Photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence

By combining generative adversarial network and black box model, the application difficulties of photovoltaic power generation prediction models in different geographical locations are solved, the prediction accuracy and robustness are improved, the calculation cost is reduced, and the high-quality enhancement and universal photovoltaic power generation data are achieved. ability to transform.

CN118899842APending Publication Date: 2024-11-05TONGJI UNIV
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411036329.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-05

AI Technical Summary

Technical Problem

In the case of difficult data collection and poor data quality, existing photovoltaic power generation prediction models are difficult to generalize to different geographical locations, resulting in poor prediction results and increasing the time and calculation cost of model application.

Method used

Using the solar radiation data enhancement method based on the generative adversarial network, combined with the black box model of the physical model, the factors of random changes in solar radiation are generated by probabilistic modeling, the enhanced solar radiation data set is expanded, and the application scope is expanded, and the timing photovoltaic power generation data set is trained Prediction model to improve prediction accuracy and robustness.

Benefits of technology

It improves the generalization capability and accuracy of photovoltaic power generation prediction models, reduces computational costs, and can provide high-quality prediction results in areas with similar atmospheric states and cloud levels.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118899842A_ABST
    Figure CN118899842A_ABST
Patent Text Reader

Abstract

The invention relates to a photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence, and the method comprises the following steps: obtaining solar radiation data which comprise direct solar radiation, scattered radiation and total radiation data; performing data enhancement on the direct solar radiation data by using a generative adversarial network; based on the direct radiation data after data enhancement, utilizing a black box model to predict scattered radiation and total radiation, and constructing an enhanced solar radiation data set; calculating the photovoltaic power generation amount based on the enhanced solar radiation data set, and constructing a time sequence photovoltaic power generation data set; and training a photovoltaic power generation prediction model by using the time sequence photovoltaic power generation data set, and realizing photovoltaic system power generation prediction based on the trained photovoltaic power generation prediction model. Compared with the prior art, the method has the advantages of high photovoltaic data augmentation capability, low calculation cost, high photovoltaic power generation prediction accuracy and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation prediction, and in particular to a photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence. Background Art

[0002] Compared with other renewable energy utilization technologies, photovoltaic power generation technology, as a relatively mature renewable energy utilization technology, has been widely used around the world. However, since photovoltaic power generation is greatly affected by weather conditions, seasonal changes and diurnal cycles, its output power has high volatility and uncertainty. Therefore, photovoltaic power generation prediction is particularly important in the stable operation and energy management of power systems: 1) Accurate photovoltaic power generation prediction helps maintain the balance and stability of the power system. Since the power system needs to maintain a real-time balance between supply and demand, accurate photovoltaic power generation prediction enables the power system to adjust the output of other power sources in advance to ensure the stable operation of the power system. For example, when it is predicted that the output of photovoltaic power generation may drop sharply, traditional thermal power or hydropower units can be dispatched in advance to make up for the lack of photovoltaic power generation. 2) Photovoltaic power generation prediction can guide the power system to optimize the scheduling and configuration of power resources and improve its renewable energy absorption capacity. With the increasing proportion of renewable energy in the power system, its volatility and uncertainty will bring great challenges to the scheduling and operation of the power system. Based on accurate photovoltaic power generation prediction, the power system can better arrange the operation mode of the power grid, optimize power production and transmission plans, reduce unnecessary start-up and shutdown, improve the absorption capacity of renewable energy and thus improve energy utilization efficiency. 3) Photovoltaic power generation prediction is a key technology for intelligent power grids: Smart grids and microgrids are based on advanced prediction and control technologies to achieve efficient management and optimal utilization of energy. As the proportion of renewable energy in the power grid increases, photovoltaic power generation prediction, especially short-term prediction, provides accurate power generation data for the power grid, helping it to better coordinate multiple energy resources and further improve the flexibility and reliability of the system.

[0003] However, various existing photovoltaic power generation prediction models require a large amount of photovoltaic power generation data or solar radiation data as input. However, the difficulties in data collection and poor data quality in actual projects hinder the application of such models in actual engineering.

