Method and system for predicting evolution process of photovoltaic power generation power probability density function

Abstract

The invention discloses a method and a system for predicting an evolution process of a photovoltaic power generation power probability density function, which are used for solving the technical problem that the reliability and the accuracy of prediction are influenced due to the fact that the fluctuation and the long-term trend of power along with time cannot be accurately captured in the prior art, and belong to the technical field of deep learning. The invention discloses a method for predicting the evolution process of a photovoltaic power generation power probability density function. The method comprises the following steps: 1, obtaining a historical probability density function; step 2, obtaining a coefficient vector of the primary function; 3, enhancing the characterization force of the meteorological characteristic information; step 4, reconstructing meteorological features; 5, probability density function evolution is predicted; s51, splicing the coefficient vector and the reconstructed meteorological features to form more comprehensive and information-rich spliced features; the splicing features and the photovoltaic historical power generation power are input into a multi-head cross attention mechanism for weighted solution, the splicing features are reconstructed, and a predicted coefficient vector is obtained through a full connection layer; and S52, carrying out reverse solution on the predicted coefficient vector through cubic B-spline interpolation to obtain a predicted photovoltaic power generation power probability density function.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and in particular relates to a prediction method and system for the evolution process of a photovoltaic power generation power probability density function. Background Art

[0002] Photovoltaic power generation forecasting provides an important reference for the dispatching and operation of power grids, helps to promote the effective integration of renewable energy and power grids, and promotes the construction of green and low-carbon power systems. Since photovoltaic power generation is affected by many uncertain factors such as weather conditions and seasonal changes, its output power has strong randomness and uncertainty. Therefore, accurate power forecasting of photovoltaic power generation is crucial for the stability and energy management of the power grid. Traditional point forecasting only provides a single power value forecast, which cannot quantify the uncertainty of the forecast value, and it is difficult to effectively deal with the forecast errors caused by power fluctuations and abnormal weather, which will lead to difficulties in power grid dispatching and increased profit and loss risks. In contrast, probabilistic forecasting quantifies potential forecast errors by providing a probability distribution of power output, and can fully describe the possibility of different power levels. The accuracy of the forecast and the accurate description of the distribution shape are the key to achieving reliable forecasting.

[0003] Among the probabilistic forecasting methods, the current probabilistic forecasting methods have the following problems: 1. Probabilistic forecasting such as quantile forecasting and interval forecasting lacks the characterization of the continuity of the overall probability distribution, and it is difficult to accurately describe the extreme situations that may occur when the uncertainty is large. 2. The prior probability distribution assumption of power is made, and the assumed distribution is usually inconsistent with the distribution of actual data. This prior assumption limits the flexibility of the model and will lead to forecasting deviations under extreme weather conditions. 3. Many probabilistic forecasting methods ignore the dependence of time series. When dealing with scenarios where photovoltaic power generation is greatly affected by weather and time periods, ignoring the time series characteristics and dynamic evolution of power changes will cause the model to be unable to accurately capture the fluctuations and long-term trends of power over time, thereby affecting the reliability and accuracy of the forecast. Summary of the invention

[0004] The purpose of the present invention is to provide a prediction method and system for the evolution process of photovoltaic power generation power probability density function, which is used to solve the technical problem that the existing technology cannot accurately capture the fluctuation and long-term trend of power over time, thereby affecting the reliability and accuracy of the prediction.

[0005] The present invention is achieved by adopting the following technical solutions:

[0006] A method for predicting the evolution process of photovoltaic power generation power probability density function comprises the following steps:

[0007] Step 1, obtain the historical probability density function;

[0008] S11, taking the historical photovoltaic power generation as input, and using kernel density estimation to obtain the probability density function (Probability Density Function, PDF) of photovoltaic power generation.

