Photovoltaic power prediction method and system based on frequency domain decoupling and multi-cycle fusion

Abstract

The invention provides a photovoltaic power prediction method and system based on frequency domain decoupling and multi-cycle fusion. The photovoltaic power prediction method comprises the following steps: firstly, acquiring photovoltaic power historical data and preprocessing the photovoltaic power historical data to obtain a photovoltaic power sequence; decoupled two-dimensional tensors of a plurality of redundancy elimination periods representing the multi-period coupling characteristic in the photovoltaic power sequence are obtained through frequency domain decoupling; on the basis of the period scale of the decoupling two-dimensional tensor, utilizing an adaptive graph convolutional neural network to capture the intra-period trend and the cross-period dependency relationship in the photovoltaic power sequence, and generating a period feature set; performing multi-period fusion on the period feature set to obtain a period fusion convolution matrix; and finally, based on the periodic fusion convolution matrix, a Transform Encoder architecture is adopted to establish a photovoltaic power prediction model, and photovoltaic power is predicted. The method fully considers the multi-period coupling characteristic of the photovoltaic power sequence, can adapt to the distribution change of the non-stationary photovoltaic time sequence based on frequency domain decoupling and multi-period fusion, and greatly improves the prediction precision and stability of the photovoltaic power.

Description

Technical Field

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

[0002] In order 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 and solar grid-connected power generation are increasingly affecting the power system. However, solar energy is an intermittent energy source. Changes in meteorological factors will cause fluctuations in photovoltaic output power and increase 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, the difficulty of dispatching and managing the power system is further increased. Accurate prediction of photovoltaic power can greatly improve the absorption capacity of photovoltaics, which is of great significance to the dispatching of power systems in a short period of time.

[0003] At present, photovoltaic power prediction methods are mainly based on mathematical statistical models or machine learning algorithms. The current methods are simple to operate and respond quickly, but their performance in the field of prediction regression is strongly positively correlated with the quantity and quality of data. However, photovoltaic power series data usually show different intra-sequence and inter-sequence correlations, and the current methods mainly focus on the time variation modeling of one-dimensional photovoltaic power series, ignoring the inherent complexity and intertwined dependencies in photovoltaic data. At the same time, short-term fluctuations, ups and downs of 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 of different time scales between multiple time series, resulting in low accuracy and poor adaptability of current photovoltaic power prediction methods. Summary of the invention

[0004] In order to solve the problems of low accuracy and poor adaptability of the current photovoltaic power prediction method described in the background technology, the present invention provides a photovoltaic power prediction method and system based on frequency domain decoupling and multi-cycle fusion.

[0005] The method of the present invention comprises: Acquire photovoltaic power history data, and perform preprocessing on the photovoltaic power history data including missing value processing, abnormal value interpolation processing and normalization processing to obtain photovoltaic power series; The photovoltaic power sequence is converted from the time domain to the frequency domain by fast Fourier transform, and multiple cycle lengths of the photovoltaic power sequence are obtained, and all cycle lengths are made to be non-repetitive, so as to obtain a selected cycle length, and 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 characterize the multi-cycle coupling characteristics in the photovoltaic power sequence; The periodic scale of the decoupled two-dimensional tensor is identified, and based on the periodic scale, an adaptive graph convolutional neural network is used to capture the intra-period trend and cross-period dependency in the photovoltaic power series to generate a periodic feature set. A periodic fusion matrix is ​​constructed based on the long-period scale tensor and the short-period scale tensor in the periodic feature set, the particle tensor of the periodic fusion matrix is ​​calculated, and a one-dimensional convolution layer is used to compress the particle size information belonging to different time series to obtain a periodic fusion convolution matrix. Based on the periodic fusion convolution matrix, the Transformer Encoder architecture is used to establish a photovoltaic power prediction model to generate long-term time series feature coding. The long-term time series feature coding is mapped to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain the photovoltaic power prediction value.

[0006] Furthermore, the preprocessing of the photovoltaic power historical data includes: For missing values, the mean of the values ​​one hour before and after the missing time and the same time of the previous day are used instead; Use 3 σ Principle detects outliers and sets the error exceeding ( μ -3 σ , μ +3 σ ) interval is considered as an outlier, and the outlier is treated as a missing value; The photovoltaic power series after missing value processing and outlier interpolation processing is completed using minimum-maximum normalization Scaling to [0,1], the calculation formula is as follows: (1), in, is the normalized photovoltaic power series, X min is the minimum value of the photovoltaic power series, X max is the maximum value of the photovoltaic power sequence.