[0004] In addition, since the solar radiation reaching the surface is affected by the geographical location of the area, the probability distribution of photovoltaic power generation sequence modeled by the generative model can only be applied to a certain geographical space. For different geographical locations and regions with different solar radiation distribution characteristics, it is necessary to re-model and train them based on their original data, which increases the time and computational cost of applying such models. Summary of the invention

[0005] The purpose of the present invention is to provide a photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence. Aiming at the shortcomings of photovoltaic data augmentation methods that use generative models to model the probability distribution of photovoltaic power generation sequences in a certain place and then obtain a large number of photovoltaic power generation scenes, such as poor generalization ability and high time and computational costs, a solar radiation data generation method based on a generative model is proposed. By probabilistically modeling the factors that cause random changes in solar radiation, the spatial scope of application is expanded from a single site to areas with similar atmospheric conditions and cloud cover levels. Aiming at the problem that actual photovoltaic data is lacking or solar radiation data cannot be obtained, making it difficult to predict photovoltaic power generation, a black box model combined with a physical model is proposed to achieve accurate prediction of solar radiation data, thereby generating photovoltaic power generation data, solving the problem of poor prediction results caused by insufficient photovoltaic power generation data. High-quality augmentation of the prediction model training set improves the accuracy and robustness of the prediction model.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence includes the following steps:

[0008] S1, obtaining solar radiation data, wherein the solar radiation data includes solar direct radiation, scattered radiation and total radiation data;

[0009] S2, using generative adversarial networks to enhance the direct solar radiation data;

[0010] S3, based on the enhanced direct radiation data, the black box model is used to predict the diffuse radiation and total radiation to construct an enhanced solar radiation dataset;

[0011] S4, calculates photovoltaic power generation based on the enhanced solar radiation dataset and constructs a time series photovoltaic power generation dataset;

[0012] S5, using the time series photovoltaic power generation data set to train the photovoltaic power generation prediction model, and realizing the photovoltaic system power generation prediction based on the trained photovoltaic power generation prediction model.

[0013] The relationship between the direct solar radiation, diffuse radiation and total radiation is:

[0014] GHI=DNI*sin(h0)+DIF

[0015] Among them, DNI is the normal solar direct radiation, DIF is the solar diffuse radiation, h0 is the solar altitude angle, DNI*sin(h) is the direct radiation DHI on the horizontal plane, and GHI is the total radiation on the horizontal plane.

[0016] The direct solar radiation includes a deterministic part and a random part, wherein the deterministic part is determined by the relative position between the earth and the sun, and the random part includes the randomness of the change of atmospheric composition described by the atmospheric transparency coefficient and the randomness of the change of cloud cover in the atmosphere described by the attenuation coefficient.

[0017] The method for determining the atmospheric transparency coefficient comprises the following steps:

[0018] A1, calculate the initial value of the atmospheric transparency coefficient according to the following formula:

[0019]

[0020] Among them, P m is the initial value of the atmospheric transparency coefficient at a certain place at a certain time, DNI clearsky It is the direct ground radiation under clear and cloudless conditions, in W / m 2 ; EDNI is the extraterrestrial normal direct radiation, in W / m 2 ; m is the atmospheric optical mass number, which is used to characterize the relative distance that sunlight travels to reach the ground, and its value varies with the solar altitude angle h;

[0021] A2, the hourly solar altitude angle is corrected to reflect only the change of atmospheric state: the atmospheric transparency under clear sky condition and solar altitude angle h=30° is divided into several levels, and the change of the initial value of the atmospheric transparency coefficient of the ground sun hourly with the solar altitude angle under different levels is statistically sorted, and the power function is used for fitting to obtain the fitting curve;

[0022] A3, further extrapolate the fitting curve and correct the atmospheric transparency coefficient. The specific method is as follows:

[0023] 1) Assume that the solar altitude angle at a certain moment is h * , the initial value of the atmospheric transparency coefficient is

[0024] 2) Calculate h=h under different typical meteorological conditions based on the fitting curve * The atmospheric transparency coefficient at time is expressed as Among them, n is the number of typical meteorological phenomena;

[0025] 3) According to the fitting curve, the atmospheric transparency coefficient at h = 30° under different typical meteorological conditions is calculated and recorded as

[0026] 4) With Y as the dependent variable and X as the independent variable, the regression equation is obtained by linear fitting using the least squares method: P2 = a*P m +b, where a and b are regression equation coefficients;

[0027] 5) The actual atmospheric transparency coefficient Substitute the regression equation obtained in the previous step to obtain the corrected atmospheric transparency coefficient P2.

[0028] The attenuation factor is defined as the ratio of the actual radiation reaching the surface to the radiation in a clear, cloudless state:

[0029]

[0030] Among them, DNI clearsky It is the direct ground radiation under cloudless conditions, and DNI is the direct solar radiation from the normal direction.

