltaic power sequence and representing the .

[0030] Furthermore, the generation method of the periodic feature set is as follows:

[0031] First, identify the period scales of the decoupled two-dimensional tensor k with redundant periods removed. Map the tensor corresponding to the n th period scale to a tensor with a period scale of through a linear transformation, where N is the number of photovoltaic decoupled sequences included in the n th period of the decoupled two-dimensional tensor with redundant periods removed. Introduce a weight matrix to capture the local features of the time series, defined as follows:

[0032] (6),

[0033] where h n and W n are learnable weight matrices belonging to the n th decoupled two-dimensional tensor with redundant periods removed;

[0034] Pass and to multiply two trainable parameters, and use the SoftMax(·) function to normalize the weights between different nodes to generate an adaptive adjacency matrix:

[0035] (7),

[0036] After obtaining the nThe adjacency matrix of the decoupled two-dimensional tensor for a redundancy removal period A n After that, MixHop graph convolution is used to capture the dependencies between period scales, and the decoupled two-dimensional tensors for multiple redundancy removal periods are fused and output to obtain a period feature set H n :

[0037] (8),

[0038] wherein, σ (·) is the Sigmoid activation function, P is an integer hyperparameter containing a set of adjacent spacings, ( A n ) j represents the adjacency matrix A n of j power, and [·] is the intermediate output in each iteration of column-level connection.

[0039] Furthermore, the method for obtaining the period fusion convolution matrix is as follows:

[0040] First, linearly reconstruct the period tensors { H 1 ,…, H k} to obtain a period fusion matrix G n :

[0041] (9),

[0042] Then, use an average pooling layer in the Scale dimension:

[0043] (10),

[0044] Finally, expand G’ N in the time series dimension to obtain the granular tensor of the period fusion matrix, and use a one-dimensional convolutional layer to compress the granularity information belonging to different time series to obtain the period fusion convolution matrix C out defined as:

[0045] (11).

[0046] Furthermore, the method for obtaining the predicted photovoltaic power value is:

[0047] Use the key modules including the multi-head attention mechanism and the feed-forward neural network in the Transformer Encoder architecture for the period fusion convolution matrixC out Perform modeling, and the definition of the multi-head attention mechanism is as follows:

[0048] (12),

[0049] where head(·) is an independent attention head, d k represents the dimension of the key vector, Q , K , V respectively represent the query, key, and value matrices, and are defined as: ;

[0050] By introducing the weight matrix W o perform a linear transformation on multiple independent attention heads head(·) after splicing to obtain the cross-period attention embedding of the periodic fusion convolution matrix MSA ( Q , K , V ), and is defined as:

[0051] (13),

[0052] Input the cross-period attention embedding MSA ( Q , K , V ) and the periodic fusion convolution matrix C out into the multi-layer linear feed-forward neural network FFN (·) for feature transformation to generate the long-term time series feature encoding ŷ en , and in this process, batch normalization BatchNorm (·) is used to normalize the input data, and is defined as:

[0053] (14),

[0054] Finally, through the linear layer Linear (·) map the long-term time series feature encoding ŷ en to the target prediction interval to obtain the final photovoltaic power prediction value ŷ pre ; where, Flatten (·) is used to flatten the two-dimensional tensor, and is defined as:

[0055] (15).

[0056] The present invention also provides a photovoltaic power prediction system based on frequency-domain decoupling and multi-period fusion, which includes a photovoltaic power sequence data acquisition module, a frequency-domain decoupling module, a periodic feature set generation module, a multi-period fusion module, and a photovoltaic power prediction module.

[0057] The photovoltaic power sequence data acquisition module is used to obtain historical photovoltaic power data and perform preprocessing on the historical photovoltaic power data, including missing value processing, outlier interpolation processing, and normalization processing, to obtain a photovoltaic power sequence.

[0058] The frequency-domain decoupling module is used to convert the photovoltaic power sequence from the time domain to the frequency domain through fast Fourier transform, obtain multiple cycle lengths of the photovoltaic power sequence, and ensure that all cycle lengths do not repeat, to obtain selected cycle lengths. Based on the selected cycle lengths, the one-dimensional input sequence is decoupled to obtain a decoupled two-dimensional tensor of multiple non-redundant cycles characterizing the multi-period coupling characteristics in the photovoltaic power sequence.