[0007] Furthermore, the photovoltaic power sequence obtained by the preprocessing is With a certain time resolution, the photovoltaic power sequence obtained by preprocessing Use a sliding window to intercept the sample strip, the window length is i + j , i represents the length of the input sequence, j is the length of the label value, the sliding step is 1, and within the sliding window, the photovoltaic power sequence input in the subsequent steps is determined according to the actual forecast demand. .

[0008] Furthermore, the method for obtaining the decoupled two-dimensional tensor is: 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: (2), 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 cycle length { p 1 ,…, p k}; When only the positive frequency interval is considered, formula (2) is summarized as: (3), use Unique (·) function ensures the uniqueness of the period length, Unique (·) Function to the front k The period lengths of the selected frequencies are checked for repeatability. If there are repetitions after rounding the period lengths, the repetitive period lengths with lower amplitudes are removed and new frequencies are reselected from the frequency list sorted by amplitude until all period lengths are unique, as defined below: (4), Based on the selected period length { p’ 1 ,…, p’ k}For a one-dimensional input sequence Decoupling is performed, and the formula is: (5), Where Padding(·) indicates that The end of the sequence is filled with 0 so that the sequence length can be p’ i divisibility; pass Function for one-dimensional input sequence Decoupling, we get the multi-cycle coupling characteristics in the photovoltaic power sequence. k Decoupled 2D tensor with redundancy-free cycles .

[0009] Furthermore, the method for generating the periodic feature set is: First, a decoupled 2D tensor is identified to remove redundant cycles of k The period scale will correspond to the n The tensor of periodic scale is mapped to the periodic scale by linear transformation On the tensor of N The decoupled two-dimensional tensor corresponding to the redundant cycle n The number of photovoltaic decoupling sequences contained in a period is calculated, and a weight matrix is ​​introduced to capture the local characteristics of the time series, which is defined as follows: (6), in, h n and W n For the n Decoupled 2D tensor with redundancy-free cycles The learnable weight matrix of pass and The two trainable parameters are multiplied together, and the weights between different nodes are normalized using the SoftMax(·) function to generate an adaptive adjacency matrix: (7), In obtaining the n Adjacency matrix of decoupled 2D tensor with redundancy cycles A n Finally, MixHop graph convolution is used to capture the dependencies between period scales, and the decoupled two-dimensional tensors of multiple de-redundant periods are fused and output to obtain the period feature set. H n : (8), in, σ (·) is the Sigmoid activation function, P is an integer hyperparameter containing a set of neighbor distances, ( A n ) jRepresents the adjacency matrix A n of j The power, [·] is the intermediate output during each iteration of the column-level connection.

[0010] Furthermore, the method for obtaining the periodic fusion convolution matrix is: First, the period tensor { H 1 ,…, H k} Linear reconstruction to obtain the periodic fusion matrix G n : (9), Then use an average pooling layer on the Scale dimension: (10), Finally G’ N Expand it in the time series dimension to get the particle tensor of the periodic fusion matrix, and use a one-dimensional convolution layer to compress the granularity information belonging to different time series to get the periodic fusion convolution matrix C out Defined as: (11).

[0011] Furthermore, the method for obtaining the photovoltaic power prediction value is: Use the key modules in the Transformer Encoder architecture, including the multi-head attention mechanism and the feed-forward neural network, to periodically fuse the convolutional matrix C out For modeling, the multi-head attention mechanism is defined as: (12), Among them, head(·) is an independent attention head, d k represents the dimension of the key vector, Q , K , V Denote query, key, and value matrices respectively, and are defined as: ; By introducing the weight matrix W o Perform linear transformation on the concatenated multiple independent attention heads (·) to obtain the cross-period attention embedding of the periodic fused convolution matrix MSA ( Q , K , V ), defined as: (13), Embedding cross-cycle attention MSA ( Q , K , V ) and the periodic fused convolution matrix C out Passed into a multi-layer linear feedforward neural network FFN (·) Perform feature conversion to generate long-term temporal feature encoding ŷ en In this process, batch normalization BatchNorm (·) is used to normalize the input data and is defined as: (14), Finally, through the linear layer Linear (·) Encode long-term temporal features ŷ en Map to the target prediction interval to obtain the final photovoltaic power prediction value ŷ pre ;in, Flatten (·) is used to flatten a two-dimensional tensor and is defined as: (15).