[0009] Step 2, obtaining the coefficient vector of the basis function;

[0010] S21, after defining the control points, the PDF of photovoltaic power generation is fitted using the cubic B-spline interpolation method;

[0011] S22, calculate the coefficient vector of the basis function corresponding to the control point.

[0012] Step 3: Enhance the representation of meteorological characteristic information;

[0013] S31, denoising the meteorological features;

[0014] S32, encoding the time information of the meteorological characteristics;

[0015] S33, performing feature splicing on the denoised meteorological features and the encoded time information to form a first splicing feature;

[0016] S34, input the first concatenated features into iTransform to obtain new meteorological features, capture the temporal relationship between different meteorological features, and enhance the representation of meteorological information.

[0017] Step 4: Reconstruct meteorological characteristics;

[0018] S41, using a multi-head cross-attention mechanism to perform weighted summation on the meteorological features output by iTransformer, so that the meteorological features have the same dimension as the coefficient vector, thereby realizing the reconstruction of the meteorological features;

[0019] S42, the input of the multi-head cross attention mechanism is the coefficient vector obtained in step S22 and the meteorological features obtained in step S34;

[0020] Step 5: predict the evolution of probability density function;

[0021] S51, concatenate the coefficient vector and the reconstructed meteorological features to form a more comprehensive and information-rich concatenated feature; weight the concatenated feature and the photovoltaic historical power generation input into the multi-head cross attention mechanism, reconstruct the concatenated feature, and obtain the predicted coefficient vector through the fully connected layer. The coefficient vector and photovoltaic historical power generation input into the multi-head cross attention mechanism are obtained after self-attention in advance;

[0022] S52, performing inverse solution of the predicted coefficient vector by cubic B-spline interpolation to obtain a predicted photovoltaic power generation power probability density function.

[0023] Further preferably, the meteorological characteristics include direct irradiance, temperature, humidity, wind speed and atmospheric pressure.

[0024] Further preferably, the kernel density estimation formula is: in, represents the probability density estimate at point x, where x is the photovoltaic power generation value to be estimated; n is the total number of data points; κ(·) represents the kernel function; h is the bandwidth parameter used to control the width of the kernel; x i is the actual power value of the ith photovoltaic cell.

[0025] Further preferably, the kernel function adopts a Gaussian kernel function, and the formula is: in is a standardized variable, indicating the actual x i The relative distance to x.

[0026] Further preferably, discrete wavelet transform (DWT) is used to perform noise reduction on the meteorological characteristics.

[0027] Further preferably, the control points are selected from the PDF curve in equal proportion.

[0028] Further preferably, S21 uses cubic B-spline to fit the obtained PDF, and the fitting formula is as follows: Where S(x) is the fitted PDF, P i is the i-th control point, B i,3 (x) is the cubic B-spline basis function corresponding to the i-th control point, which is constructed through a recursive relationship, and the specific form is:

[0029]

[0030] Among them, x i is the i-th node vector, and at a given x value, the coefficient vector M is solved by the interpolation condition.

[0031] The relationship between the coefficient vector of the S22 cubic B-spline and the control points is expressed by the following linear equation system: M×{P0,P1,…,P n}={S0,S1,…,S n}, the coefficient vector M is obtained by the basis function satisfying the conditions of value continuity, first-order derivative continuity and second-order derivative continuity, which represents the relationship between the control point and the interpolation curve, and the coefficient vector is extracted as the coefficient vector.

[0032] A prediction system for the evolution process of photovoltaic power generation power probability density function is used to implement any of the above-mentioned prediction methods for the evolution process of photovoltaic power generation power probability density function, including a weight extraction module, a noise reduction module, a multivariate attention mechanism and a prediction module.

[0033] The weight extraction module realizes the extraction of basis function coefficient vectors. Specifically, in order to obtain the coefficient vectors corresponding to the control points on the PDF curve, the PDF of the historical photovoltaic power generation power is first obtained by kernel density estimation using the Gaussian kernel function, and then the PDF is fitted using cubic B-spline interpolation. The coefficient vectors corresponding to the control points calculated during the fitting process are extracted as the coefficient vectors.