[0059] The periodic feature set generation module is used to identify the periodic scales of the decoupled two-dimensional tensor, and capture the intra-period trend and cross-period dependence relationships in the photovoltaic power sequence by using an adaptive graph convolutional neural network based on the periodic scales, to generate a periodic feature set.

[0060] The multi-period fusion module is used to construct a period fusion matrix according to the long-period scale tensor and short-period scale tensor in the periodic feature set, calculate the granular tensor of the period fusion matrix, and use a one-dimensional convolutional layer to compress the granularity information belonging to different time series to obtain a period fusion convolutional matrix.

[0061] The photovoltaic power prediction module is used to establish a photovoltaic power prediction model based on the period fusion convolutional matrix by adopting a Transformer Encoder architecture, generate long-term time series feature encoding, and map the long-term time series feature encoding to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain a photovoltaic power prediction value.

[0062] The present invention also provides an electronic device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor realizes the photovoltaic power prediction method based on frequency-domain decoupling and multi-period fusion as described above by executing the computer instructions.

[0063] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it realizes the photovoltaic power prediction method based on frequency-domain decoupling and multi-period fusion as described above.

[0064] Compared with the prior art, the present invention first obtains historical photovoltaic power data and preprocesses it to obtain a photovoltaic power sequence; then, through frequency-domain decoupling, a decoupled two-dimensional tensor of multiple redundant-free periods characterizing the multi-period coupling characteristics in the photovoltaic power sequence is obtained; then, based on the period scale of the decoupled two-dimensional tensor, an adaptive graph convolutional neural network is used to capture the intra-period trend and cross-period dependence in the photovoltaic power sequence, generating a period feature set; then, through multi-period fusion of the period feature set, a period fusion convolutional matrix is obtained; finally, based on the period fusion convolutional matrix, a photovoltaic power prediction model is established using the Transformer Encoder architecture to predict the photovoltaic power. The present invention can extract key spectral features in the photovoltaic power sequence. The decoupled two-dimensional tensor obtained through frequency-domain decoupling can characterize the multi-period coupling characteristics in the photovoltaic power sequence, thus being able to adapt to the distribution changes of non-stationary photovoltaic time series; at the same time, an adaptive graph convolutional neural network is introduced to capture the intra-period trend and cross-period dependence in the photovoltaic power sequence, so that the generated period feature set has the key period features in the decoupled two-dimensional tensor; through multi-period fusion of the period feature set, the feature information of tensors with different period scales can be integrated, and the multi-period correlation of the obtained period fusion convolutional matrix is greatly enhanced. In summary, the present invention fully considers the multi-period coupling characteristics of the photovoltaic power sequence, and based on frequency-domain decoupling and multi-period fusion, can adapt to the distribution changes of non-stationary photovoltaic time series, greatly improving the prediction accuracy and stability of photovoltaic power. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a flowchart of the method of the present invention.

[0066] Figure 2 is a flowchart of the method for obtaining the decoupled two-dimensional tensor.

[0067] Figure 3 is a flowchart of the method for generating the period feature set.

[0068] Figure 4 is a flowchart of the method for obtaining the period fusion convolutional matrix.

[0069] Figure 5 is the photovoltaic power prediction curve of an embodiment of the present invention.

[0070] Figure 6 is the system architecture diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0071] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0072] Photovoltaic power prediction method based on frequency-domain decoupling and multi-cycle fusion. The flowchart is as follows Figure 1 shown, and the specific steps are as follows.

[0073] First, obtain the historical photovoltaic power data, and perform preprocessing on the historical photovoltaic power data, including missing value processing, outlier interpolation processing, and normalization processing, to obtain a photovoltaic power sequence.

