Photovoltaic scene intelligent generation method and system

Abstract

The present application relates to the technical field of data processing, in particular to a photovoltaic scene intelligent generation method and system, the method comprising: obtaining historical photovoltaic data; using a wavelet decomposition module to perform multi-scale frequency domain decomposition on the historical photovoltaic data to extract signal components at different frequency scales; using a frequency domain reparameterization mechanism to embed each scale component into the structure of a deep learning model to construct a frequency-aware multi-head attention mechanism; performing context modeling and structure fusion under frequency conditions in the frequency spectrum space to realize structural generation of photovoltaic scenes; and based on a frequency spectrum structure loss and statistical feature constraint, optimizing and verifying the generation result to output photovoltaic power scene sequences that conform to actual power fluctuation rules. According to the scheme of the present application, the photovoltaic output scene data engineering applicability is effectively improved, and the real situation in an industrial environment is reflected.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing. More particularly, the present application relates to a photovoltaic scene intelligent generation method and system. BACKGROUND

[0002] With large-scale access of renewable energy such as photovoltaic, the power grid puts forward higher requirements for the simulation and generation of new energy output scene. Photovoltaic output has obvious periodicity, burstiness and instability characteristics. Its power time series data often contains multiple scale fluctuation modes, such as trend changes driven by solar intensity, medium-frequency fluctuations caused by cloud shadow disturbance, and high-frequency disturbances caused by device dynamic response, etc. These time series characteristics are distributed in different frequency ranges, have non-stationary and multi-scale overlapping characteristics, and are difficult to be effectively described by traditional single-scale modeling methods.

[0003] At present, the mainstream time series generation method mainly depends on time domain modeling, which cannot explicitly distinguish and control the fluctuation behavior in different frequency bands, resulting in deviations in global trend, local disturbance and detail consistency of the generated results, and it is difficult to meet the comprehensive requirements of reality, controllability and rationality.

[0004] In recent years, deep learning models have shown strong context modeling capabilities in processing long sequence data due to their self-attention structure, and have been gradually introduced into new energy scene generation tasks. However, the standard self-attention mechanism acts on the original time domain signal, lacks explicit perception of frequency structure, and is difficult to decouple and reconstruct multi-scale information. Especially when facing photovoltaic output, which has obvious spectral hierarchy, the time domain attention structure cannot effectively extract the statistical structure and evolution law of different frequency bands, which easily leads to suppression of high-frequency details or drift of low-frequency trend, resulting in unreasonable photovoltaic scene.

[0005] Therefore, the current urgent problem to be solved is that the photovoltaic output scene data (such as power generation) has poor engineering applicability and cannot reflect the real situation. SUMMARY

[0006] To solve the above technical problems of poor engineering applicability of photovoltaic output scene data (such as power generation) and inability to reflect the real situation, the present application provides solutions in the following aspects.

[0007] In a first aspect, the present application provides a photovoltaic scene intelligent generation method, comprising: acquiring historical photovoltaic data, decomposing the historical photovoltaic data in a multi-scale frequency domain by using a wavelet decomposition module to extract signal components at different frequency scales; embedding each scale component into the structure of a deep learning model by using a frequency domain reparameterization mechanism to construct a frequency-aware multi-head attention mechanism; modeling the context under the frequency condition and fusing the structure in the frequency spectrum space to realize the structural generation of the photovoltaic scene; and optimizing and verifying the generation result based on the spectral structure loss and statistical feature constraint to output a photovoltaic power scene sequence conforming to the actual power fluctuation law.

[0008] Preferably, the wavelet decomposition module is used to decompose the historical photovoltaic data in a multi-scale frequency domain to extract signal components at different frequency scales, which comprises: performing multi-level discrete wavelet transform on the historical photovoltaic data to obtain frequency domain components at multiple scales; wherein the wavelet basis can be any one of the Daubechies, Symlet or Coiflet family; and the frequency domain components extracted at each frequency scale are used as independent input channels to construct a multi-frequency scale embedding tensor.

