Photovoltaic scene intelligent generation method and system

Abstract

The invention relates to the technical field of data processing, in particular to a photovoltaic scene intelligent generation method and system, and the method comprises the steps: obtaining historical photovoltaic data, and carrying out the multi-scale frequency domain decomposition of the historical photovoltaic data through a wavelet decomposition module, so as to extract signal components under different frequency scales; a frequency domain re-parameterization mechanism is utilized to embed each scale component into the structure of the deep learning model, and a frequency sensing multi-head attention mechanism is constructed; performing context modeling and structure fusion under a frequency condition in a frequency spectrum space so as to realize structural generation of a photovoltaic scene; and on the basis of the spectrum structure loss and the statistical feature constraint, optimizing and checking the generated result to output a photovoltaic power scene sequence conforming to an actual power fluctuation rule. According to the scheme of the invention, the applicability of the photovoltaic output scene data engineering is effectively improved, and the real situation in the industrial environment is reflected.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology. More specifically, the present invention relates to a method and system for intelligently generating photovoltaic scenes. Background Art

[0002] With the large-scale integration of renewable energy sources such as photovoltaics, the power grid has placed higher demands on the simulation and generation of renewable energy output scenarios. Photovoltaic output exhibits significant periodicity, suddenness, and instability. Its power time series data often contains fluctuation patterns at multiple scales, such as trend changes driven by sunlight intensity, medium-frequency fluctuations caused by cloud shadows, and high-frequency disturbances caused by the dynamic response of equipment. These time series characteristics are distributed across different frequency ranges, exhibit non-stationarity, and exhibit multi-scale overlap, making them difficult to effectively capture using traditional single-scale modeling methods.

[0003] Currently, mainstream time series generation methods mostly rely on time domain modeling and are unable to explicitly identify and control the fluctuation behavior of different frequency bands. This leads to deviations in the generated results in terms of global trends, local disturbances, and detail consistency, making it difficult to meet the comprehensive requirements of realism, controllability, and rationality.

[0004] In recent years, deep learning models, thanks to their self-attention architecture, have demonstrated strong contextual modeling capabilities in processing long sequences of data and have been gradually introduced into new energy scenario generation tasks. However, standard self-attention mechanisms, operating on raw time-domain signals, lack explicit perception of frequency structure, making it difficult to decouple and reconstruct multi-scale information. Especially when faced with time-series signals with significant spectral hierarchy, such as photovoltaic output, the time-domain attention architecture cannot effectively extract the statistical structure and evolution patterns of different frequency bands, easily resulting in the suppression of high-frequency details or the drift of low-frequency trends, leading to irrational photovoltaic scenarios.

[0005] Therefore, what needs to be urgently addressed is the problem that photovoltaic output scenario data (such as power generation) has poor engineering applicability and cannot reflect the real situation. Summary of the Invention

[0006] In order to solve the technical problem that the above-mentioned photovoltaic output scenario data (such as power generation) has poor engineering applicability and cannot reflect the actual situation, the present invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides a method for intelligent generation of photovoltaic scenes, the method comprising: acquiring historical photovoltaic data, performing multi-scale frequency domain decomposition on the historical photovoltaic data 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 using a frequency domain reparameterization mechanism to construct a frequency-aware multi-head attention mechanism; performing context modeling and structural fusion under frequency conditions in the spectrum space to achieve structural generation of photovoltaic scenes; and optimizing and verifying the generation results based on spectrum structure loss and statistical feature constraints to output a photovoltaic power scene sequence that conforms to the actual power fluctuation law.

[0008] Preferably, using a wavelet decomposition module to perform multi-scale frequency domain decomposition on historical photovoltaic data to extract signal components at different frequency scales includes: performing a 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 families; the frequency domain components extracted at each frequency scale are respectively used as independent input channels to construct a multi-frequency scale embedding tensor.

[0009] Preferably, the frequency-aware multi-head attention mechanism includes: constructing multi-frequency band attention heads, each attention head processing the frequency component corresponding to the set wavelet band; dynamically weighting and converging the output of each frequency band attention head based on the spectrum gating mechanism to achieve frequency-adaptive information fusion.