[0012] The present invention also proposes a photovoltaic power prediction system based on frequency domain decoupling and multi-cycle fusion, including a photovoltaic power sequence data acquisition module, a frequency domain decoupling module, a periodic feature set generation module, a multi-cycle fusion module and a photovoltaic power prediction module.

[0013] The photovoltaic power sequence data acquisition module is used to acquire photovoltaic power history data, and perform preprocessing on the photovoltaic power history data including missing value processing, abnormal value interpolation processing and normalization processing to obtain a photovoltaic power sequence.

[0014] 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 make all cycle lengths non-repetitive to obtain a selected cycle length, decouple the one-dimensional input sequence based on the selected cycle length, and obtain a decoupled two-dimensional tensor of multiple de-redundant cycles that characterize the multi-cycle coupling characteristics in the photovoltaic power sequence.

[0015] The cycle feature set generation module is used to identify the cycle scale of the decoupled two-dimensional tensor, and to capture the intra-cycle trend and cross-cycle dependency in the photovoltaic power sequence by using an adaptive graph convolutional neural network based on the cycle scale to generate a cycle feature set.

[0016] The multi-period fusion module is used to construct a periodic fusion matrix based on the long-period scale tensor and the short-period scale tensor in the periodic feature set, calculate the particle tensor of the periodic fusion matrix, and use a one-dimensional convolution layer to compress the granularity information belonging to different time series to obtain a periodic fusion convolution matrix.

[0017] The photovoltaic power prediction module is used to establish a photovoltaic power prediction model based on a periodic fusion convolution matrix and a Transformer Encoder architecture, generate long-term time series feature coding, map the long-term time series feature coding to a target prediction interval through a linear layer of the photovoltaic power prediction model, and obtain a photovoltaic power prediction value.

[0018] The present invention also proposes an electronic device, comprising: a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to implement the photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion as described above.

[0019] The present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion as described above is implemented.

[0020] Compared with the prior art, the present invention first obtains photovoltaic power historical data and performs preprocessing to obtain a photovoltaic power sequence; then obtains a decoupled two-dimensional tensor of multiple de-redundant cycles that characterizes the multi-cycle coupling characteristics in the photovoltaic power sequence through frequency domain decoupling; then uses an adaptive graph convolutional neural network based on the period scale of the decoupled two-dimensional tensor to capture the intra-cycle trend and cross-cycle dependency in the photovoltaic power sequence, and generates a period feature set; then obtains a period fusion convolution matrix by multi-cycle fusion of the period feature set; finally, based on the period fusion convolution matrix, a photovoltaic power prediction model is established using a TransformerEncoder architecture to predict photovoltaic power. The present invention can extract key spectral features in the photovoltaic power sequence, and the decoupled two-dimensional tensor obtained by frequency domain decoupling can characterize the multi-cycle coupling characteristics in the photovoltaic power sequence, thereby 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-cycle trend and cross-cycle dependency in the photovoltaic power sequence, so that the generated period feature set has the period key features in the decoupled two-dimensional tensor; by multi-cycle fusion of the period feature set, the feature information of tensors of different period scales can be integrated, and the multi-cycle correlation of the obtained period fusion convolution matrix is ​​greatly enhanced. In summary, the present invention fully considers the multi-cycle coupling characteristics of photovoltaic power series, and based on frequency domain decoupling and multi-cycle 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

[0021] Figure 1 The figure is a flow chart of the method of the present invention.

[0022] Figure 2 Flowchart of a method for obtaining a decoupled two-dimensional tensor.

[0023] Figure 3 Flowchart of the method for generating a periodic feature set.

[0024] Figure 4 Flowchart of the method for obtaining a periodic fused convolution matrix.

[0025] Figure 5 This is a photovoltaic power prediction curve according to an embodiment of the present invention.