[0009] Preferably, the frequency-aware multi-head attention mechanism comprises: constructing multiple frequency band attention heads, each of which processes the frequency components corresponding to a set wavelet frequency band; and dynamically weighting and converging the outputs of each frequency band attention head based on a spectral gating mechanism to realize frequency-adaptive information fusion.

[0010] Preferably, the context modeling and structure fusion under the frequency condition in the frequency spectrum space comprises: converting the signal components at each frequency scale into Gaussian distribution variables and determining their mean and variance; sampling from the variational distribution at each frequency scale as input to the generation module; and introducing a learnable scale selection function to dynamically adjust the role of different frequency components in the generation.

[0011] Preferably, the optimization and verification of the generation result based on the spectral structure loss and statistical feature constraint comprises: constraining the frequency energy distribution of the generated signal by a spectral structure loss function to make it consistent with the real historical data in the main frequency band; evaluating the authenticity of the generated result in the frequency spectrum space by using a multi-scale discriminator to identify whether it has the trend and disturbance of the actual power signal; and combining the statistical quantities of the time domain mean, volatility and autocorrelation coefficient to conduct overall testing on the generated data to ensure that it meets the physical constraints and engineering characteristics.

[0012] Preferably, the historical photovoltaic data is preprocessed, including time alignment of data from different data sources, linear interpolation of missing data, and filtering and correction of outliers; and the complete sequence of the corrected historical photovoltaic data is normalized.

[0013] Preferably, the formula for calculating the signal components at different frequency scales is:

[0014]

[0015]

[0016] wherein A J (t) represents the signal component obtained by decomposition, D j (t) represents a plurality of detail components, c J,k =∫f(t)φ J,k (t)dt represents the Jth layer low-frequency approximation component coefficient, d j,k =∫f(t)ψ j,k (t)dt represents the jth layer high-frequency disturbance component coefficient, φ J,k (t) represents a low-frequency approximation base, ψ j,k (t) represents a high-frequency disturbance base.

[0017] Preferably, the output of information fusion is:

[0018]

[0019] wherein, represents a fusion vector, represents the sampling result of the disturbance latent variable at the jth scale, α j represents the structure generation weight of the jth frequency scale.

[0020] Preferably, the spectral structure loss is:

[0021]

[0022] wherein, represents the loss function of the spectral structure loss, and |F(f)| represent the amplitude spectrum generated with the real sequence at frequency f; w(f) represents a frequency weight function.

[0023] In a second aspect, the present application also provides a photovoltaic scene intelligent generation device, comprising: a processor; a memory, the memory storing computer program instructions, when the computer program instructions are executed by the processor, realizing a photovoltaic scene intelligent generation method according to one or more embodiments of the foregoing.

[0024] The application has the beneficial effects that: the application realizes unified modeling and accurate generation of photovoltaic power time series signals on trend components and disturbance components by constructing a frequency domain multi-scale structure basis with wavelet decomposition, combined with a self-attention modeling mechanism in the frequency spectrum space. Specifically, the application first decomposes historical photovoltaic power data into multiple scale frequency components through discrete wavelet transform, capturing the change structure at different levels; then uses frequency domain reparameterization technology to embed each scale frequency feature as input into a multi-head self-attention Transformer, and performs context modeling and structure coupling in the frequency spectrum space; then through frequency condition modeling and context construction method, the collaborative synthesis of multi-scale signals is realized; finally, the frequency spectrum structure loss and statistical feature constraint are introduced to constrain the generated output from the frequency distribution, consistency and statistical structure, etc., to ensure that the generated scene has high authenticity and physical reasonableness in terms of trend and fluctuation, effectively improving the engineering applicability of photovoltaic output scene data (such as power generation) and reflecting the real situation in the industrial environment. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a flow chart illustrating a photovoltaic scene intelligent generation method according to an embodiment of the application;

[0026] Figure 2 is a structural diagram of a photovoltaic scene intelligent generation device according to an embodiment of the application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0028] The specific embodiments of the application will be described in detail below with reference to the drawings.