[0010] Preferably, context modeling and structural fusion under frequency conditions are performed in the spectral space, including: 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 the input of the generation module; introducing a learnable scale selection function to dynamically adjust the intensity of the effect of different frequency components in the generation.

[0011] Preferably, the optimization and verification of the generated results based on the spectrum structure loss and statistical feature constraints include: constraining the frequency energy distribution of the generated signal through the 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 generated results in the spectrum space to distinguish 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 perform an overall test on the generated data to ensure that it meets the physical constraints and engineering characteristics.

[0012] Preferably, the method further includes: preprocessing the historical photovoltaic data, wherein the preprocessing includes time alignment of data from different data sources, linear interpolation of missing data, and filtering correction of outliers; and normalizing the complete sequence of the corrected historical photovoltaic data.

[0013] Preferably, the calculation formula of the signal components at different frequency scales is:

[0014]

[0015]

[0016] Where A J (t) indicates that the signal component obtained by decomposition is an approximate component, D j (t) represents multiple detail components, c J,k =∫f(t)φ J,k (t)dt represents the coefficient of the low-frequency approximate component of the Jth layer, d j,k =∫f(t)ψ j,k (t)dt represents the coefficient of the high-frequency disturbance component of the jth layer, φ J,k (t) represents the low-frequency approximation basis, ψ j,k (t) represents the high-frequency perturbation basis.

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

[0018]

[0019] Where, represents the fusion vector, represents the perturbation latent variable sampling result at the jth scale, α j represents the structure generation weight of the j-th frequency scale.

[0020] Preferably, the spectrum structure loss is:

[0021]

[0022] in, The loss function representing the loss of spectral structure, and |F(f)| represent the amplitude spectra of the generated and true sequences at frequency f; w(f) represents the frequency weight function.

[0023] In a second aspect, the present invention further provides a photovoltaic scene intelligent generation device, comprising: a processor; a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a photovoltaic scene intelligent generation method according to one or more embodiments described above is implemented.

[0024] The beneficial effects of the present invention are as follows: the present invention constructs a frequency domain multi-scale structure foundation by wavelet decomposition, and combines the self-attention modeling mechanism in the spectrum space to achieve unified modeling and precise generation of photovoltaic power time series signals in trend components and disturbance components. Specifically, the present invention first deconstructs historical photovoltaic power data into multiple scale frequency components through discrete wavelet transform to capture its change structure at different levels; then uses frequency domain reparameterization technology to embed the frequency features of each scale as input into the multi-head self-attention Transformer, and performs context modeling and structural coupling in the spectrum space; then realizes the collaborative synthesis of multi-scale signals through frequency condition modeling and context construction methods; finally, introduces spectrum structure loss and statistical feature constraints, and constrains the generated output from aspects such as frequency distribution, consistency and statistical structure, ensuring that the generated scene has high authenticity and physical rationality in both trend and fluctuation levels, effectively improving the engineering applicability of photovoltaic output scenario data (such as power generation) and reflecting the real situation in industrial environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart illustrating a method for intelligently generating a photovoltaic scene according to an embodiment of the present invention;

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

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

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

[0029] Figure 1 is a flow chart illustrating a method for intelligently generating a photovoltaic scene according to an embodiment of the present invention.

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

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

[0032]

[0033]

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

[0035] At step S102, the frequency domain reparameterization mechanism is used to embed each scale component into the structure of the deep learning model, constructing a frequency-aware multi-head attention mechanism. In some embodiments, multiple frequency band attention heads are constructed, each of which processes the frequency components corresponding to a set wavelet frequency band. The outputs of each frequency band attention head are dynamically weighted and aggregated based on a spectral gating mechanism to achieve frequency-adaptive information fusion.

[0036] The output of information fusion is:

[0037]

[0038] Where, represents the fusion vector, represents the perturbation latent variable sampling result at the jth scale, α j represents the structure generation weight of the j-th frequency scale.