[0026] Figure 6 It is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0027] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with 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.

[0028] Photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion, the flow chart is as follows Figure 1 As shown, the specific steps are described as follows.

[0029] Firstly, the historical photovoltaic power data is obtained and preprocessed including missing value processing, outlier interpolation processing and normalization processing to obtain the photovoltaic power series.

[0030] Specifically, the method for preprocessing photovoltaic power historical data is: For missing values, the mean of the values ​​one hour before and after the missing time and the same time of the previous day are used instead; Use 3 σ Principle detects outliers and sets the error exceeding ( μ -3 σ , μ +3 σ ) interval is considered as an outlier, and the outlier is treated as a missing value; The photovoltaic power series after missing value processing and outlier interpolation processing is completed using minimum-maximum normalization Scaling to [0,1], the calculation formula is as follows: (1), in, is the normalized photovoltaic power series, X min is the minimum value of the photovoltaic power series, X max is the maximum value of the photovoltaic power sequence.

[0031] Preprocessing to obtain photovoltaic power series It has a certain time resolution, such as 15 minutes, 1 hour, or 1 day, which can be adjusted according to the demand for prediction accuracy. Use a sliding window to intercept the sample strip, 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 reduce the situation where the prediction effect of the present invention is poor at a specific moment. Within the sliding window, the photovoltaic power sequence input in the subsequent step is determined according to the actual prediction requirements. .

[0032] In one embodiment of the present invention, the window length is 144, wherein i =96, j =48.

[0033] 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.

[0034] 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: 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: (2), 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 cycle length { p 1 ,…, p k}; When only the positive frequency interval is considered, formula (2) is summarized as: (3), use Unique (·) function ensures the uniqueness of the period length, Unique (·) Function to the front k The period lengths of the selected frequencies are checked for repeatability. If there are repetitions after rounding the period lengths, the repetitive period lengths with lower amplitudes are removed and new frequencies are reselected from the frequency list sorted by amplitude until all period lengths are unique, as defined below: (4), Based on the selected period length { p’ 1 ,…, p’ k}For a one-dimensional input sequence Decoupling is performed, and the formula is: (5), Where Padding(·) indicates that The end of the sequence is filled with 0 so that the sequence length can be p’ i divisibility; pass Function for one-dimensional input sequence Decoupling, we get the multi-cycle coupling characteristics in the photovoltaic power sequence. k Decoupled 2D tensor with redundancy-free cycles .

[0035] In one embodiment of the present invention, select k =6.

[0036] Next, the periodic scale of the decoupled two-dimensional tensor is identified, and an adaptive graph convolutional neural network is used based on the periodic scale to capture the intra-period trend and cross-period dependency in the photovoltaic power series to generate a periodic feature set.

[0037] Specifically, the flow chart of the method for generating a periodic feature set is as follows: Figure 3 As shown in Figure 2, the method for generating periodic feature sets is: First, a decoupled 2D tensor is identified to remove redundant cycles of k The period scale will correspond to the n The tensor of periodic scale is mapped to the periodic scale by linear transformation On the tensor of N The decoupled two-dimensional tensor corresponding to the redundant cycle n The number of photovoltaic decoupling sequences contained in a period is calculated, and a weight matrix is ​​introduced to capture the local characteristics of the time series, which is defined as follows: (6), in, h n and W n For the n Decoupled 2D tensor with redundancy-free cycles The learnable weight matrix of pass and The two trainable parameters are multiplied together, and the weights between different nodes are normalized using the SoftMax(·) function to generate an adaptive adjacency matrix: (7), In obtaining the n Adjacency matrix of decoupled 2D tensor with redundancy cycles A n Finally, MixHop graph convolution is used to capture the dependencies between period scales, and the decoupled two-dimensional tensors of multiple de-redundant periods are fused and output to obtain the period feature set. H n : (8), in, σ (·) is the Sigmoid activation function, P is an integer hyperparameter containing a set of neighbor distances, ( A n ) j Represents the adjacency matrix A n of j The power, [·] is the intermediate output during each iteration of the column-level connection.