[0031] The attenuation coefficient is divided into three categories according to the weakening effect of cloud cover on direct radiation. The first category is the case where the daytime normal solar direct radiation is 0, the second category is the case where the daytime normal solar direct radiation is not weakened by clouds, and the third category is the case where the daytime normal solar direct radiation changes randomly. The first and second categories are directly represented by definition, and the third category is represented by learning using a generative adversarial network.

[0032] Based on the upper / lower quartiles of the mean and the median of the variance of the cloud description parameters, the diurnal variation characteristics of cloud cover are divided into 8 categories. The generative adversarial network of cloud cover and attenuation coefficient is trained based on the categories of the diurnal variation characteristics of cloud cover to generate the diurnal variation feature sequence and attenuation coefficient sequence of cloud cover.

[0033] The method of predicting scattered radiation and total radiation using a black box model based on the direct radiation data after data enhancement is specifically as follows: the maximum information coefficient between direct radiation and scattered radiation, and between direct radiation and total radiation are calculated respectively, and according to the comparison results of the maximum information coefficients, a black box model for predicting scattered radiation based on direct radiation and a black box model for predicting total radiation based on direct radiation are trained respectively.

[0034] The black box model is a generative adversarial network prediction model, and the generator of the generative adversarial network prediction model adopts a transformer structure.

[0035] The step S4 is specifically as follows: according to the enhanced solar radiation data set, the total radiation on the slope per unit area is calculated by the photovoltaic inclination angle, and then multiplied by the total area of ​​the photovoltaic panels and the total efficiency coefficient of the system to obtain the photovoltaic power generation calculation amount, the actual photovoltaic power generation is divided by the calculated photovoltaic power generation to obtain the total efficiency coefficient at that moment, the calculated efficiency coefficient frequency is counted to obtain the efficiency coefficient frequency matrix; the rows where the coefficients in the matrix have an efficiency coefficient greater than 20% are removed; sampling is performed from the efficiency coefficient sequence after removal to obtain the corresponding photovoltaic power generation sequence, and a time series photovoltaic power generation data set is constructed.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1) Through the analysis of solar radiation data, the present invention proposes a parameter to describe the random variation of solar radiation and a method for generating solar radiation data based on a generative model. By probabilistically modeling the factors causing the random variation of solar radiation, compared with the related models for generating photovoltaic power generation data, the spatial scope of application of this model is expanded from a single site to areas with similar atmospheric conditions and cloud cover levels. It has the advantages of strong model generalization ability and low time and computational cost.

[0038] 2) The present invention enhances the training set of the photovoltaic power generation prediction model through the proposed solar radiation generation method, which can improve the accuracy and robustness of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0040] Figure 2 Schematic diagram of fitting curves of atmospheric transparency coefficient changing with altitude angle under four typical atmospheric transparency levels in an embodiment;

[0041] Figure 3 A scatter plot of the hourly series mean-variance of daily cloud cover types in one embodiment;

[0042] Figure 4 A schematic diagram of a generator structure of a weakening coefficient generation model in an embodiment;

[0043] Figure 5 A schematic diagram of a discriminator structure of a weakening coefficient generation model in an embodiment;

[0044] Figure 6 Generating a statistical similarity evaluation result of the attenuation coefficient and the true attenuation coefficient in one embodiment;

[0045] Figure 7 A schematic diagram of the results of randomly sampling the attenuation coefficient sequence and the cloud amount sequence in an embodiment;

[0046] Figure 8 Schematic diagram of the distribution of daily DNI maximum values ​​under different weather conditions on the summer solstice in one embodiment;

[0047] Fig. 9 A structure diagram of a Transformer-based generative adversarial network prediction model in an embodiment;

[0048] Fig.10 A flowchart for constructing a dataset for a Transformer-based generative adversarial network prediction model in one embodiment;

[0049] Fig.11 A schematic diagram of a test set performance of a generative adversarial network prediction model in one embodiment;

[0050] Fig.12 is a frequency distribution histogram of efficiency coefficient in one embodiment;

[0051] Fig.13 is a schematic diagram of the photovoltaic power generation prediction model structure in an embodiment;

[0052] Fig.14 A flowchart of the formation and application of an enhanced training set in one embodiment of the invention for repeated experiments. DETAILED DESCRIPTION

[0053] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0054] This embodiment first analyzes the solar radiation data reaching the earth's surface.