[0034] The denoising module realizes meteorological feature denoising and time information encoding. The collected meteorological information is denoised using DWT to minimize the impact of noise. At the same time, the time information is encoded and the meteorological information is concatenated with the encoded time information. Finally, the meteorological features containing time series information are obtained in the denoising module.

[0035] The multivariate attention mechanism realizes the reconstruction of meteorological features and the reconstruction after splicing the coefficient vector and meteorological features.

[0036] Specifically, iTransformer is used to embed each feature independently, so that each feature is considered equally, while preventing confusion and information loss between channels. iTransformer can also learn fine-grained relationships and multi-dimensional dependency structures between different features, as well as deep associations between time information and meteorological features.

[0037] By considering the correlation between the coefficient vector and meteorological information, a multi-head cross-attention mechanism is adopted to reconstruct the features of meteorological information according to the correlation, so that the meteorological features are aligned with the coefficient vector in dimension and time, ensuring that they can be effectively fused in the same space.

[0038] The reconstructed meteorological features are spliced ​​with the coefficient vector after the self-attention mechanism, and then input into the multi-head cross-attention mechanism again. The spliced ​​features are reconstructed again using the correlation between the coefficient vector and the historical power.

[0039] In the multivariate attention mechanism, these features are weighted and combined according to different attention heads to achieve more accurate feature representation.

[0040] The prediction module obtains the predicted coefficient vector and the predicted photovoltaic power PDF.

[0041] The final reconstructed features are input into the linear fully connected layer, and the input features are linearly transformed to achieve the prediction of the coefficient vector. The predicted coefficient vector is inversely solved using cubic B-spline interpolation to obtain the predicted photovoltaic power PDF.

[0042] This system can capture and predict the changing characteristics of photovoltaic power PDF over time, and can dynamically adjust the probability distribution of power generation to obtain the global probability distribution of photovoltaic power, providing a more comprehensive risk assessment tool for the power grid.

[0043] The present invention has the following beneficial effects:

[0044] (1) Significant results have been achieved in the prediction of photovoltaic power PDF, which can more comprehensively reflect the randomness and uncertainty caused by changes in meteorological conditions in the photovoltaic power generation process. The predicted photovoltaic power PDF can define the possible range and probability distribution of power generation, so that the power grid dispatching center can better estimate the fluctuation range of photovoltaic power generation, providing reliable data support for dispatching optimization and energy management.

[0045] (2) This innovative method can improve the accuracy and reliability of photovoltaic power generation forecasts, allowing the power grid to not only access renewable energy more efficiently, but also take timely response measures when photovoltaic power generation suddenly increases or decreases, thereby ensuring the stable operation of the power grid.

[0046] (3) iTransformer is used to model temporal relationships, and a multi-head cross-attention mechanism is used to effectively capture and utilize the complex patterns and relationships in time series. PDF, with its comprehensive probabilistic information description, provides a more flexible and detailed characterization of power uncertainty. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 A structural diagram showing the photovoltaic power generation power probability density function evolution model of the present invention.

[0050] Figure 2 The flowchart of the photovoltaic power generation power probability density function of the present invention is shown.

[0051] Figure 3 Represents the photovoltaic power generation and meteorological characteristic data of the present invention.

[0052] Figure 4 A diagram showing the prediction results of the photovoltaic power generation power probability density function evolution model of the present invention.

[0053] Figure 5 The actual evolution diagram of the photovoltaic power generation power probability density function of the present invention is shown.

[0054] Figure 6 The diagram represents the predicted evolution diagram of the photovoltaic power generation power probability density function of the present invention. DETAILED DESCRIPTION

[0055] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the scheme of the present invention will be further described below. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0056] In the description, it should be noted that the terms "first" and "second" are only used for descriptive purposes and should not be understood as indicating or implying relative importance. It should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be internal communication between two elements. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0057] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present invention, rather than all of the embodiments.