[0029] Figure 1 is a flow chart illustrating a photovoltaic scene intelligent generation method according to an embodiment of the application;

[0030] As Figure 1As shown, at step S101, historical photovoltaic data is acquired, and the historical photovoltaic data is decomposed in a multi-scale frequency domain by using a wavelet decomposition module to extract signal components at different frequency scales. In some embodiments, the historical photovoltaic data can be subjected to multi-level discrete wavelet transform to obtain frequency domain components at multiple scales; wherein the wavelet basis can be any one of the Daubechies, Symlet or Coiflet families. The frequency domain components extracted at each frequency scale are respectively taken as independent input channels to construct a multi-frequency scale embedded tensor.

[0031] The calculation formula of the signal components at different frequency scales is:

[0032]

[0033]

[0034] wherein A J (t) represents an approximate component obtained by decomposition, D j (t) represents a plurality of detail components, c J,k =∫f(t)φ J,k (t)dt represents the Jth layer low-frequency approximate component coefficient, d j,k =∫f(t)ψ j,k (t)dt represents the jth layer high-frequency disturbance component coefficient, φ J,k (t) represents a low-frequency approximation basis, ψ j,k (t) represents a high-frequency disturbance basis.

[0035] At step S102, each scale component is embedded into the structure of the deep learning model by using a frequency domain reparameterization mechanism to construct a frequency-aware multi-head attention mechanism. In some embodiments, a multi-frequency band attention head is constructed, and each attention head processes the frequency components corresponding to a set wavelet frequency band; the outputs of each frequency band attention head are dynamically weighted and converged based on a frequency spectrum gating mechanism to realize frequency adaptive information fusion.

[0036] The output of the information fusion is:

[0037]

[0038] wherein, represents a fusion vector, represents the jth scale disturbance latent variable sampling result, α j represents the structure generation weight of the jth frequency scale.

[0039] At step S103, context modeling and structure fusion under frequency condition are performed in the spectral space to realize structural generation of the photovoltaic scene. In some embodiments, the signal component under each frequency scale can be converted into a Gaussian distribution variable, and its mean and variance are determined. Sampling is performed from the variational distribution under each frequency scale as the input of the generation module. A learnable scale selection function is introduced to dynamically adjust the strength of different frequency components in the generation.

[0040] At step S104, the generation result is optimized and verified based on the spectral structure loss and statistical feature constraint to output a photovoltaic power scene sequence conforming to the actual power fluctuation law. In some embodiments, the frequency energy distribution of the generated signal can be constrained by the spectral structure loss function to make it consistent with the real historical data in the main frequency band. The authenticity of the generated result is evaluated in the spectral space by using a multi-scale discriminator to identify whether it has the trend and disturbance of the actual power signal. The statistical quantities of the time domain mean, volatility and autocorrelation coefficient are combined to test the generated data as a whole to ensure that it meets the physical constraints and engineering characteristics.

[0041] The spectral structure loss is:

[0042]

[0043] wherein, Ls represents the loss function of the spectral structure loss, and |F(f)| represent the amplitude spectrum of the generated and real sequences at frequency f; w(f) represents the frequency weight function.

[0044] Further, before wavelet decomposition of the historical photovoltaic data, the historical photovoltaic data can also be preprocessed, and the preprocessing includes time alignment of data from different data sources, linear interpolation of missing data, and filtering and correction of abnormal values. The complete sequence of the corrected historical photovoltaic data is normalized.

[0045] Next, the scheme of the present application will be further described in detail in combination with specific embodiments.

[0046] Step one, the historical photovoltaic time series output data is preprocessed, specifically including the following steps:

[0047] First, according to the operation record of the photovoltaic power station, the historical output data is collected, with hourly or minute-by-minute as the time granularity, and the time step Δt (such as 5 minutes, 15 minutes or 30 minutes) is uniformly set. After time stamp alignment of different data sources, the original data sequence is converted into a standard time series format, denoted as P t wherein P tP(t) represents the photovoltaic output power at time step t, in kilowatt (kW).

[0048] Secondly, the normalized original power sequence is input into the preprocessing module for integrity and legality verification. For each time step t, the system checks whether there are missing values or non-physical abnormal values (such as P t <0 or P t >P max , where P max is the rated capacity of the site). After detecting missing or abnormal items, enter the repair stage.