[0055] The amount of solar radiation reaching the earth's surface is not only related to the relative position of the earth and the sun, but also to the atmosphere above the earth's surface. Due to the existence of the atmosphere, the total solar radiation reaching the earth's surface consists of two parts: one part is the part of the sun directly shining on the ground, which is direct radiation (DNI), and the other part is the part that reaches the ground after being scattered by the atmosphere, which is scattered radiation (DIF). The total radiation reaching the horizontal plane GHI = DNI*sin(h0)+DIF. When the relative position of the earth and the sun is determined, the amount of normal solar radiation reaching the outside of the earth is certain, that is, the extraterrestrial normal solar irradiance (EDNI) can be calculated by formula 1:

[0056]

[0057] The cloud cover in the sky and the presence of various gases and dust in the sky reduce the solar radiation reaching the surface, and at the same time, the randomness of their changes causes the randomness of the changes in the solar radiation reaching the surface. Therefore, the direct solar radiation reaching the surface can be decomposed into a deterministic part and a random part. The deterministic part is determined by the relative position between the earth and the sun, that is, the extraterrestrial normal solar irradiance (EDNI) on a certain day represents the maximum irradiance reaching the earth's surface on that day. The random part also consists of two parts: 1) The randomness of the changes in the components of the atmosphere (aerogel, dust and other gas molecules), which can be described using the atmospheric transparency coefficient that has been widely used in various studies. 2) The random changes in the amount of clouds in the atmosphere, which is described using the attenuation coefficient proposed in this embodiment.

[0058] ① Randomness of direct solar radiation Part 1: Analysis of atmospheric transparency coefficient

[0059] The atmospheric transparency coefficient is a comprehensive reflection of the attenuation coefficient of solar radiation by various factors that are difficult to determine, such as atmospheric thickness and composition. Currently, various studies generally use the Bonguer-Lambert law (as shown in formula (2)). The atmospheric transparency coefficient is obtained by measuring and statistically analyzing radiation data. It is widely used in practice because it avoids analyzing the influence of many meteorological factors and has a certain theoretical basis.

[0060]

[0061] in:

[0062] P m : atmospheric transparency coefficient at a certain place at a certain time;

[0063] DNI clearsky : Direct ground radiation under clear and cloudless conditions, W / m 2 ;

[0064] EDNI: Extraterrestrial normal direct radiation, W / m 2 ;

[0065] m: Atmospheric optical mass number, used to characterize the relative distance that sunlight travels to reach the ground. Its value varies with the solar altitude angle h(°).

[0066] The atmospheric transparency coefficient sequence of a certain location every day and hour {P m0 , P m1 , ..., P m23} represents the change of atmospheric transparency to a certain extent, but this sequence cannot directly represent the atmospheric transparency within a certain spatial range of the location. The main reason is that the atmospheric transparency coefficient calculated by formula (2) is a function of the solar altitude angle, that is, P = f(h). The different distances between different points on the earth and the sun lead to different initial values ​​of the solar altitude angle. At the same time, the solar altitude angle of the location will continue to change with the rotation and revolution of the earth within a year. Therefore, although the atmospheric transparency and its changes within a certain spatial range (Qinghai-Tibet Plateau, Sichuan Basin, Yangtze River Delta, etc.) are similar due to the similarity of landform altitude, it is impossible to use the atmospheric transparency sequence {P calculated by formula (2) at several points within this spatial range. m0 , P m1 , ..., P m23} directly represents the hourly change in atmospheric transparency. The hourly solar altitude angle needs to be corrected to reflect only the change in atmospheric state.

[0067] This embodiment divides the atmospheric transparency under sunny conditions and when the solar altitude angle is h=30° into four levels, and statistically organizes the changes in the atmospheric transparency coefficient of the ground sun hourly with the solar altitude angle under different levels, and uses a power function for fitting. The fitting coefficients are shown in Table 1, and the fitting curves are shown in Table 2:

[0068] Table 1 Fitting curve coefficients of atmospheric transparency coefficients changing with altitude angle under four typical atmospheric transparency levels

[0069]

[0070] Since Table 1 only gives the curves of atmospheric transparency coefficient with altitude angle under four atmospheric transparency levels, and the solar altitude angle and atmospheric transparency coefficient at a certain moment in reality may not be on any of the curves, the curves are further extrapolated. The specific method is as follows:

[0071] 1) Assume that the solar altitude angle at a certain moment is h * , the atmospheric transparency coefficient is

[0072] 2) According to Table 1, h=h under four typical meteorological conditions is calculated. * The atmospheric transparency coefficient at time is expressed as

[0073] 3) According to Table 1, h = 30° under four typical meteorological conditions is calculated. The atmospheric transparency coefficient at time is expressed as

[0074] 4) With Y as the dependent variable and X as the independent variable, the regression equation is obtained by linear fitting using the least squares method: P2 = a*P m +b where a and b are regression equation coefficients;

[0075] 5) The actual atmospheric transparency coefficient Substitute the regression equation obtained in the previous step to obtain the corrected atmospheric transparency coefficient P2.