[0074] Specifically, the method for preprocessing the historical photovoltaic power data is as follows:

[0075] For missing values, use the mean value of the hour before and after the missing moment and the same moment of the previous day to replace them;

[0076] Use the 3 σ sigma principle to detect outliers, and regard the values outside the error range of ( μ -3 σ , μ +3 σ ) as outliers, and treat the outliers as missing values;

[0077] Use min-max normalization to scale the photovoltaic power sequence after missing value processing and outlier interpolation processing to between [0,1]. The calculation formula is as follows:

[0078] (1),

[0079] where, is the photovoltaic power sequence after normalization processing, X min is the minimum value of the photovoltaic power sequence, X max is the maximum value of the photovoltaic power sequence.

[0080] The preprocessing obtains a photovoltaic power sequence with a certain time resolution, such as 15 minutes, 1 hour, 1 day, which can be specifically adjusted according to the requirements for prediction accuracy. For the photovoltaic power sequence obtained by preprocessing use a sliding window to intercept sample strips, and the window length is i + j , i represents the length of the input sequence, j is the length of the label value, and the sliding step is 1, so as to make full use of the data set and at the same time reduce the poor prediction effect of the present invention at a specific moment. Inside the sliding window, determine the photovoltaic power sequence input for the subsequent steps according to the actual prediction requirements.

[0081] In one embodiment of the present invention, the window length is 144, wherein i =96, j =48.

[0082] Then, the photovoltaic power sequence is converted from the time domain to the frequency domain through fast Fourier transform to obtain multiple cycle lengths of the photovoltaic power sequence, and all cycle lengths are made to be non-repetitive to obtain the selected cycle length. The one-dimensional input sequence is decoupled based on the selected cycle length to obtain a decoupled two-dimensional tensor of multiple de-redundant cycles that characterizes the multi-cycle coupling characteristics in the photovoltaic power sequence.

[0083] Specifically, the flowchart of the method for obtaining the decoupled two-dimensional tensor is as follows: Figure 2 As shown, the method for obtaining the decoupled two-dimensional tensor is:

[0084] For photovoltaic power sequence , Fast Fourier Transform is used to detect the prominent periodicity as the period length { p 1,…, p k}, the calculation formula is:

[0085] (2),

[0086] in, F represent The intensity of each frequency component in , FFT(·) and Amp(·) represent fast Fourier transform and amplitude calculation respectively, Avg(·) is the average function, L Indicates the length of the input photovoltaic power sequence; through arg Top (·) Before function selection k amplitude values, and obtain the most significant frequency with unnormalized amplitude { f 1,…, f k}, the most significant frequency { f 1,…, f k} corresponds to k The length of the cycle { p 1,…, p k};

[0087] When only the positive frequency interval is considered, formula (2) is summarized as:

[0088] (3),

[0089] use Unique (·) function ensures the uniqueness of the period length, Unique (·) Function to the front kPerform a repeatability check on the cycle lengths of the selected frequencies. If there are duplicates after rounding the cycle lengths, remove the duplicate cycle lengths with lower amplitudes, and reselect new frequencies from the frequency list sorted by amplitude until all cycle lengths are unique, as defined below:

[0090] (4),

[0091] Based on the selected cycle lengths { p’ 1 ,…, p’ k} decouple the one-dimensional input sequence The formula is:

[0092] (5),

[0093] where Padding(·) means padding 0 at the end of so that the sequence length can be divisible by p’ i ;

[0094] Through the function, decouple the one-dimensional input sequence to obtain a decoupled two-dimensional tensor k of redundancy-free cycles, which characterizes the multi-cycle coupling characteristics in the photovoltaic power sequence.

[0095] In an embodiment of the present invention, k is selected as 6.

[0096] Next, identify the cycle scales of the decoupled two-dimensional tensor. Based on the cycle scales, use an adaptive graph convolutional neural network to capture the intra-cycle trends and cross-cycle dependencies in the photovoltaic power sequence, and generate a cycle feature set.