[0058] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0059] Embodiment 1, a method for predicting the evolution process of photovoltaic power generation power probability density function, such as Figure 1-6 As shown, the following steps are included:

[0060] Step 1: Get the historical PDF;

[0061] S11, taking the historical photovoltaic power generation as input, and using kernel density estimation to obtain the PDF of photovoltaic power generation;

[0062] The kernel density estimation formula is: in, represents the probability density estimate at point x, where x is the photovoltaic power generation value to be estimated; n is the total number of data points; κ(·) represents the kernel function; h is the bandwidth parameter used to control the width of the kernel; x i is the actual power value of the ith photovoltaic cell.

[0063] The kernel function uses the Gaussian kernel function, and the formula is in is a standardized variable, indicating the actual x i The relative distance to x.

[0064] The dataset used in this example comes from a photovoltaic power station in Hebei Province, China, and the dataset was collected from July 2018 to June 2019. The dataset contains photovoltaic power and seven meteorological characteristics, including global horizontal irradiance (GHI), direct irradiance (DNI), temperature, relative humidity, wind speed, wind direction, and atmospheric pressure. The temporal resolution of the data is 15 minutes.

[0065] Step 2, obtaining the coefficient vector of the basis function;

[0066] S21, after defining the control points, use the cubic B-spline interpolation method to fit the PDF of photovoltaic power generation; in this embodiment, the control points are selected from the PDF curve in equal proportion.

[0067] After the PDF of photovoltaic power generation is obtained by kernel density estimation, cubic B-spline is used to fit the obtained PDF. The fitting formula is as follows: Where S(x) is the fitted PDF, P i is the i-th control point, B i,3 (x) is the cubic B-spline basis function corresponding to the i-th control point, which is constructed through a recursive relationship, and the specific form is:

[0068]

[0069]

[0070] Among them, x i is the i-th node vector, and under a given x value, the coefficient vector M is solved by the interpolation condition;

[0071] S22, calculating the coefficient vector of the basis function corresponding to the control point;

[0072] The relationship between the coefficient vector of the cubic B-spline and the control points is expressed by the following linear equation system: M×{P0,P1,…,P n}={S0,S1,…,S n}, the coefficient vector M is obtained by satisfying the conditions of value continuity, first-order derivative continuity and second-order derivative continuity of the basis function, which represents the relationship between the control point and the interpolation curve.

[0073] Step 3: Enhance the representation of meteorological characteristic information;

[0074] S31, DWT is used to reduce the noise of meteorological features.

[0075] Meteorological characteristics include direct irradiance, temperature, humidity, wind speed, and atmospheric pressure.

[0076] S32, encoding the time information of the meteorological characteristics;

[0077] S33, performing feature splicing on the denoised meteorological features and the encoded time information to form a first splicing feature;

[0078] After the discrete wavelet transform (DWT), the meteorological characteristics will be divided into two parts: a high-frequency part and a low-frequency part. The high-frequency part is usually noise, and the low-frequency part is a useful feature. This embodiment adopts three-level decomposition and reconstruction to achieve the purpose of noise reduction. The first decomposition: the original meteorological characteristic signal is decomposed into a low-frequency part A1 and a high-frequency part D1. The low-frequency part A1 contains the rough information of the signal, and the high-frequency part D1 contains the detailed information of the signal (usually noise). The low-frequency part A1 obtained by the first decomposition is decomposed again to obtain a new low-frequency part A2 and a high-frequency part D2. Similarly, the low-frequency part A3 and the high-frequency part D3 are obtained after the third decomposition. After three-level decomposition, the signal is decomposed into the following four parts: the lowest frequency part A3, which contains the most important and smoothest information of the meteorological characteristics. The high-frequency parts D1, D2, and D3 contain layer-by-layer details and noise. The noise reduction effect is achieved by thresholding the high-frequency part. Finally, the processed parts are re-merged to obtain a denoised signal. The reconstructed signal will retain the main features of the original signal while removing most of the noise. Then, the denoised meteorological features are concatenated with the encoded time information so that the final prediction PDF can take into account the changes in meteorological features over time, and the interaction between meteorological features and time is considered simultaneously when processing time series data.