[0039] In step S103, frequency-dependent context modeling and structural fusion are performed within the spectral space to achieve structured generation of the photovoltaic scene. In some embodiments, the signal components at each frequency scale can be converted into Gaussian distribution variables, and their mean and variance can be determined. Samples are then taken from the variational distribution at each frequency scale and used as input to the generation module. A learnable scale selection function is introduced to dynamically adjust the influence of different frequency components in the generation.

[0040] At step S104, the generated results are optimized and verified based on spectral structure loss and statistical feature constraints to output a photovoltaic power scenario sequence that conforms to actual power fluctuation patterns. In some embodiments, the frequency energy distribution of the generated signal can be constrained by a spectral structure loss function to ensure consistency with real historical data within the main frequency band. A multi-scale discriminator is used to evaluate the authenticity of the generated results in the spectral space to determine whether they exhibit the trend and perturbations of the actual power signal. The generated data is holistically verified using statistics such as the time domain mean, volatility, and autocorrelation coefficient to ensure that it meets physical constraints and engineering characteristics.

[0041] The spectral structure loss is:

[0042]

[0043] in, The loss function representing the loss of spectral structure, and |F(f)| represent the amplitude spectra of the generated and true sequences at frequency f; w(f) represents the frequency weight function.

[0044] Furthermore, before performing the wavelet decomposition on the historical photovoltaic data, the historical photovoltaic data may be preprocessed. The preprocessing includes time alignment of data from different data sources, linear interpolation of missing data, and filtering correction of outliers. The complete sequence of the corrected historical photovoltaic data is then normalized.

[0045] Next, the solution of the present invention will be further described in detail with reference to specific embodiments.

[0046] Step 1: Preprocess the historical photovoltaic time-series output data, which specifically includes the following steps:

[0047] First, historical output data is collected based on the operation records of the photovoltaic power station. The time granularity is hourly or minutely, and the time step Δt is uniformly set (such as 5 minutes, 15 minutes or 30 minutes). The timestamps of different data sources are aligned. After data alignment, the original data sequence is converted into a standard time series format, which is recorded as P t , where P tRepresents the photovoltaic output power at time step t, in kilowatts (kW).

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

[0049] For missing data, linear interpolation is used to fill in the gaps. Let time points t1 and t2 be the two valid sampling moments before and after the missing point, and record is the photovoltaic power value at the corresponding moment, then the missing point P t The interpolation calculation formula is:

[0050]

[0051] Among them, P t Indicates the power value after interpolation, t is the time of the missing point, t1 <t<t2。

[0052] For the detected abnormal values, the statistical filtering method is used to correct them. Assuming the mean of historical data is μ and the standard deviation is σ, μ±3σ is used as the abnormality judgment threshold. If the power P at a certain moment is t If the value exceeds this range, it is considered an outlier. The replacement value of the outlier is reconstructed using the sliding mean of adjacent windows. The sliding mean calculation formula is:

[0053]

[0054] in, Represents the smoothed power value, and the window radius k is usually 2 to 4.

[0055] Finally, the repaired complete sequence is normalized to eliminate the interference caused by the absolute power difference on the frequency domain decomposition. Let the maximum value of the entire sequence be P max , the minimum value is p min , then the normalized power at each time step is The calculation formula is:

[0056]

[0057] Normalized power series Maintained in the interval [0,1], it serves as the input for subsequent multi-scale frequency domain analysis.

[0058] Step 2: Use the wavelet decomposition module to perform multi-scale frequency domain decomposition on the pre-processed time series data to extract the trend and disturbance signal components at different frequency scales. The specific steps include the following:

[0059] First, the normalized photovoltaic power time series data obtained in step 1 is Discrete Wavelet Transform (DWT) is applied to perform multi-scale frequency domain decomposition. DWT constructs a set of time-frequency localized orthogonal basis functions through the scaling function φ(t) and the wavelet function ψ(t):

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

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

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

[0063]

[0064] Among them, c J,k =∫f(t)φ J,k (t)dt represents the coefficient of the low-frequency approximate component of the Jth layer; d j,k =∫f(t)ψ j,k (t)dt represents the coefficient of the high-frequency disturbance component of the jth layer; φ J,k (t) constitutes a low-frequency approximation basis, ψ j,k (t) constitutes a high-frequency perturbation basis. Assume that the input sequence is the normalized photovoltaic power sequence The decomposition result is the approximate component A J (t) and several detail components D j (t), where:

[0065]

[0066]

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

[0068] Then, all components {A J (t),D1(t),…,D J(t)} as multi-channel feature input to construct a unified tensor Each column represents a component at a scale. This tensor forms the basis for the multi-frequency input of the Transformer model, ensuring that the subsequent network structure can separately model and jointly construct trends and disturbances in 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: The 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 scaling 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 set to 4–6 to ensure time-frequency resolution. The Symlet wavelet (symN) family is a symmetry-enhanced Daubechies wavelet with the same vanishing moment but improved filter symmetry, so that the wavelet function satisfies the minimum phase symmetry condition, reducing edge effects while maintaining analytical ability and satisfying the same recursive expression:

[0072]

[0073] in It is an improved low-pass filter whose design goal is to maximize symmetry and frequency band locality. The Coiflet wavelet (coifN) family has the characteristics of N vanishing moments of both the scaling function and the wavelet function, and has good time-frequency dual localization capability. 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 where long-term trends and transient changes coexist, and it is particularly effective in modeling photovoltaic power in scenarios with both stationary components and disturbances. All three wavelet bases possess compact support, orthogonality, and complete reconstruction properties. Wavelet basis functions are selected based on the actual application task: Daubechies wavelets are used in scenarios requiring enhanced high-frequency disturbance detection, Symlet wavelets are preferred for trend smoothing modeling, and Coiflet wavelets are preferred for scenarios involving the coordinated expression of high and low frequencies. This achieves controllable signal structure and optimized expression within the decomposed hierarchical structure.

[0077] Step 3: Use the frequency domain reparameterization mechanism to embed the frequency components of each scale into the Transformer structure, build a frequency-aware multi-head self-attention mechanism, and realize contextual representation under scale conditions. The specific steps include the following:

[0078] First, in the encoding structure of the model, the input tensor Construct a multi-head attention structure corresponding to the frequency segment. Assume that the model contains H attention heads, each attention head h j Focusing on the jth frequency scale channel in the tensor (corresponding to the jth layer of wavelet components), each attention head receives input:

[0079] X (j) ∈R T×d ,j=1,2,…,J+1

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

[0081]

[0082] Among them, 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×dare the transformation matrices for query, key, and value, respectively. Each attention head independently captures the contextual dependencies within the frequency band through the self-attention mechanism, constructing a refined response to the trend segment or disturbance segment.

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

[0084]

[0085] Among them, W g ∈R d 、b g ∈R is a trainable parameter, mean(Z j ) indicates Z j The operation of finding the mean in the time dimension, σ(·) is the Sigmoid activation function. The final fusion output is expressed as:

[0086]

[0087] The above structure realizes the dynamic weighting of the attention head outputs in different frequency bands and has the frequency-adaptive context aggregation capability.

[0088] Finally, a cross-frequency information exchange mechanism is added to the Transformer encoding structure, specifically through context communication between cross-heads. Assuming that there is a potential dependency between different frequency bands, for example, the local perturbation at the jth scale depends on the trend background information at the ith scale, a cross-attention mechanism can be introduced:

[0089]

[0090] The result is Z j The enhanced term enables coupled information propagation between frequency channels. This mechanism supports deep structural fusion between components of different scales within the model, helping to capture both global trends and local mutations when generating photovoltaic power series, significantly improving the learning ability and expression completeness of multi-scale dependencies.