[0097] Specifically, the method flowchart for generating the cycle feature set is as shown in Figure 3 The method for generating the cycle feature set is as follows:

[0098] First, identify the cycle scales of the decoupled two-dimensional tensor of redundancy-free cycles. Map the tensor corresponding to the k th cycle scale to a tensor with a cycle scale of n through a linear transformation, where is the number of photovoltaic decoupled sequences included in the N th cycle corresponding to the decoupled two-dimensional tensor of redundancy-free cycles. Introduce a weight matrix to capture the local features of the time series, as defined below: n where

[0099] (6),

[0100] wherein, h n and W n are learnable weight matrices of the decoupled two-dimensional tensor belonging to the n th redundancy removal period; The learnable weight matrix of the decoupled two-dimensional tensor belonging to the redundancy removal period is;

[0101] By and multiplying two trainable parameters and normalizing the weights between different nodes using the SoftMax(·) function to generate an adaptive adjacency matrix:

[0102] (7),

[0103] After obtaining the adjacency matrix n of the decoupled two-dimensional tensor of the A n th redundancy removal period, use MixHop graph convolution to capture the dependencies between period scales, fuse the decoupled two-dimensional tensors of multiple redundancy removal periods and output them to obtain a period feature set H n :

[0104] (8),

[0105] wherein, σ (·) is the Sigmoid activation function, P is an integer hyperparameter containing a set of adjacent spacings,( A n ) j represents the A n th power of the adjacency matrix j , and [·] is the intermediate output in each iteration of column-level connection.

[0106] Next, construct a period fusion matrix based on the long-period scale tensor and the short-period scale tensor in the period feature set, calculate the granular tensor of the period fusion matrix, and use a one-dimensional convolutional layer to compress the granularity information belonging to different time series to obtain a period fusion convolutional matrix.

[0107] The method flowchart for obtaining the period fusion convolutional matrix is as shown in Figure 4 , and the method for obtaining the period fusion convolutional matrix is:

[0108] First, linearly reconstruct the period tensors { H 1 ,…, H k} to obtain a period fusion matrixG n :

[0109] (9),

[0110] Then, use an average pooling layer in the Scale dimension:

[0111] (10),

[0112] Finally, G’ n Unfold in the time series dimension to obtain the granular tensor of the periodic fusion matrix, and use a one-dimensional convolutional layer to compress the granularity information belonging to different time series to obtain the periodic fusion convolutional matrix C out Defined as:

[0113] (11).

[0114] Finally, based on the periodic fusion convolutional matrix, use the Transformer Encoder architecture to establish a photovoltaic power prediction model, generate long-term time series feature encodings, and map the long-term time series feature encodings to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain the photovoltaic power prediction value.

[0115] Specifically, the method for obtaining the photovoltaic power prediction value is:

[0116] Use the key modules including the multi-head attention mechanism and the feed-forward neural network in the Transformer Encoder architecture to model the periodic fusion convolutional matrix C out The definition of the multi-head attention mechanism is:

[0117] (12),

[0118] where head(·) is an independent attention head, d k represents the dimension of the key vector, Q , K , V represent the query, key, and value matrices respectively, and are defined as: ;

[0119] By introducing the weight matrix W o perform a linear transformation on the concatenated multiple independent attention heads head(·) to obtain the cross-period attention embedding of the periodic fusion convolutional matrix MSA ( Q , K ,V ) is defined as:

[0120] (13),

[0121] Embed cross - cycle attention MSA ( Q , K , V ) and the cycle fusion convolution matrix C out Feed them into a multi - layer linear feed - forward neural network FFN (·) for feature transformation to generate long - term time - series feature encoding ŷ en , during which batch normalization BatchNorm (·) is used to normalize the input data and is defined as:

[0122] (14),

[0123] Finally, through the linear layer Linear (·) map the long - term time - series feature encoding ŷ en to the target prediction interval to obtain the final photovoltaic power prediction value ŷ pre ; where Flatten (·) is used to flatten the two - dimensional tensor and is defined as:

[0124] (15).

[0125] In an embodiment of the present invention, the photovoltaic power prediction curve is as Figure 5 shown.

[0126] The photovoltaic power prediction system based on frequency - domain decoupling and multi - cycle fusion, the system architecture diagram is as Figure 6 shown, and it is composed of a photovoltaic power sequence data acquisition module, a frequency - domain decoupling module, a cycle feature set generation module, a multi - cycle fusion module, and a photovoltaic power prediction module.