[0079] S34, input the first concatenated features into iTransform to obtain new meteorological features, capture the temporal relationship between different meteorological features, learn the fine-grained relationship between different features, avoid confusion and information loss between channels, and improve the representation of features;

[0080] In order to identify the contribution of each meteorological feature under the complex relationship between meteorological features, this embodiment uses iTransformer to consider each meteorological feature separately and embed each feature independently, so that each meteorological feature can be given equal attention. The formula is as follows:

[0081] Among them, z k It is the result of embedding the kth meteorological feature.

[0082] iTransformer mainly consists of three parts: multi-head attention, layer normalization, and feedforward network. The encoded meteorological features will be sent to the multi-head attention module. Each head will focus on different input meteorological features in different ways. Through multi-head parallelism, different attention weights and richer contextual information are learned. Then, each feature is independently standardized through layer normalization to ensure that the input mean of each layer is 0 and the standard deviation is 1, reducing the impact caused by differences in measurement units. The calculation formula is as follows:

[0083] Where Z = {z1,z2,···,z K} means that it contains k coded meteorological features, z k represents the kth meteorological feature that has been independently encoded.

[0084] Then, the feedforward network is used to perform nonlinear transformation on the independent embedding of each position of the output feature of the multi-head attention, which helps the model better understand complex time series data, as shown in the following formula:

[0085] in, It is the feature after multi-head attention and layer normalization processing. W1, W2 are weight matrices, b1, b2 are biases, and f1, f2 represent activation functions. After being processed by the feedforward network, it is layer normalized again, and finally a multi-layer perceptron is used to obtain the final meteorological feature containing time series information.

[0086] Step 4: Reconstruct meteorological characteristics;

[0087] S41, using a multi-head cross-attention mechanism to perform weighted summation on the meteorological features output by iTransformer, so that the meteorological features have the same dimension as the coefficient vector, thereby realizing the reconstruction of the meteorological features;

[0088] Here, the input of the multi-head cross attention mechanism is the coefficient vector obtained in step S22 and the meteorological features obtained in step S34.

[0089] The multi-head cross-attention mechanism can effectively capture the complex relationships between different features and reconstruct the features through weighted summation, thereby enhancing the expressiveness of the features.

[0090] The calculation formula is as follows:

[0091] Q i =X c W i Q

[0092] K i =X m W i K

[0093] V i =X m W i V

[0094]

[0095] Among them, X c represents the input coefficient vector, X m represents the input meteorological characteristics, W i Q , W i K , W i V is the weight matrix of the i-th head (built-in in the multi-head cross attention mechanism). In order to implement multiple heads, each head uses a different weight matrix; softmax is a normalization function used to normalize the output so that it falls within the interval [0,1]; output i represents the output of the i-th head, d n Represents the dimension of the key, which is used for scaling to avoid excessive dot product values. The output of all heads is concatenated and transformed through a linear layer to obtain the output. The formula is as follows:

[0096] F=Concat(output1,output2,···,output i )W o

[0097] Among them, F represents the reconstructed features obtained by multi-head cross attention, Concat(·) represents the concatenation function, and W o represents the output weight matrix.

[0098] Step 5, predict PDF evolution;

[0099] S51, concatenate the coefficient vector and the reconstructed meteorological features to form a more comprehensive and information-rich concatenated feature; weight the concatenated feature and the photovoltaic historical power generation input multi-head cross attention mechanism, reconstruct the concatenated feature, and obtain the predicted coefficient vector through the fully connected layer (FCNN);

[0100] Here, the coefficient vector input to the multi-head cross-attention mechanism and the photovoltaic historical power generation are obtained after self-attention in advance.

