Photovoltaic power generation power condition data generation method based on improved diffusion model

Abstract

The invention discloses a photovoltaic power generation power working condition data generation method based on an improved diffusion model. The method comprises the following steps: S1, obtaining photovoltaic power generation power working condition data, photovoltaic power station geographical environment data and photovoltaic power station geographical environment data; s2, preprocessing the photovoltaic power generation power working condition data and the photovoltaic power station geographical environment data; s3, dividing a geographic region where the photovoltaic power station is located according to the equal-proportion power generation area to obtain regional photovoltaic power generation power working condition data and regional photovoltaic power station geographic environment data, performing clustering analysis on the regional photovoltaic power generation power working condition data, and constructing prompt words according to a clustering result and the regional photovoltaic power station geographic environment data; and S4, constructing a power prediction model according to the time step, the photovoltaic power generation power working condition data, the photovoltaic power station geographical environment data and the cue words, and outputting photovoltaic power generation power working condition prediction data. The meaning of the photovoltaic working condition is understood through the text encoder, and the interactivity between people and the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation power condition generation, and in particular to a photovoltaic power generation power condition data generation method based on an improved diffusion model. Background Art

[0002] As a key infrastructure, the power system plays a vital role in the development of modern society. In recent years, with the continuous development of society and the growing demand for electricity, the field of new energy has also developed vigorously. In particular, in the field of photovoltaic power generation, the installed capacity has firmly stabilized in the world's first position, and the innovation and application of photovoltaic technology are also at the forefront of the world.

[0003] The main goal of photovoltaic power generation is to obtain a stable and sufficiently large output power. Only when photovoltaic power generation equipment guarantees a stable and long-term output power can sufficient economic benefits be obtained. Photovoltaic power is particularly dependent on the surrounding environmental factors, that is, the working conditions of the power generation, especially external factors such as weather that will significantly affect the intensity of solar radiation. Therefore, for large-scale analysis and prediction of photovoltaic power generation, it is crucial to study various data of photovoltaic power generation under different environmental factors. However, when practitioners make specific analysis based on photovoltaic environmental information, they often find that the relevant data is insufficient, such as the case where the amount of data under certain special weather conditions is significantly smaller. As a result, subsequent large-scale data analysis causes large deviations due to data imbalance.

[0004] At present, the industry mainly uses two types of methods for photovoltaic data augmentation. One is the relatively traditional data upsampling method, such as repeated sampling, Lagrange interpolation, spline interpolation, etc. This type of method is widely used because it is relatively simple and has a fast calculation speed. However, the disadvantages are very obvious: repeated sampling directly reuses the original data without changing the data itself. This method can easily cause the step effect and distortion of the data; Lagrange interpolation is a method based on polynomial interpolation, which is prone to Runge effect and has poor numerical stability. Spline interpolation is a piecewise low-order polynomial interpolation method, which has certain deficiencies in boundary processing and numerical accuracy. At the same time, most interpolation methods find it difficult to take into account the mutual influence between all global variables. Traditional methods are also generally difficult to generate new data. When using a data set constructed with the help of an interpolation method to make data public, it is easy for the public to reverse the real data situation, which causes the leakage of confidential information.

[0005] The other is the sample synthesis method. Most of these methods are new data augmentation technologies explored in the past 20 years. They are mainly improvements made to traditional data augmentation, alleviating the shortcomings of traditional methods in terms of accuracy, stability, and understanding of coupling between variables. For example, the SMOTE method, although this method is also based on the interpolation method, combines the nearest neighbor method in machine learning technology and perceives global variable information to a certain extent. However, it still fails to solve the boundary and result stability problems. Methods such as CTGAN use the generative adversarial network algorithm in deep learning, which has made progress in boundary problems and deep feature extraction. However, new problems such as unstable training and complex structural design have emerged. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a method for generating photovoltaic power condition data based on an improved diffusion model.