[0127] The photovoltaic power sequence data acquisition module is used to obtain the historical photovoltaic power data and perform pre - processing on the historical photovoltaic power data including missing value processing, outlier interpolation processing, and normalization processing to obtain the photovoltaic power sequence.

[0128] The frequency - domain decoupling module is used to convert the photovoltaic power sequence from the time domain to the frequency domain through the fast Fourier transform, obtain multiple cycle lengths of the photovoltaic power sequence, and make all cycle lengths non - repeating to obtain the selected cycle lengths. Based on the selected cycle lengths, decouple the one - dimensional input sequence to obtain a decoupled two - dimensional tensor of multiple non - redundant cycles characterizing the multi - cycle coupling characteristics in the photovoltaic power sequence.

[0129] A periodic feature set generation module, which is used to identify the periodic scales of the decoupled two-dimensional tensor, and based on the periodic scales, use an adaptive graph convolutional neural network to capture the intra-period trends and cross-period dependencies in the photovoltaic power sequence, and generate a periodic feature set.

[0130] A multi-period fusion module, which is used to construct a period fusion matrix according to the long-period scale tensor and short-period scale tensor in the periodic feature set, calculate the granular tensors of the period fusion matrix, and use a one-dimensional convolutional layer to compress the granularity information belonging to different time series to obtain a period fusion convolutional matrix.

[0131] A photovoltaic power prediction module, which is used to establish a photovoltaic power prediction model based on the period fusion convolutional matrix by adopting a Transformer Encoder architecture, generate long-term time series feature encodings, and map the long-term time series feature encodings to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain the photovoltaic power prediction value.

[0132] The specific implementation manners of each module in this system are the same as those described in the above method, and will not be elaborated here.

[0133] The present invention also provides an electronic device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor realizes the photovoltaic power prediction method based on frequency domain decoupling and multi-period fusion and the photovoltaic power prediction system based on frequency domain decoupling and multi-period fusion as described above by executing the computer instructions.

[0134] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it realizes the photovoltaic power prediction method based on frequency domain decoupling and multi-period fusion and the photovoltaic power prediction system based on frequency domain decoupling and multi-period fusion as described above.

[0135] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming languages Java, C++, Python, and the interpreted scripting language JavaScript, etc.

[0136] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, as well as the combination of flows and / or blocks in the flowchart and / or block diagram. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in multiple blocks.

[0137] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in multiple blocks.

[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in multiple blocks.

[0139] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present application.

[0140] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these changes and modifications.

Claims

1. A photovoltaic power prediction method based on frequency-domain decoupling and multi-period fusion, characterized in that Including: Obtain historical photovoltaic power data, and perform preprocessing on the historical photovoltaic power data including missing value processing, outlier interpolation processing, and normalization processing to obtain a photovoltaic power sequence; Convert the photovoltaic power sequence from the time domain to the frequency domain through fast Fourier transform to obtain multiple cycle lengths of the photovoltaic power sequence, and ensure that all cycle lengths do not repeat to obtain selected cycle lengths. Decouple the one-dimensional input sequence based on the selected cycle lengths to obtain a decoupled two-dimensional tensor of multiple redundancy-free cycles characterizing the multi-cycle coupling characteristics in the photovoltaic power sequence; Identify the cycle scale of the decoupled two-dimensional tensor, and use an adaptive graph convolutional neural network to capture the intra-cycle trend and cross-cycle dependence relationship in the photovoltaic power sequence based on the cycle scale to generate a cycle feature set; Construct a cycle fusion matrix based on the long-cycle scale tensor and short-cycle scale tensor in the cycle feature set, calculate the granular tensor of the cycle fusion matrix, and use a one-dimensional convolutional layer to compress the granularity information belonging to different time series to obtain a cycle fusion convolutional matrix; Based on the cycle fusion convolutional matrix, establish a photovoltaic power prediction model using the Transformer Encoder architecture to generate long-term time series feature encodings, and map the long-term time series feature encodings to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain the photovoltaic power prediction value.