[0049] For missing data, linear interpolation is used for completion. Let t1 and t2 be the two valid sampling time points before and after the missing point, and let P be the photovoltaic power value at the corresponding time, then the interpolation formula of the missing point P t is:

[0050]

[0051] where P t represents the interpolated power value, t is the missing point time, and t1 < t < t2.

[0052] For detected abnormal values, a statistical filtering method is used for correction. Let the historical data mean be μ and the standard deviation be σ, then take μ ± 3σ as the abnormal judgment threshold, if the power P t at a certain time step exceeds this interval, it is considered as an abnormal point. The replacement value of the abnormal point is reconstructed by using the sliding mean of the adjacent window, and the sliding mean calculation formula is:

[0053]

[0054] where P represents the smoothed power value, and the window radius k is usually taken as 2 to 4.

[0055] Finally, the normalized processing is performed on the complete sequence after repair to eliminate the interference of absolute power difference on frequency domain decomposition. Let the maximum value of the whole sequence be P max , and the minimum value be p min , then the normalized power P of each time step is calculated as:

[0056]

[0057] The normalized power sequence P is kept in the interval [0, 1] and is used as the input of subsequent multi-scale frequency domain analysis.

[0058] Step two, using wavelet decomposition module to preprocess the time series data in multi-scale frequency domain decomposition, extract different frequency scale trend and disturbance signal components, including the following steps:

[0059] First, the normalized photovoltaic power time series data sequence obtained in step one Discrete wavelet transform (Discrete Wavelet Transform, DWT) is applied to multi-scale frequency domain decomposition. DWT constructs a set of time-frequency localized orthogonal basis function family through scale function φ(t) and wavelet function ψ(t):

[0060] φ j,k (t)=2 -j / 2 ·φ(2 -j t-k)

[0061] ψ j,k (t)=2 -j / 2 ·ψ(2 -j t-k)

[0062] Wherein, j∈Z represents the scale number, k∈Z represents the translation parameter, φ(t) is the scale function (construct low frequency approximation), ψ(t) is the wavelet function (construct high frequency detail). The wavelet decomposition of any signal f(t) can be expressed as:

[0063]

[0064] Wherein, c J,k =∫f(t)φ J,k (t)dt represents the Jth layer low frequency approximation component coefficient; d j,k =∫f(t)ψ j,k (t)dt represents the jth layer high frequency disturbance component coefficient; φ J,k (t) constitutes a low frequency approximation base, ψ j,k (t) constitutes a high frequency disturbance base. Let the input sequence be the normalized photovoltaic power sequence Its decomposition result is the approximation component A J (t) and several detail components D j (t), wherein:

[0065]

[0066]

[0067] The above formula constitutes the multi-scale frequency domain representation after decomposition, and has complete reconstruction.

[0068] Then, all components {A J (t), D1(t), …, D J(t)} as multi-channel feature input, constructing unified tensor where each column is a component at one scale. This tensor constitutes the multi-frequency input basis of the Transformer model, ensuring that the subsequent network structure separates and jointly constructs the trend and disturbance in the photovoltaic output.

[0069] Finally, to construct the above-mentioned basis function family, the following three types of standard orthogonal wavelet bases are used, whose mathematical definitions are as follows: Daubechies wavelet (dbN) family is the shortest support orthogonal wavelet with N vanishing moments, which satisfies the polynomial orthogonal relationship between the wavelet function ψ(t) and the scale function φ(t), and its recursive relationship satisfies:

[0070]

[0071] where h n is the low-pass filter coefficient, g n = (-1) n h 2N-1-n is the high-pass filter coefficient, and N is usually taken as 4-6 to ensure the time-frequency resolution; Symlet wavelet (symN) family is the Daubechies wavelet with enhanced symmetry, which has the same vanishing moments but improved filter symmetry, so that the wavelet function satisfies the minimum phase symmetry condition, reduces the edge effect while maintaining the analytical ability, and satisfies the same recursive expression:

[0072]

[0073] where is the improved low-pass filter, and the design goal is to maximize the symmetry and frequency band locality; Coiflet wavelet (coifN) family has the characteristic that the scale function and the wavelet function have N vanishing moments at the same time, which has good time-frequency dual positioning ability, and its wavelet function satisfies:

[0074] ∫t k ·φ(t)dt=0

[0075] ∫t k ·ψ(t)dt=0

[0076] where k = 0, 1, …, N-1. This condition makes Coiflet suitable for extracting mixed structures with both long-term trends and transient changes, which is prominent in modeling photovoltaic power with both stationary components and disturbance scenarios. The three wavelet bases all have compact support, orthogonality and perfect reconstruction characteristics, and the wavelet base function is selected according to the actual application task: Daubechies wavelet is used in the scene where high-frequency disturbance detection needs to be strengthened, Symlet wavelet is preferred in trend smoothing modeling, and Coiflet wavelet is preferred in high-low frequency collaborative expression scene, so as to realize the controllability of signal structure and expression optimization of decomposition level structure.

[0077] Step three, embed each scale frequency component into the Transformer structure by using the frequency domain reparameterization mechanism, construct the frequency-aware multi-head self-attention mechanism, and realize the context representation under the scale condition, which includes the following steps:

[0078] First, in the encoding structure of the model, the input tensor X is constructed as follows: Construct the multi-head attention structure corresponding to the frequency segmentation. Suppose the model contains H attention heads, and each attention head h j Focus on the jth frequency scale channel in the tensor (corresponding to the jth wavelet component), then each attention head receives the input:

[0079] X (j) ∈R T×d ,j=1,2,…,J+1

[0080] where d is the feature dimension, X (j) is the embedded representation of the jth scale component after linear mapping. Each attention head performs self-attention calculation:

[0081]

[0082] where Q j = X (j) W j Q , K j = X (j) W j K , V j = X (j) W j V , W j Q , W j K , W j V ∈R d×dTransform matrix of query, key and value respectively. Each attention head independently captures the contextual dependencies within the frequency band through self-attention mechanism, and constructs a fine response to the trend segment or the disturbance segment.

[0083] Secondly, a spectral gating mechanism is introduced after the multi-head attention, which is used to evaluate the importance of each frequency scale attention output in the current generation task. Let the output of attention head j be Then the gating weight a of it is calculated j The expression of a is:

[0084]

[0085] Where, W g ∈R d , b g ∈R are trainable parameters, mean(Z j ) represents the mean of Z j The operation of taking the mean in the time dimension, and σ(·) is the Sigmoid activation function. The final fusion output is represented as:

[0086]

[0087] The above structure realizes the dynamic weighting of the outputs of different frequency band attention heads, and has the ability of frequency adaptive context aggregation.

[0088] Finally, a cross-frequency band information interaction mechanism is added to the encoding structure of the Transformer, which is specifically realized through the context communication between cross-heads. Let there be a potential dependency relationship between different frequency bands, for example, the local disturbance of the jth scale depends on the trend background information of the ith scale, a cross-attention mechanism can be introduced:

[0089]

[0090] The result is taken as an enhancement of Z j , realizing the coupling information propagation between frequency channels. This mechanism supports the deep structural fusion between different scale components within the model, which helps to capture both global trends and local mutation behaviors when generating photovoltaic power sequences, and significantly improves the learning ability and expression integrity of multi-scale dependency relationships.

[0091] Step four, frequency condition modeling and structural generation in spectral space, through scale response control function to dynamically adjust the generation contribution of different frequency bands, including the following steps:

[0092] Firstly, the frequency scale representation As a kind of conditional latent variables in frequency domain, they are input into the frequency domain variational inference to conduct statistical modeling. The representation Z j ∈R T ×d is regarded as a sample subject to conditional Gaussian distribution , and the variational prior is constructed as follows:

[0093]

[0094] where z∈R d represents the latent feature vector in the frequency scale; μ j ∈R d is the mean vector in the jth frequency band; σ is the corresponding variance vector; diag(·) represents the construction of a diagonal covariance matrix. The variational modeling process is realized by a neural network, where μ j = f μ (Z j ), are two fully connected transformation modules, respectively, and the statistical feature structure of different scales is learned.