[0101] The calculation formula is as follows:

[0102] Q i =X p W i Q

[0103] K i =X j W i K

[0104] V i =X j W i V

[0105] Among them, X p represents the historical power of the input, X j Represents the concatenated features of the input

[0106]

[0107] F=Concat(output1,output2,···,output i )W o

[0108] The mutual influence between features can improve the feature expression ability, thereby helping the model learn richer information and making the prediction accuracy of PDF more accurate.

[0109] S52, the predicted coefficient vector is inversely solved by cubic B-spline interpolation to obtain the predicted photovoltaic power PDF.

[0110] Embodiment 2, a prediction system for the evolution process of a probability density function of photovoltaic power generation power, is used to implement a prediction method for the evolution process of a probability density function of photovoltaic power generation power, including a weight extraction module, a noise reduction module, a multivariate attention mechanism and a prediction module; the weight extraction module realizes the extraction of basis function coefficient vectors, the noise reduction module realizes meteorological feature noise reduction and time information encoding, the multivariate attention mechanism realizes the reconstruction of meteorological features and the reconstruction after the coefficient vector and the meteorological features are spliced, and the prediction module realizes the obtained predicted coefficient vector and the predicted photovoltaic power generation power PDF.

[0111] Photovoltaic power generation is affected by direct irradiance, temperature, humidity and other meteorological characteristics, and has randomness and uncertainty. The present invention reflects the randomness and uncertainty of power changes over time by constructing a probability density function (PDF) evolution model of photovoltaic power generation.

[0112] The present invention uses kernel density estimation to obtain the PDF of photovoltaic power generation, cubic B-spline to calculate the coefficient vector, DWT to perform denoising on historical features, and combines iTransformer and multi-head cross attention mechanism to realize PDF prediction of photovoltaic power generation.

[0113] The present invention captures the relationship between time and the probability distribution of photovoltaic power generation by establishing a photovoltaic power generation probability density evolution model that takes meteorological information into account. It can provide a global probability distribution of photovoltaic power generation and reflect the power change trend and its uncertainty at different times.

[0114] By converting the PDF prediction into the prediction of the coefficient vector, the key features of the PV power generation PDF can be effectively captured as a low-dimensional representation, avoiding redundancy in computation and storage.

[0115] The present invention fully considers the relationship between photovoltaic power generation, meteorological characteristics and coefficient vectors, and adopts a multi-head cross-attention mechanism to reconstruct features. By using a multivariate attention mechanism, the model can obtain multiple perspectives on meteorological features from different attention heads, thereby more accurately capturing important patterns of meteorological data in time and space dimensions. Through a multi-stage attention mechanism, the model can adaptively learn the complex relationship between important features in meteorological information, coefficient vectors and historical power, and ultimately provide accurate and representative feature inputs for prediction tasks, achieving more accurate photovoltaic power PDF prediction.