2. The photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion according to claim 1, wherein: The preprocessing of the historical photovoltaic power data includes: For missing values, use the mean of the hour before and after the missing moment and the same moment of the previous day to replace; Use the 3σ principle to detect outliers, regard the values with errors exceeding the interval (μ - 3σ, μ + 3σ) as outliers, and treat the outliers as missing values; The photovoltaic power sequence X after missing value processing and outlier interpolation is scaled to the range of [0, 1] using min-max normalization, and its calculation formula is as follows: 1D ​ Among them, X' 1D is the photovoltaic power sequence after normalization processing, X min is the minimum value of the photovoltaic power sequence, X max is the maximum value of the photovoltaic power sequence.

3. The photovoltaic power prediction method based on frequency-domain decoupling and multi-cycle fusion according to claim 2, wherein: The preprocessing obtains a photovoltaic power sequence X' 1D With a certain time resolution, for the photovoltaic power sequence X'obtained by preprocessing 1D Use a sliding window to intercept sample strips. The window length is i + j, where i represents the length of the input sequence and j is the length of the label value. The sliding step is 1. Inside the sliding window, determine the photovoltaic power sequence input for subsequent steps according to the actual prediction requirements 4. The photovoltaic power prediction method based on frequency-domain decoupling and multi-cycle fusion according to claim 3, characterized in that: The method for obtaining the decoupled two-dimensional tensor is: For the photovoltaic power sequence The fast Fourier transform is used to detect prominent periodicities as the cycle lengths {p1, …, p k}, and the calculation formula is as follows: where F represents the intensity of each frequency component in, FFT(·) and Amp(·) represent the fast Fourier transform and amplitude calculation respectively, Avg(·) is the average function, the top k amplitude values are selected by the argTopk(·) function, and the most significant frequencies {f1, …, f k} with unnormalized amplitudes are obtained, and the most significant frequencies {f1, …, f k} correspond to the k cycle lengths {p1, …, p k}; When only considering the positive frequency interval, formula (2) summarizes to: Use the Unique(·) function to ensure the uniqueness of the cycle lengths. The Unique(·) function performs a repeatability check on the cycle lengths of the first k selected frequencies. If there are repeated cycle lengths after rounding, remove the repeated cycle lengths with lower amplitudes, and reselect new frequencies from the frequency list sorted by amplitude until all cycle lengths are unique, defined as follows: Decouple the one-dimensional input sequence based on the selected cycle lengths {p’1,…,p’ k}, and the formula is: ​ where Padding(·) means padding 0 at the end of to make the sequence length divisible by p’ i ; By decoupling the one-dimensional input sequence through a function a decoupled two-dimensional tensor of k redundancy-removed periods characterizing the multi-period coupling characteristics in the photovoltaic power sequence is obtained 5. The photovoltaic power prediction method based on frequency-domain decoupling and multi-cycle fusion according to claim 4, wherein: The method for generating the cycle feature set is: First, identify the decoupled two-dimensional tensor of the redundancy removal period for k periodic scales, and map the tensor corresponding to the n-th periodic scale among them to a tensor with a periodic scale of p’ n *N, where N is the number of photovoltaic decoupled sequences included in the n-th period corresponding to the decoupled two-dimensional tensor of the redundancy removal period. Introduce a weight matrix to capture the local features of the time series, which is defined as follows: where h n and W n are learnable weight matrices of the decoupled two-dimensional tensors belonging to the n-th redundancy removal period ; By and multiplying two trainable parameters and normalizing the weights between different nodes using the SoftMax(·) function to generate an adaptive adjacency matrix: After obtaining the adjacency matrix A of the decoupled two-dimensional tensor in the nth redundancy-removing period n MixHop graph convolution is used to capture the dependencies between period scales, and the decoupled two-dimensional tensors of multiple redundancy-removing periods are fused and output to obtain the period feature set H n : H n = σ([(A n ) j h n j∈P ) (8),​ Among them, σ(·) is the Sigmoid activation function, P is an integer hyperparameter containing a set of adjacent spacings, and (A n ) j represents the j-th power of the adjacency matrix A n , and [·] is the intermediate output during each iteration of column-level concatenation.