[0095] Then, in the structural generation stage, random sampling is performed from the above frequency domain Gaussian distribution to obtain the intermediate latent variable with frequency disturbance characteristics as the input basis of the generation module. The sampling process is realized by the reparameterization technique to ensure that the model is derivable during back propagation:

[0096]

[0097] where ∈∈R d represents a noise sample from a standard normal distribution (I is the identity matrix), ⊙ represents element-wise multiplication, is the disturbance type latent variable sampling result in the jth scale.

[0098] Finally, a scale response control function is introduced to adjust the contribution strength of different frequency bands to the final generated output. The control function is realized by a learnable module g(·), and the fusion weight α j ∈[0, 1] of each scale component under the current context condition is calculated, which is expressed as:

[0099]

[0100] where W α ∈R d , b α ∈R are trainable parameters, and α j represents the structural generation weight of the jth frequency scale. The final structural generation feature vector is:

[0101]

[0102] The fusion vector The dynamic generation of multi-scale frequency components under different context conditions is comprehensively expressed, and adaptive coupling of trend structure (low frequency) and disturbance component (high frequency) is realized.

[0103] Step five, based on the spectrum structure loss and statistical feature constraint, the generation result is optimized and verified, including the following steps:

[0104] Firstly, the spectral structure loss function is used to constrain the energy distribution of the generated photovoltaic power sequence in the frequency space, to ensure its structural consistency in the main frequency band and the adjacent frequency band. Let the generated sequence be The real historical sequence is P t Discrete Fourier transform (DFT) is performed on both to obtain the spectral representation:

[0105]

[0106] Among them, denotes the Fourier transform operation, and f is the frequency variable. Then the spectral structure loss function is defined as:

[0107]

[0108] Among them: |F(f)| represents the amplitude spectrum of the generated and real sequences at frequency f; w(f)∈[0,1] is the frequency weight function, which is used to emphasize the main frequency band (such as the daily cycle, cloud shadow disturbance bandwidth); the loss function measures the amplitude difference of the frequency domain structure, and constrains the model to maintain reasonable frequency distribution in the generation process.

[0109] Secondly, a multi-scale discriminator structure is introduced to evaluate the authenticity of the generated results in the frequency space. The discriminator is composed of multiple sub-discriminators Each sub-discriminator focuses on a specific frequency scale component D j (t) or A J (t). Each discriminator discriminates the input spectral segment and the corresponding real component F j (f), and its loss can be defined as:

[0110]

[0111] Among them, D j (·) is the output of the jth frequency band discriminator; λ is the weight of the gradient penalty term; The intermediate spectrum sample is interpolated between the real and generated samples; the discriminator is used to simultaneously identify whether the trend and disturbance structure are consistent with the real photovoltaic scene.

[0112] Finally, at the time domain level of the generated sequence, a set of statistical quantities are introduced to verify the overall consistency of the output results, including volatility and autocorrelation structure indicators. The first-order autocorrelation coefficient is used to evaluate the instantaneous fluctuation rate. The sequence stationarity and trend consistency are evaluated. If the generated data falls within the statistical tolerance interval set based on the real data in the above indicators, it is determined that it has good engineering usability and reliability; otherwise, the model will optimize and adjust the generated structure through the loss function joint feedback mechanism.

[0113] The present application realizes the intelligent process of extracting multi-scale structure features from historical data and guiding generation through the physical scale extraction of the wavelet decomposition module, the context feature learning of the frequency control type structure generation and the spectrum-statistics double constraint optimization mechanism. The generated photovoltaic time series scene data not only retains the long-term trend and short-time disturbance characteristics in the original data, but also has good frequency consistency, physical rationality and statistical stability, which can effectively support the prediction modeling and scheduling decision process under the new energy scene, and significantly improve the robustness and efficiency of system operation.

[0114] Figure 2 is a structural diagram of a photovoltaic scene intelligent generation device according to an embodiment of the present application.

[0115] The present application also provides a photovoltaic scene intelligent generation device. As shown in Figure 2 The device includes a processor and a memory, and the memory stores computer program instructions, which, when executed by the processor, implement a photovoltaic scene intelligent generation method according to one or more embodiments described above.

[0116] The device also includes a communication bus and a communication interface and other components familiar to those skilled in the art, the settings and functions of which are known in the art, and therefore will not be described here.