[0038] Next, a periodic fusion matrix is ​​constructed based on the long-period scale tensor and the short-period scale tensor in the periodic feature set, the particle tensor of the periodic fusion matrix is ​​calculated, and a one-dimensional convolutional layer is used to compress the granularity information belonging to different time series to obtain the periodic fusion convolution matrix.

[0039] The flow chart of the method for obtaining the periodic fusion convolution matrix is ​​as follows Figure 4 As shown, the method for obtaining the periodic fusion convolution matrix is: First, the period tensor { H 1 ,…, H k} Linear reconstruction to obtain the periodic fusion matrix G n : (9), Then use an average pooling layer on the Scale dimension: (10), Finally G’ n Expand it in the time series dimension to get the particle tensor of the periodic fusion matrix, and use a one-dimensional convolution layer to compress the granularity information belonging to different time series to get the periodic fusion convolution matrix C out Defined as: (11).

[0040] Finally, based on the periodic fusion convolution matrix, the Transformer Encoder architecture is used to establish a photovoltaic power prediction model to generate long-term time series feature coding. The long-term time series feature coding is mapped to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain the photovoltaic power prediction value.

[0041] Specifically, the method for obtaining the photovoltaic power prediction value is: Use the key modules in the Transformer Encoder architecture, including the multi-head attention mechanism and the feed-forward neural network, to periodically fuse the convolutional matrix C out For modeling, the multi-head attention mechanism is defined as: (12), Among them, head(·) is an independent attention head, d k represents the dimension of the key vector, Q , K , V Denote query, key, and value matrices respectively, and are defined as: ; By introducing the weight matrix W o Perform linear transformation on the concatenated multiple independent attention heads (·) to obtain the cross-period attention embedding of the periodic fused convolution matrix MSA ( Q , K , V ), defined as: (13), Embedding cross-cycle attention MSA ( Q , K , V ) and the periodic fused convolution matrix C out Passed into a multi-layer linear feedforward neural network FFN (·) Perform feature conversion to generate long-term temporal feature encoding ŷ en In this process, batch normalization BatchNorm (·) is used to normalize the input data and is defined as: (14), Finally, through the linear layer Linear (·) Encode long-term temporal features ŷ enMap to the target prediction interval to obtain the final photovoltaic power prediction value ŷ pre ;in, Flatten (·) is used to flatten a two-dimensional tensor and is defined as: (15).

[0042] In one embodiment of the present invention, the photovoltaic power prediction curve is as follows: Figure 5 shown.

[0043] Photovoltaic power prediction system based on frequency domain decoupling and multi-cycle fusion, the system architecture is shown in the figure below: Figure 6 As shown, it consists of a photovoltaic power series 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.

[0044] The photovoltaic power sequence data acquisition module is used to acquire photovoltaic power history data, and pre-process the photovoltaic power history data including missing value processing, abnormal value interpolation processing and normalization processing to obtain the photovoltaic power sequence.

[0045] 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 make all cycle lengths non-repetitive to obtain the selected cycle length, decouple the one-dimensional input sequence based on the selected cycle length, and obtain a decoupled two-dimensional tensor of multiple de-redundant cycles that characterize the multi-cycle coupling characteristics in the photovoltaic power sequence.

[0046] The periodic feature set generation module is used to identify the periodic scale of the decoupled two-dimensional tensor, and to capture the intra-period trend and cross-period dependency in the photovoltaic power series based on the periodic scale using an adaptive graph convolutional neural network to generate a periodic feature set.

[0047] The multi-period fusion module is used to construct a periodic fusion matrix based on the long-period scale tensor and the short-period scale tensor in the periodic feature set, calculate the particle tensor of the periodic fusion matrix, and use a one-dimensional convolution layer to compress the granularity information belonging to different time series to obtain the periodic fusion convolution matrix.

[0048] The photovoltaic power prediction module is used to establish a photovoltaic power prediction model based on the periodic fusion convolution matrix and the Transformer Encoder architecture to generate long-term time series feature coding. The long-term time series feature coding is mapped to the target prediction interval through the linear layer of the photovoltaic power prediction model to obtain the photovoltaic power prediction value.

[0049] The specific implementation method of each module in the system is consistent with that described in the above method and will not be repeated here.

[0050] The present invention also proposes an electronic device, comprising: a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to implement the photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion and the photovoltaic power prediction system based on frequency domain decoupling and multi-cycle fusion as described above.