[0007] A method for generating photovoltaic power condition data based on an improved diffusion model comprises the following steps: S1: Acquire photovoltaic power generation power condition data and photovoltaic power station geographical environment data, the photovoltaic power station geographical environment data includes weather data, timestamp, geographical location information, equipment parameters and photovoltaic power station environmental data; S2: Preprocessing of photovoltaic power generation data and photovoltaic power station geographical environment data, including feature selection, outlier processing, feature coding, data standardization and missing value processing; S3: Divide the geographical area where the photovoltaic power station is located according to the proportional power generation area, obtain the regional photovoltaic power generation power condition data and the regional photovoltaic power station geographical environment data, perform cluster analysis on the regional photovoltaic system power condition data, and construct prompt words based on the clustering results and the regional photovoltaic power station geographical environment data; S4: Based on the time step, photovoltaic power generation condition data, photovoltaic power station geographical environment data and prompt words, and using machine learning algorithm, a power prediction model is constructed to output photovoltaic power generation condition prediction data.

[0008] Preferably, in S2, feature selection is specifically to use the Boruta algorithm to remove data features in the photovoltaic power station geographical environment data whose impact on photovoltaic power generation power condition data is less than a preset importance threshold, and select data features whose impact importance is greater than the preset importance threshold.

[0009] Preferably, in S2, outlier processing is specifically performed by calculating the Z score of each data point in the photovoltaic power generation condition data and the photovoltaic power station geographical environment data. If the absolute value of the Z score is greater than a preset threshold, the data point is determined to be an outlier and the outlier is deleted.

[0010] Preferably, in S2, feature encoding is specifically, natural number encoding is used for ordered features in photovoltaic power generation condition data and photovoltaic power station geographical environment data; and unique hot encoding is used for unordered features in photovoltaic power station geographical environment data.

[0011] Preferably, in S3, the specific calculation steps of the proportional power generation area are: S31: Analyze the weather data, environmental data and geographical location information of the photovoltaic power station to obtain the external factor coefficient of the power station; S32: Analyze the photovoltaic conversion efficiency, inverter efficiency and cable loss ratio of the photovoltaic module to obtain the internal factor coefficient of the power station; S33: multiplying the normalized coefficient of external factors of the power station and the internal factors coefficient of the power station; S34: Convert the multiplication result in S33 to the standard photovoltaic power generation area in proportion to obtain a proportional power generation area.

[0012] Preferably, it is characterized in that, in S4, the specific steps of constructing the power prediction model are: S41: preset time step; S42: Use CLIP text encoder to convert the prompt words in S3 into vector representation; S43: using the Transformer encoder to encode the photovoltaic power condition data within the time step range, extract the condition mode and condition characteristics, and form a time series representation vector; S44: using a machine learning algorithm, inputting the time step, the CLIP text encoder output, and the Transformer encoder output into a power prediction model; S45: When the power prediction model is output, different output layer configurations are selected according to the selected preset task type.

[0013] Preferably, in S45, different output layer configurations are selected, specifically: If the preset task type is a classification task, use the Softmax activation function to output the power generation mode; If the preset task type is a regression task, a linear activation function is used to output the photovoltaic power generation condition prediction data.

[0014] The invention discloses a method for generating photovoltaic power condition data based on an improved diffusion model, which has the following beneficial effects: selecting features through a Boruta algorithm, using a Z score method to clear free variables, using a natural number / unique hot encoding method to encode classification and discrete features, using Gaussian quantile changes to convert continuous data distribution into Gaussian distribution, using a random forest-based chain equation multiple interpolation method to fill in default values, greatly simplifying the original data and enriching the integrity of the data; when dividing the data, the external and internal influencing factors of the photovoltaic power station are included in the data consideration range. And 24 hours is used as the data interception size, and the time granularity size is one minute. More real data can be accurately analyzed and generated; the meaning of photovoltaic conditions is understood through a text encoder, including but not limited to the understanding and analysis of prompt words such as geographical location, weather factors, and seasonal factors, which improves the interactivity between people and models and enhances the accuracy of model data results. By using a transformer encoder, the influence of the preceding time on the subsequent results is combined in the generation process, and the mutual influence between various variables is also enhanced, and the results are more in line with reality. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] 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, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 A method flow chart of a method for generating photovoltaic power condition data based on an improved diffusion model provided by the present invention; DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0019] refer to Figure 1 , a method for generating photovoltaic power condition data based on an improved diffusion model, comprising the following steps: S1: Acquire photovoltaic power generation power condition data and photovoltaic power station geographical environment data, wherein the photovoltaic power station geographical environment data includes weather data, timestamp, geographical location information, equipment parameters and photovoltaic power station environmental data; wherein the weather data includes annual average sunshine intensity, rainfall and wind speed; the geographical location information of the photovoltaic power station includes latitude, longitude, altitude, etc.; the geographical environment data includes geographical environment such as terrain, soil type, vegetation coverage, etc.; S2: Preprocessing of photovoltaic power generation data and photovoltaic power station geographical environment data, including feature selection, outlier processing, feature coding, data standardization and missing value processing; S3: Divide the geographical area where the photovoltaic power station is located according to the proportional power generation area, obtain the regional photovoltaic power generation power condition data and the regional photovoltaic power station geographical environment data, perform cluster analysis on the regional photovoltaic system power condition data, and construct prompt words based on the clustering results and the regional photovoltaic power station geographical environment data; Among them, the regional photovoltaic power generation power condition data and the regional photovoltaic power station geographical environment data are intercepted every 24 hours, and the data of each area is summarized and intercepted every 24 hours. The time granularity is unified to 1 minute, and a matrix of n* (60*24=1440) dimensions can be obtained as a single sample, where n is the number of characteristic variables finally selected.