[0091] Step 4: Perform frequency condition modeling and structural generation in the spectrum space, and dynamically adjust the generation contribution of different frequency bands through the scale response control function. The specific steps include the following:

[0092] First, the frequency scale representation obtained by the multi-head attention mechanism in step 3 is As a type of conditional latent variable in the frequency domain, they are input into the frequency domain variational inference machine for statistical modeling. Each scale corresponds to the representation Z j ∈R T×d is considered to be subject to a conditional Gaussian distribution For samples, we construct a variational prior of the following form:

[0093]

[0094] Where: z∈R d represents the potential eigenvector at this frequency scale; μ j ∈R d is the mean vector under the j-th frequency band; is the corresponding variance vector; diag(·) represents the construction of the diagonal covariance matrix. The variational modeling process is implemented through a neural network, where μ j =f μ (Z j ), They are output by two fully connected transformation modules respectively, and statistical feature structures of different scales are 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 perturbation characteristics As the input basis of the generation module. The sampling process is implemented using the reparameterization technique to ensure that the model is differentiable during back propagation:

[0096]

[0097] Among them, ∈∈R d Represents a standard normal distribution Noise samples (I is the unit matrix), ⊙ represents the element-wise multiplication operation, is the sampling result of the perturbation latent variable at 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. This control function is implemented by a learnable module g(·), which calculates the fusion weight α of each scale component under the current context. j ∈[0,1], which is expressed as:

[0099]

[0100] Among them, W α ∈R d 、b α ∈R is a trainable parameter, α j represents the structural generation weight of the jth frequency scale. The final structural generation feature vector is:

[0101]

[0102] The fusion vector It comprehensively expresses the dynamic generation effect of multi-scale frequency components under different contextual conditions and realizes the adaptive coupling of trend structure (low frequency) and disturbance component (high frequency).

[0103] Step 5: Based on the spectral structure loss and statistical feature constraints, the generated results are optimized and verified. The specific steps include the following:

[0104] First, the spectrum 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 neighboring frequency band. Suppose the generated sequence is The real historical sequence is P t , perform discrete Fourier transform (DFT) on both of them to get the spectrum representation:

[0105]

[0106] in, represents the Fourier transform operation, and f is the frequency variable. The spectral structure loss function is defined as:

[0107]

[0108] in: |F(f)| represents the amplitude spectrum of the generated and true 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 diurnal cycle and the cloud shadow perturbation bandwidth); the loss function measures the amplitude difference of the frequency domain structure and constrains the model to maintain a reasonable frequency distribution during the generation process.

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

[0110]

[0111] Among them D j (·) is the output of the j-th frequency band discriminator; λ is the weight of the gradient penalty term; It is the intermediate spectrum sample obtained by interpolation 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 statistics are introduced to check the overall consistency of the output results, including indicators such as volatility and autocorrelation structure. Used to evaluate instantaneous rate of change; first-order autocorrelation coefficient Evaluate the stationarity and trend coherence of the series. If the generated data falls within the statistical tolerance interval set based on real data in terms of the above indicators, it is considered to have good engineering usability and credibility; otherwise, the model will optimize and adjust the generated structure through a loss function combined with a feedback mechanism.

[0113] This paper utilizes a wavelet decomposition module for physical scale extraction, frequency-domain reparameterization and contextual feature learning in the attention modeling module, frequency-controlled structure generation, and a spectral-statistical dual-constraint optimization mechanism to intelligently extract multi-scale structural features from historical data and use them to guide generation. The resulting photovoltaic time-series scenario data not only retains the long-term trends and short-term disturbance characteristics of the original data, but also exhibits excellent frequency consistency, physical rationality, and statistical stability. This effectively supports predictive modeling and scheduling decision-making in new energy scenarios, significantly improving the robustness and efficiency of system operations.

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

[0115] The present invention also provides a photovoltaic scene intelligent generation device. Figure 2 As shown, the device includes a processor and a memory, wherein the memory stores computer program instructions. When the computer program instructions are executed by the processor, a photovoltaic scene intelligent generation method according to one or more embodiments described above is implemented.

[0116] The device also includes other components well known to those skilled in the art, such as a communication bus and a communication interface. The configuration and functions of these components are known in the art and will not be described in detail here.

[0117] While several embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Numerous modifications, variations, and alternatives will occur to those skilled in the art without departing from the concept and spirit of the present invention. It should be understood that various alternatives to the embodiments of the present invention described herein may be employed in practicing the present invention.