[0051] The present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion and the photovoltaic power prediction system based on frequency domain decoupling and multi-cycle fusion as described above.

[0052] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The scheme in the embodiments of the present application may be implemented in various computer languages, for example, object-oriented programming languages ​​Java, C++, Python, and literal scripting languages ​​JavaScript, etc.

[0053] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0054] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0055] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0056] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0057] 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 equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion, characterized in that: include: Acquire photovoltaic power history data, and perform preprocessing on the photovoltaic power history data including missing value processing, abnormal value interpolation processing and normalization processing to obtain photovoltaic power series; The photovoltaic power sequence is converted from the time domain to the frequency domain by fast Fourier transform, and multiple cycle lengths of the photovoltaic power sequence are obtained, and all cycle lengths are made to be non-repetitive, so as to obtain a selected cycle length, and 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 characterize the multi-cycle coupling characteristics in the photovoltaic power sequence; The periodic scale of the decoupled two-dimensional tensor is identified, and based on the periodic scale, an adaptive graph convolutional neural network is used to capture the intra-period trend and cross-period dependency in the photovoltaic power series to generate a periodic feature set. A periodic fusion matrix is ​​constructed based on the long-period scale tensor and the short-period scale tensor in the periodic feature set, the particle tensor of the periodic fusion matrix is ​​calculated, and a one-dimensional convolution layer is used to compress the particle size information belonging to different time series to obtain a periodic fusion convolution matrix. Based on the periodic fusion convolution matrix, the Transformer Encoder architecture is used to establish a photovoltaic power prediction model to generate long-term time series feature coding. The long-term time series feature coding is mapped 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 is characterized in that: The preprocessing of photovoltaic power historical data includes: For missing values, the mean of the values ​​one hour before and after the missing time and the same time of the previous day are used instead; Use 3 σ Principle detects outliers and sets the error exceeding ( μ -3 σ , μ +3 σ ) interval is considered as an outlier, and the outlier is treated as a missing value; The photovoltaic power series after missing value processing and outlier interpolation processing is completed using minimum-maximum normalization Scaling to [0,1], the calculation formula is as follows: （1）， in, is the normalized photovoltaic power series, X min is the minimum value of the photovoltaic power series, 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 is characterized in that: The photovoltaic power sequence is obtained by preprocessing With a certain time resolution, the photovoltaic power sequence obtained by preprocessing Use a sliding window to intercept the sample strip, the window length is i + j , i represents the length of the input sequence, j is the length of the label value, the sliding step is 1, and within the sliding window, the photovoltaic power sequence input in the subsequent steps is determined according to the actual forecast demand. .

4. The photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion according to claim 3 is characterized in that: The method for obtaining the decoupled two-dimensional tensor is: 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: （2）， 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 cycle length { p 1,…, p k }; When only the positive frequency interval is considered, formula (2) is summarized as: （3）， use Unique (·) function ensures the uniqueness of the period length, Unique (·) Function to the front k The period lengths of the selected frequencies are checked for repeatability. If there are repetitions after rounding the period lengths, the repetitive period lengths with lower amplitudes are removed and new frequencies are reselected from the frequency list sorted by amplitude until all period lengths are unique, as defined below: （4）， Based on the selected cycle length { p’ 1 ,…, p’ k For a one-dimensional input sequence Decoupling is performed, and the formula is: （5）， Where Padding(·) indicates that The end of the sequence is filled with 0 so that the sequence length can be p’ i divisibility; pass Function for one-dimensional input sequence Decoupling, we get the multi-cycle coupling characteristics in the photovoltaic power sequence. k Decoupled 2D tensor with redundant cycles .