[0020] Specifically, clustering algorithms such as K-means and hierarchical clustering are used to cluster the power operating condition data of regional photovoltaic systems, identify areas with different power output modes, and find out groups of power stations with similar power generation capabilities; based on the clustering results, the number of samples in each cluster, the characteristic distribution of power operating condition data and other information are checked to confirm the characteristics of different clusters such as high-yield power areas or low-yield power areas; combined with weather data, timestamps, geographical location information of photovoltaic power stations and geographical environment data, prompt words are constructed, such as high-yield power areas with sufficient sunshine and flatness; south-facing slopes suitable for winter power generation; and high-yield power areas with a combination of high sunshine intensity and low wind speed.

[0021] S4: According to the time step, photovoltaic power generation power condition data, photovoltaic power station geographical environment data and the prompt words, and using a machine learning algorithm, a power prediction model is constructed to output photovoltaic power generation power condition prediction data.

[0022] The invention discloses a method for generating photovoltaic power condition data based on an improved diffusion model. The method selects features through a Boruta algorithm, uses a Z-score method to clear free variables, uses a natural number / unique hot encoding method to encode classification and discrete features, uses Gaussian quantile changes to convert continuous data distribution into Gaussian distribution, and uses a chain equation multiple interpolation method based on random forests to fill in default values, thereby greatly simplifying original data and enriching data integrity. When dividing data, external and internal influencing factors of photovoltaic power stations are included in the data consideration scope. 24 hours is used as the data interception size, and the time granularity size is one minute, so that more real data can be accurately analyzed and generated. A text encoder is used to understand the meaning of photovoltaic conditions, including but not limited to the understanding and analysis of prompt words such as geographical location, weather factors, and seasonal factors, so as to improve the interactivity between people and models and enhance the accuracy of model data results. By using a transformer encoder, the influence of the preceding time on the subsequent results is combined in the generation process, and the mutual influence between various variables is also enhanced, so that the results are more in line with reality.

[0023] In a preferred embodiment, in S2, feature selection, specifically, uses the Boruta algorithm to remove data features in the photovoltaic power station's geographical environment data whose importance to the photovoltaic power generation power condition data is less than a preset importance threshold, such as soil type, and selects data features whose importance is greater than the preset importance threshold, such as the location latitude, longitude, and altitude of the photovoltaic power station.

[0024] In a preferred embodiment, in S2, outlier processing is specifically performed by calculating the Z score of each data point in the photovoltaic power generation power condition data and the photovoltaic power station geographical environment data. If the absolute value of the Z score is greater than a preset threshold, the data point is determined to be an outlier and the outlier is deleted.

[0025] Among them, points with absolute values ​​greater than 3 are considered outliers, and the calculation formula of Z score is as follows: Z=((X-μ)) / σ, where X is the data point, μ is the mean, and σ is the standard deviation.