[0117] Although the present application has shown and described several embodiments of the present application, it will be apparent to those skilled in the art that such embodiments are provided only by way of example. Those skilled in the art will think of many changes, changes and alternatives without departing from the idea and spirit of the present application. It should be understood that various alternatives to the embodiments of the present application described herein can be employed in practicing the present application.

Claims

1. A photovoltaic scene intelligent generation method, characterized in that, The method comprises: Obtaining historical photovoltaic data, and performing multi-scale frequency domain decomposition on the historical photovoltaic data by using a wavelet decomposition module to extract signal components at different frequency scales; the signal components extracted at each frequency scale are taken as independent input channels respectively to construct multi-frequency scale components; Embedding each frequency scale component into the structure of a deep learning model to construct a frequency-aware multi-head attention mechanism, which comprises: constructing multi-frequency band attention heads, each of which processes frequency components corresponding to a set wavelet frequency band; and dynamically weighting and converging the outputs of each frequency band attention head based on a frequency spectrum gating mechanism to realize frequency-adaptive information fusion; Performing context modeling under frequency conditions in the frequency spectrum space to realize structural generation of photovoltaic scenes, which comprises: converting the signal components at each frequency scale into Gaussian distribution variables and determining their mean and variance; sampling from the variational distribution at each frequency scale to obtain intermediate latent variables with frequency disturbance characteristics as input of a generation module; and introducing a learnable scale selection function to dynamically adjust the role intensity of different frequency components in generation; Optimizing and verifying the generation result based on a frequency spectrum structure loss and a statistical feature constraint to output photovoltaic power scene sequences conforming to actual power fluctuation rules.

2. The photovoltaic scene intelligent generation method according to claim 1, characterized in that, The multi-scale frequency domain decomposition of historical photovoltaic data by using a wavelet decomposition module to extract signal components at different frequency scales comprises: Performing multi-level discrete wavelet transform on historical photovoltaic data to obtain frequency domain components at multiple scales; wherein the wavelet basis can be any one of the Daubechies, Symlet or Coiflet families.

3. The photovoltaic scene intelligent generation method according to claim 1, characterized in that, The optimization and verification of the generation result based on a frequency spectrum structure loss and a statistical feature constraint comprises: Constraining the frequency energy distribution of the generated signal by using a frequency spectrum structure loss function to make it consistent with the real historical data in the main frequency band; Using a multi-scale discriminator to evaluate the authenticity of the generation result in the frequency spectrum space to identify whether it has the trend and disturbance of the actual power signal; Combining the statistics of the time domain mean, volatility and autocorrelation coefficient to conduct overall testing on the generated data to ensure that it meets the physical constraints and engineering characteristics.

4. The photovoltaic scene intelligent generation method of claim 1, wherein, Further comprising: Pretreating the historical photovoltaic data, which comprises time alignment of data from different data sources, linear interpolation of missing data and filtering correction of abnormal values; Normalizing the complete sequence of the corrected historical photovoltaic data.

5. The photovoltaic scene intelligent generation method according to claim 1, characterized in that, The calculation formula of the signal components at different frequency scales is: wherein denotes that the signal component resulting from the decomposition is an approximation component, denotes a plurality of detail components, denotes a first layer low frequency approximation component coefficient, denotes a first layer high frequency perturbation component coefficient, denotes a low frequency approximation basis, denotes a high frequency perturbation basis, denotes a normalized photovoltaic power sequence.

6. The photovoltaic scene intelligent generation method according to claim 1, characterized in that, The output of the information fusion is: In the formula, represents a fusion vector, represents the intermediate latent variable with frequency perturbation characteristics at the dimension, represents the structure generation weight of the frequency dimension.

7. The photovoltaic scene intelligent generation method of claim 1, wherein, The frequency spectrum structure loss is: wherein, a loss function representing a loss of spectral structure, and represents generating an amplitude spectrum at frequencies of the real sequence; represents a frequency weighting function.

8. An apparatus for photovoltaic scene intelligent generation, characterized in that, Comprise: A processor; A memory storing computer program instructions, which realize the photovoltaic scene intelligent generation method according to any one of claims 1-7 when executed by the processor.

Images