5. The photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion according to claim 4 is characterized in that: The method for generating the periodic feature set is: First, a decoupled 2D tensor is identified to remove redundant cycles of k The period scale will correspond to the n The tensor of periodic scale is mapped to the periodic scale by linear transformation On the tensor of N The decoupled two-dimensional tensor corresponding to the redundant cycle n The number of photovoltaic decoupling sequences contained in a period is calculated, and a weight matrix is ​​introduced to capture the local characteristics of the time series, which is defined as follows: （6）， in, h n and W n For the n Decoupled 2D tensor with redundant cycles The learnable weight matrix of pass and The two trainable parameters are multiplied together, and the weights between different nodes are normalized using the SoftMax(·) function to generate an adaptive adjacency matrix: （7）， In obtaining the n Adjacency matrix of decoupled 2D tensor with redundancy cycles A n Finally, MixHop graph convolution is used to capture the dependencies between period scales, and the decoupled two-dimensional tensors of multiple de-redundant periods are fused and output to obtain the period feature set. H n : （8）， in, σ (·) is the Sigmoid activation function, P is an integer hyperparameter containing a set of neighbor distances, ( A n ) j Represents the adjacency matrix A n of j The power, [·] is the intermediate output of each iteration of the column-level connection.

6. The photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion according to claim 5 is characterized in that: The method for obtaining the periodic fusion convolution matrix is: First, the period tensor { H 1 ,…, H k } Linear reconstruction to obtain the periodic fusion matrix G n : （9）， Then use an average pooling layer on the Scale dimension: （10）， Finally G’ N Expand it in the time series dimension to get the particle tensor of the periodic fusion matrix, and use a one-dimensional convolution layer to compress the granularity information belonging to different time series to get the periodic fusion convolution matrix C out Defined as: （11）。 7. The photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion according to claim 6 is characterized in that: The method for obtaining the photovoltaic power prediction value is: Use the key modules in the Transformer Encoder architecture, including the multi-head attention mechanism and the feed-forward neural network, to periodically fuse the convolutional matrix C out For modeling, the multi-head attention mechanism is defined as: （12）， Among them, head(·) is an independent attention head, d k represents the dimension of the key vector, Q , K , V Denote query, key, and value matrices respectively, and are defined as: ; By introducing the weight matrix W o Perform linear transformation on the concatenated multiple independent attention heads (·) to obtain the cross-period attention embedding of the periodic fused convolution matrix MSA ( Q , K , V ), defined as: （13）， Embedding cross-cycle attention MSA ( Q , K , V ) and the periodic fused convolution matrix C out Passed into a multi-layer linear feedforward neural network FFN (·) Perform feature conversion to generate long-term temporal feature encoding ŷ en In this process, batch normalization BatchNorm (·) is used to normalize the input data and is defined as: （14）， Finally, through the linear layer Linear (·) Encode long-term temporal features ŷ en Map to the target prediction interval to obtain the final photovoltaic power prediction value ŷ pre ;in, Flatten (·) is used to flatten a two-dimensional tensor and is defined as: （15）。 8. A photovoltaic power prediction system based on frequency domain decoupling and multi-cycle fusion that implements the method described in any one of claims 1 to 7, characterized in that: It includes photovoltaic power sequence data acquisition module, frequency domain decoupling module, period feature set generation module, multi-period fusion module and photovoltaic power prediction module; The photovoltaic power sequence data acquisition module is used to acquire photovoltaic power history data, and perform preprocessing on the photovoltaic power history data including missing value processing, abnormal value 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 make all cycle lengths non-repetitive, obtain a selected cycle length, decouple the one-dimensional input sequence based on the selected cycle length, and obtain a decoupled two-dimensional tensor of multiple de-redundant cycles that characterize 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 to capture the intra-cycle trend and cross-cycle dependency in the photovoltaic power sequence by using an adaptive graph convolutional neural network based on the cycle scale to generate a cycle feature set; The multi-period fusion module is used to construct a periodic fusion matrix according to the long-period scale tensor and the short-period scale tensor in the periodic feature set, calculate the particle tensor of the periodic fusion matrix, and use a one-dimensional convolution layer to compress the particle size information belonging to different time series to obtain a periodic fusion convolution matrix; The photovoltaic power prediction module is used to establish a photovoltaic power prediction model based on a periodic fusion convolution matrix and a Transformer Encoder architecture, generate long-term time series feature coding, map the long-term time series feature coding to a target prediction interval through a linear layer of the photovoltaic power prediction model, and obtain a photovoltaic power prediction value.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor implements the photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion as described in any one of claims 1 to 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 a processor, the photovoltaic power prediction method based on frequency domain decoupling and multi-cycle fusion as described in any one of claims 1 to 7 is implemented.

Images