Claims

1. A photovoltaic scene intelligent generation method, characterized in that: The method comprises: Obtain historical photovoltaic data and use the wavelet decomposition module to perform multi-scale frequency domain decomposition on the historical photovoltaic data to extract signal components at different frequency scales; The frequency domain reparameterization mechanism is used to embed the scale components into the structure of the deep learning model and build a frequency-aware multi-head attention mechanism. Conduct frequency-conditioned context modeling and structural fusion in the spectrum space to achieve structured generation of photovoltaic scenarios; Based on the spectrum structure loss and statistical feature constraints, the generated results are optimized and verified to output a photovoltaic power scenario sequence that conforms to the actual power fluctuation law.

2. The photovoltaic scene intelligent generation method according to claim 1, characterized in that: The wavelet decomposition module is used to perform multi-scale frequency domain decomposition on historical photovoltaic data to extract signal components at different frequency scales, including: Perform multi-level discrete wavelet transform on historical photovoltaic data to obtain frequency domain components at multiple scales; the wavelet basis can be any one of the Daubechies, Symlet or Coiflet families; The frequency domain components extracted at each frequency scale are used as independent input channels to construct a multi-frequency scale embedding tensor.

3. The photovoltaic scene intelligent generation method according to claim 1, characterized in that: The frequency-aware multi-head attention mechanism includes: Construct multi-frequency band attention heads, each of which processes the frequency components corresponding to the set wavelet band; Based on the spectrum gating mechanism, the output of the attention head in each frequency segment is dynamically weighted and aggregated to achieve frequency-adaptive information fusion.

4. The photovoltaic scene intelligent generation method according to claim 1, characterized in that: Frequency-conditioned context modeling and structural fusion in the spectral space, including: Convert the signal components at each frequency scale into Gaussian distribution variables and determine their mean and variance; Sampling from the variational distribution at each frequency scale as input to the generation module; A learnable scale selection function is introduced to dynamically adjust the strength of different frequency components in generation.

5. The photovoltaic scene intelligent generation method according to claim 1, characterized in that: The optimization and verification of the generated results based on the spectrum structure loss and statistical feature constraints include: The frequency energy distribution of the generated signal is constrained by the spectrum structure loss function to make it consistent with the real historical data within the main frequency band; A multi-scale discriminator is used to evaluate the authenticity of the generated results in the spectral space to determine whether they have the trend and perturbation of the actual power signal; The generated data are comprehensively tested by combining the statistics of time domain mean, volatility and autocorrelation coefficient to ensure that they meet the physical constraints and engineering characteristics.

6. The photovoltaic scene intelligent generation method according to claim 1, characterized in that: Also includes: Preprocessing historical photovoltaic data, including time alignment of data from different data sources, linear interpolation of missing data, and filtering and correction of outliers; The complete series of corrected historical PV data is normalized.

7. The photovoltaic scene intelligent generation method according to claim 1, characterized in that: The calculation formula for signal components at different frequency scales is: Where A J (t) indicates that the signal component obtained by decomposition is an approximate component, D j (t) represents multiple detail components, c J,k =∫f(t)φ J,k (t)dt represents the coefficient of the low-frequency approximate component of the Jth layer, d j,k =∫f(t)ψ j,k (t)dt represents the coefficient of the high-frequency disturbance component of the jth layer, φ J,k (t) represents the low-frequency approximation basis, ψ j,k (t) represents the high-frequency perturbation basis.

8. The photovoltaic scene intelligent generation method according to claim 3, characterized in that: The output of information fusion is: Where, represents the fusion vector, represents the perturbation latent variable sampling result at the jth scale, α j represents the structure generation weight of the j-th frequency scale.

9. The photovoltaic scene intelligent generation method according to claim 1, characterized in that: The spectral structure loss is: in, The loss function representing the loss of spectral structure, and |F(f)| represent the amplitude spectra of the generated and true sequences at frequency f; w(f) represents the frequency weight function.

10. A photovoltaic scene intelligent generation device, characterized in that: include: processor; A memory storing computer program instructions, wherein when the computer program instructions are executed by the processor, a photovoltaic scene intelligent generation method according to any one of claims 1 to 9 is implemented.

Images