6. The photovoltaic power prediction method based on frequency-domain decoupling and multi-cycle fusion according to claim 5, wherein: The method for obtaining the cycle fusion convolutional matrix is: First, linearly reconstruct the periodic tensors {H 1 , …, H k} to obtain the periodic fusion matrix G n : G n = Concat(H 1 ,..., H k ) (9), Then use an average pooling layer in the Scale dimension: Finally, G’ N is unfolded in the time series dimension to obtain the granular tensor of the periodic fusion matrix, and a one-dimensional convolutional layer is used to compress the granularity information belonging to different time series to obtain the periodic fusion convolutional matrix C out is defined as: Cout = Conv1D(G N ′, K) (11).

7. The photovoltaic power prediction method based on frequency-domain decoupling and multi-period fusion according to claim 6, wherein: The method for obtaining the photovoltaic power prediction value is: Model the periodic fusion convolution matrix C using key modules including the multi-head attention mechanism and the feed-forward neural network in the Transformer Encoder architecture out The multi-head attention mechanism is defined as follows: where head(·) is an independent attention head, d k represents the dimension of the key vector, and Q, K, and V represent the query, key, and value matrices respectively, which are defined as: Q = C out W Q , K = C out W K , V = C out W V ; By introducing the weight matrix W o Perform a linear transformation on multiple independent attention heads head(·) after splicing to obtain the cross-period attention embedding MSA(Q, K, V) of the periodic fusion convolution matrix, which is defined as: MSA(Q, K, V) = Concat(head1,..., head h )W o (13), Embed cross-cycle attention into MSA(Q, K, V) and the cycle fusion convolution matrix C out Feed it into the multi-layer linear feed-forward neural network FFN(·) for feature transformation to generate long-term time series feature encoding During this process, batch normalization BatchNorm(·) is used to normalize the input data, which is defined as: Finally, the long-term time series features are encoded through the linear layer Linear(·) and mapped to the target prediction interval to obtain the final photovoltaic power prediction value where Flatten(·) is used to flatten the two-dimensional tensor and is defined as:

8. A photovoltaic power prediction system based on frequency-domain decoupling and multi-cycle fusion for implementing the method according to any one of claims 1 to 7, characterized in that: Including a photovoltaic power sequence data acquisition module, a frequency domain decoupling module, a cycle feature set generation module, a multi-cycle fusion module, and a photovoltaic power prediction module; The photovoltaic power sequence data acquisition module is used to obtain historical photovoltaic power data, and perform preprocessing on the historical photovoltaic power data including missing value processing, outlier interpolation processing, and normalization processing to obtain a photovoltaic power sequence; The frequency-domain decoupling module is used to convert the photovoltaic power sequence from the time domain to the frequency domain through fast Fourier transform, obtain multiple cycle lengths of the photovoltaic power sequence, and ensure that all cycle lengths are non-repetitive, so as to obtain the selected cycle length. Based on the selected cycle length, the one-dimensional input sequence is decoupled to obtain a decoupled two-dimensional tensor of multiple redundancy-removed cycles characterizing the multi-cycle coupling characteristics in the photovoltaic power sequence; The cycle feature set generation module is used to identify the cycle scale of the decoupled two-dimensional tensor, and capture the intra-cycle trend and cross-cycle dependence relationships in the photovoltaic power sequence by using an adaptive graph convolutional neural network based on the cycle scale, so as to generate a cycle feature set; The multi-cycle fusion module is used to construct a cycle fusion matrix according to the long-cycle scale tensor and the short-cycle scale tensor in the cycle feature set, calculate the granular tensor of the cycle fusion matrix, and use a one-dimensional convolutional layer to compress the granularity information belonging to different time series to obtain a cycle fusion convolutional matrix; The photovoltaic power prediction module is used to establish a photovoltaic power prediction model based on the cycle fusion convolutional matrix by adopting a Transformer Encoder architecture, generate long-term time series feature encoding, and map the long-term time series feature encoding to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain the photovoltaic power prediction value.

9. An electronic device, characterized in that, Including: A memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor realizes the photovoltaic power prediction method based on frequency-domain decoupling and multi-cycle fusion according to any one of claims 1-7 by executing the computer instructions.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it realizes the photovoltaic power prediction method based on frequency-domain decoupling and multi-cycle fusion according to any one of claims 1-7.

Images