[0116] The above is only a specific implementation of the present invention, which enables those skilled in the art to understand or implement the present invention. Although detailed descriptions are given with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. A method for predicting the evolution process of photovoltaic power generation power probability density function, characterized by: The following steps are involved: Step 1, obtain the historical probability density function; S11, taking the historical photovoltaic power generation as input, and using kernel density estimation to obtain the probability density function of photovoltaic power generation; Step 2, obtaining the coefficient vector of the basis function; S21, after defining the control points, the probability density function of photovoltaic power generation is fitted using the cubic B-spline interpolation method; S22, calculating the coefficient vector of the basis function corresponding to the control point; Step 3: Enhance the representation of meteorological characteristic information; S31, denoising the meteorological features; S32, encoding the time information of the meteorological characteristics; S33, performing feature splicing on the denoised meteorological features and the encoded time information to form a first splicing feature; S34, inputting the first concatenated features into iTransform to obtain new meteorological features, capturing the temporal relationship between different meteorological features, and enhancing the representation of meteorological information; Step 4: Reconstruct meteorological characteristics; S41, using a multi-head cross-attention mechanism to perform weighted summation on the meteorological features output by iTransformer, so that the meteorological features have the same dimension as the coefficient vector, thereby realizing the reconstruction of the meteorological features; S42, the input of the multi-head cross attention mechanism is the coefficient vector obtained in step S22 and the meteorological features obtained in step S34; Step 5: predict the evolution of probability density function; S51, splicing the coefficient vector and the reconstructed meteorological features to form a more comprehensive and information-rich splicing feature; the splicing feature and the photovoltaic historical power generation are input into the multi-head cross attention mechanism for weighted calculation, reconstructing the splicing feature, and obtaining the predicted coefficient vector through the fully connected layer. The coefficient vector and photovoltaic historical power generation input by the multi-head cross attention mechanism are obtained after self-attention in advance; S52, performing inverse solution of the predicted coefficient vector by cubic B-spline interpolation to obtain a predicted photovoltaic power generation power probability density function.

2. A method for predicting the evolution process of photovoltaic power generation power probability density function according to claim 1, characterized in that: Meteorological characteristics include direct irradiance, temperature, humidity, wind speed, and atmospheric pressure.

3. A method for predicting the evolution process of photovoltaic power generation power probability density function according to claim 1, characterized in that: The kernel density estimation formula is: in, represents the probability density estimate at point x, where x is the photovoltaic power generation value to be estimated; n is the total number of data points; κ(·) represents the kernel function; h is the bandwidth parameter used to control the width of the kernel; x i is the actual power value of the ith photovoltaic cell.

4. A method for predicting the evolution process of photovoltaic power generation power probability density function according to claim 1, characterized in that: The kernel function uses the Gaussian kernel function, and the formula is in is a standardized variable, indicating the actual x i The relative distance to x.

5. A method for predicting the evolution process of photovoltaic power generation power probability density function according to claim 1, characterized in that: Discrete wavelet transform is used to reduce the noise of meteorological characteristics.

6. A method for predicting the evolution process of photovoltaic power generation power probability density function according to any one of claims 1 to 5, characterized in that: The control points are selected in equal proportion from the probability density function curve.

7. A method for predicting the evolution process of photovoltaic power generation power probability density function according to claim 6, characterized in that: S21 uses cubic B-spline to fit the obtained PDF. The fitting formula is as follows: Where S(x) is the fitted PDF, P i is the i-th control point, B i,3 (x) is the cubic B-spline basis function corresponding to the ith control point, which is constructed through a recursive relationship. The specific form is: Among them, x i is the i-th node vector, and under a given x value, the coefficient vector M is solved by the interpolation condition; The relationship between the coefficient vector of the S22 cubic B-spline and the control points is expressed by the following linear equation system: M×{P0,P1,…,P n }={S0,S1,…,S n }, the coefficient vector M is obtained by satisfying the conditions of value continuity, first-order derivative continuity and second-order derivative continuity of the basis function, which represents the relationship between the control point and the interpolation curve, and the coefficient vector is extracted as the coefficient vector.

8. A prediction system for the evolution process of photovoltaic power generation power probability density function, characterized by: A method for predicting the evolution process of a photovoltaic power probability density function as described in any one of claims 1 to 7, comprising a weight extraction module, a denoising module, a multivariate attention mechanism and a prediction module; the weight extraction module realizes the extraction of basis function coefficient vectors, the denoising module realizes meteorological feature denoising and time information encoding, the multivariate attention mechanism realizes the reconstruction of meteorological features and the reconstruction after the coefficient vector and the meteorological features are spliced, and the prediction module realizes the obtained predicted coefficient vector and the predicted photovoltaic power PDF.

Images