[0026] In a preferred embodiment, in S2, feature encoding is specifically, for ordered features in photovoltaic power generation condition data and photovoltaic power station geographical environment data, such as seasonal change data, natural number encoding is used; for features without obvious order in photovoltaic power station geographical environment data, such as weather data including rainy and sunny days, one-hot encoding is used.

[0027] In a preferred embodiment, data standardization is to perform Gaussian quantile transformation on continuous features and features encoded with natural numbers, arrange all data from large to small, and calculate the quantile P of the i-th data. i The formula is as follows: , Where Φ is the cumulative distribution function of the standard normal distribution, is the quantile function, is the starting quantile.

[0028] Specifically, missing value processing involves filling in the default value using a random forest-based chain equation multiple interpolation method. For any variable with a default value, a random forest is constructed, and other variables are used as predictors. The constructed random forest is then used to predict the default value, and the above process is repeated multiple times, and the results of multiple interpolations are analyzed and integrated.

[0029] In a preferred embodiment, in S3, the specific calculation steps of the proportional power generation area are: S31: Analyze the weather data, environmental data and geographical location information of the photovoltaic power station to obtain the external factor normalization coefficient of the power station; S32: Analyze the photovoltaic conversion efficiency, inverter efficiency and cable loss ratio of the photovoltaic module to obtain the internal factor coefficient of the power station; S33: multiplying the normalized coefficient of external factors of the power station and the internal factors coefficient of the power station; Among them, the calculation formula of the total coefficient of internal factors outside the power station η_total is: η_total = η_outside × η_inside, η_external is the normalized coefficient of external factors of the power station, and η_internal is the coefficient of internal factors of the power station; S34: Convert the multiplication result in S33 to the standard photovoltaic power generation area in proportion to obtain a proportional power generation area.

[0030] Among them, the standard photovoltaic power generation area [Area]_standard is 10 hectares, and the calculation formula for the ratio conversion [Area]_conversion is as follows: [Area]_conversion = [Area]_standard × η_total, where η_total is the total factor of the power station.

[0031] In a preferred embodiment, in S4, the specific steps of constructing a power prediction model are: S41: preset time step; S42: Use CLIP text encoder to convert the prompt words in S3 into vector representation; S43: Use the Transformer encoder to encode the photovoltaic power condition data from 0 hours to the previous time step, extract the operating mode and operating characteristics, and form a time series representation vector; the operating mode refers to different power generation efficiencies, power output fluctuations, and power generation performance under specific weather conditions, including high-efficiency power generation mode and low-efficiency power generation mode; the operating characteristics include the temperature of the photovoltaic module, sunshine intensity, wind speed, tilt angle and time of the photovoltaic module; S44: using a machine learning algorithm, inputting the time step, the CLIP text encoder output, and the Transformer encoder output into a power prediction model; S45: When the power prediction model is output, a multi-layer perceptron is used to select different output layer configurations according to the selected preset task type.

[0032] In a preferred embodiment, in S45, different output layer configurations are selected, specifically: If the preset task type is selected as a classification task, the Softmax activation function is used to output the power generation mode, such as whether it is high power generation, low power generation, or normal power generation at a certain moment; The final output of the classified data is passed through the Softmax activation function to determine the one-hot encoding result. The variables encoded with natural numbers are transformed by the inverse Gaussian quantile and then rounded.

[0033] Other continuous variables are directly output after Gaussian quantile transformation. Finally, the results are summarized and output.

[0034] If the preset task type is selected as a regression task, a linear activation function is used to output photovoltaic power generation condition prediction data, such as predicting the power output in the next hour.

[0035] Construct a power prediction model, specifically: First, the parameters of the power prediction model should be initialized. For continuous variables that have been Gaussian initialized, the original noise diffusion method is selected. Here, the diffusion time step T and the diffusion coefficient β_t are initialized first, and the noise ε that conforms to the standard normal distribution is set. For variables encoded using one-hot codes, the polynomial diffusion method is selected, and a cubic polynomial function is selected: N(t)=at^3+bt^2+ct+d. The least squares method is used to find the appropriate polynomial parameters, thereby determining the four parameters a, b, c, and d. To construct the input layer of the power prediction model, prompt words are needed as prompts to adjust the output results. A pre-trained text encoder is added. Here, the CLILP text encoder is used. Then, an additional transformer encoder is used to encode the content from 0 hour to the previous time t'-1 in a sample. Specifically, a 24-time loop structure is used, and each loop processes the data content of one hour. When in the next loop, the results from the first hour to the previous loop are processed to capture the mutual influence between long-distance elements in the time series, obtain the patterns and characteristics in the historical working conditions, and thus better predict the working condition information of the next hour. Among them, the historical operating condition sequence is: h_(1:t'-1)=[h_1,h_2,...,h_(t'-1)], The encoded representation vector after the transformer encoder: e_(t'-1)=TRANSFORMER_ENCODER(h 1:t'-1 ), A multi-layer perceptron model is used, where the input is the time step t, the output of the transformer encoder e_(t'-1), and the output of the text encoder. The final output is the preliminary prediction result of 24 hours; Finally, the output of the classified data is passed through the Softmax activation function to determine the one-hot code result. The variables encoded with natural numbers are rounded after the inverse Gaussian quantile change. Other continuous variables are directly output after the Gaussian quantile change. Finally, the output results are summarized.

[0036] During the training of the power prediction model, the mean square error of ordinary noise and polynomial noise is used as the loss function: , in, is the noise of standard normal distribution, is the polynomial noise, L is the loss function (θ is a form of parameter representation), To predict normally distributed noise, To predict polynomial noise, To find the expected value formula, we average over all possible initial states X0, noise ϵ, and time steps t.

[0037] For continuous variables, the diffusion process is as follows: First, the noise coefficient: , is the diffusion coefficient, is the noise coefficient at time step t, is the cumulative noise coefficient at time step t; , in, a s is the noise coefficient at time step s; Then the diffusion result of the continuous numerical matrix X at each time step t can be expressed as follows: ， in, is the diffusion result of the continuous numerical matrix X at time step t, is the noise of standard normal distribution.

[0038] Specifically, the continuous numerical matrix X actually represents the ordered features in the photovoltaic power generation condition data encoded by natural numbers and the geographical environment data of the photovoltaic power station, such as seasonal variation data. The Gaussian diffusion model used here.

[0039] For discrete variables, the diffusion process is as follows: The noise factor N(t) formula is: , Where Z is a random matrix that follows a standard normal distribution.

[0040] Then the diffusion result of the one-hot encoding matrix O at each time step t can be expressed as follows: , Specifically, O represents the use of unique hot encoding for the features of the photovoltaic power station geographical environment data without obvious order, such as weather data including rainy and sunny data. Here, the polynomial diffusion model is used, N(t) is the noise coefficient, O t It is the diffusion result of the one-hot encoding matrix O at each time step t.

[0041] Reverse denoising process: perform the following steps from time step t=T to 0: Estimate Noise: Use the model to estimate the noise at time step t and , Backward update: using estimated noise Update data, , in, The continuous numerical matrix X diffuses the result at time step t, represents the state of the continuous numerical matrix X at time t-1 after it is updated, is the noise coefficient at time step t, is the cumulative noise coefficient at time step t. To predict normally distributed noise; , Among them, O ｔis the diffusion result of the one-hot encoding matrix O at each time step t, O ｔ－１ is the result of the previous time step of the one-hot encoding matrix O at each time step t, is the prediction polynomial noise.

[0042] Randomly select the power condition change of a block on any day from the training data, and the corresponding condition prompt word and a randomly selected noise time step, randomly extract a noise according to the noise addition formula, and input them into the power prediction model together. Calculate the loss, calculate the gradient, back propagate, and finally converge.

[0043] Use the trained power prediction model for large-scale batch generation. Determine the working condition prompt word of the sample type that needs to be amplified, input it with Gaussian noise and the current time step to the trained model, repeat the set number of time steps, and finally output the amplification result.

[0044] Finally, the batch generated data is verified to see if they meet the requirements. The present invention uses a variety of methods to verify the data. The first method is to build a machine learning model using the original data and the augmented data respectively. If the final evaluation index is small enough, the data is considered to meet the requirements.

[0045] Specifically, we use a variety of machine learning models such as random forest, support vector machine, ridge regression, etc. for targeted training. Finally, we calculate the accuracy, F1 score, explained variance score and other evaluation indicators on the original test set to explore whether the model built with augmented data can easily cope with the test set of the original data set. The smaller the gap between the last two model indicators, the better.

[0046] The second method is to calculate the evaluation index of the distribution of the original data set and the generated data set. If it is small enough, it also meets the requirements. If both verification methods meet the index requirements, it proves that the data can be directly applied and the model is successfully trained. The evaluation indicators used here include the difference between Jensen-Shannon divergence, optimal transmission distance and pairwise correlation for comparative judgment.

Claims

1. A method for generating photovoltaic power condition data based on an improved diffusion model, characterized in that: The steps include: S1: Acquire photovoltaic power generation power condition data and photovoltaic power station geographical environment data, wherein the photovoltaic power station geographical environment data includes weather data, timestamp, geographical location information, equipment parameters and photovoltaic power station environmental data; S2: preprocessing the photovoltaic power generation power condition data and the photovoltaic power station geographical environment data, including feature selection, outlier processing, feature coding, data standardization and missing value processing; S3: Divide the geographical area where the photovoltaic power station is located according to the proportional power generation area, obtain regional photovoltaic power generation power condition data and regional photovoltaic power station geographical environment data, perform cluster analysis on the regional photovoltaic system power condition data, and construct prompt words according to the clustering results and the regional photovoltaic power station geographical environment data; S4: According to the time step, the photovoltaic power generation condition data, the photovoltaic power station geographical environment data and the prompt word, a machine learning algorithm is used to construct a power prediction model, which can output photovoltaic power generation condition prediction data.

2. The method for generating photovoltaic power condition data based on an improved diffusion model according to claim 1, characterized in that: In S2, the feature selection is specifically to use the Boruta algorithm to remove data features in the photovoltaic power station geographical environment data whose impact on the photovoltaic power generation power condition data is less than a preset importance threshold, and select data features whose impact importance is greater than the preset importance threshold.

3. The method for generating photovoltaic power condition data based on an improved diffusion model according to claim 1, characterized in that: In S2, the outlier processing is specifically to calculate the Z score of each data point in the photovoltaic power generation power condition data and the photovoltaic power station geographical environment data. If the absolute value of the Z score is greater than a preset threshold, the data point is determined to be an outlier and the outlier is deleted.

4. The method for generating photovoltaic power condition data based on an improved diffusion model according to claim 1, characterized in that: In S2, the feature coding is specifically that the ordered features in the photovoltaic power generation power condition data and the photovoltaic power station geographical environment data are coded using natural numbers; and the unordered features in the photovoltaic power station geographical environment data are coded using one-hot coding.

5. The method for generating photovoltaic power condition data based on an improved diffusion model according to claim 1, characterized in that: In S3, the specific calculation steps of the proportional power generation area are: S31: Analyze the weather data, environmental data and geographical location information of the photovoltaic power station to obtain the external factor coefficient of the power station; S32: Analyze the photovoltaic conversion efficiency, inverter efficiency and cable loss ratio of the photovoltaic module to obtain the internal factor coefficient of the power station; S33: multiplying the power station external factor normalization coefficient and the power station internal factor coefficient; S34: converting the multiplication result in S33 into a standard photovoltaic power generation area in proportion to obtain a proportional power generation area.

6. The method for generating photovoltaic power condition data based on an improved diffusion model according to claim 1, characterized in that: In S4, the specific steps of constructing the power prediction model are: S41: preset time step; S42: using CLIP text encoder to convert the prompt word in S3 into a vector representation; S43: using the Transformer encoder to encode the photovoltaic power condition data within the time step range, extract the condition mode and condition characteristics, and form a time series representation vector; S44: using a machine learning algorithm, inputting the time step, the CLIP text encoder output, and the Transformer encoder output into a power prediction model; S45: When the power prediction model is output, different output layer configurations are selected according to the selected preset task type.

7. The method for generating photovoltaic power condition data based on an improved diffusion model according to claim 6, characterized in that: In S45, different output layer configurations are selected, specifically: If the preset task type is a classification task, use the Softmax activation function to output the power generation mode; If the preset task type is a regression task, a linear activation function is used to output the photovoltaic power generation condition prediction data.

Images