[0076] The above method is used to correct the altitude angle of the atmospheric transparency sequence of multiple stations within a certain geographical space, turning it into a sequence that only reflects the degree of atmospheric transparency.

[0077] ② Randomness of direct solar radiation Part 2: Cloud cover analysis and its weakening effect on direct radiation

[0078] The corrected atmospheric transparency is no longer affected by cloud cover and sun altitude, and becomes only a parameter that reflects the random changes in atmospheric conditions. Its random changes affect the direct radiation in a clear and cloudless state, namely DNI. clearskyAt this time, the randomness of cloud cover can be described by the hourly sequence of "Cloud type", and the effect of random changes in cloud cover on the direct solar radiation reaching the surface is described by the ratio of the actual radiation reaching the surface to the radiation in a clear and cloudless state. That is, the attenuation coefficient is defined as:

[0079]

[0080] In order to analyze the randomness of solar radiation changes under certain weather conditions, the parameter describing cloud cover, namely "Cloud type", is analyzed and the daily hourly series are classified based on the quantiles of its daily mean and variance, and used as conditional control to generate the results. The daily mean-variance scatter plot of "Cloud type" is shown in Figure 3 As shown in the figure, based on the upper / lower quartiles of the mean and the median of the variance, the daily variation characteristics of cloud cover can be divided into 8 categories.

[0081] Next, the attenuation coefficient sequence reflecting the weakening effect of cloud cover on direct radiation is analyzed. It is found that the change law of the attenuation coefficient can be summarized into three categories. Among them, the first category is the case where the daytime normal solar direct radiation is 0, the second category is the case where the daytime normal solar direct radiation is not weakened by clouds, and the third category is the case where the daytime normal solar direct radiation changes randomly. In different cloud cover mean ranges, the proportions of the three types of attenuation coefficients are different. For the three types of attenuation coefficients, the rules of the first and second categories can be directly expressed, and the third type of attenuation coefficient is expressed by generative model learning. In order to facilitate the application of the model to obtain the corresponding cloud cover change sequence and corresponding attenuation coefficient under the specified weather state. The daily variation characteristics of cloud cover are divided into 8 categories based on the upper, middle and lower quartiles of the cloud cover mean and the median of the variance, and the generative model of cloud cover and attenuation coefficient is trained based on them. Table 2 summarizes the three types of attenuation coefficients and their proportions in different cloud cover ranges. The third and fourth columns of the table are the corresponding relationship between the generative model training and the label information in the generation stage.

[0082] Table 2 Three types of attenuation coefficients and the proportions of various sequences in different cloud ranges

[0083]

[0084]

[0085] After completing the analysis of the attenuation coefficient and cloud cover, it is clear that the modeling object of the generated object is the third type of sequence; at the same time, according to the conditional information proposed above to guide the generation of conditional control of the generation model, the data is labeled and the generation model network architecture is designed and established.

[0086] Based on the above analysis process, this embodiment provides a photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence, such as Figure 1 As shown, the following steps are included:

[0087] S1, obtain solar radiation data.

[0088] S2, using generative adversarial networks to enhance the direct solar radiation data.

[0089] Based on the above analysis results, this embodiment constructs a generative adversarial network, whose generator and discriminator structures are as follows: Figure 4 and Figure 5 The model is used to model the attenuation coefficient, and the modeling effect is evaluated as follows:

[0090] 1) Statistical similarity evaluation

[0091] The principal component analysis of the generated attenuation coefficient and the original attenuation coefficient is performed, and the results are as follows: Figure 6 As shown in the figure, the principal component analysis of the generated attenuation coefficient is basically consistent with that of the original attenuation coefficient, indicating that the trend of random changes in the attenuation coefficient is captured well. The cumulative probability density distribution also has a high degree of overlap, indicating that the probability distribution of the generated data is similar to that of the original data points.

[0092] 2) Data diversity evaluation

[0093] Randomly sample some attenuation coefficient sequences and cloud cover sequences from the generated data, such as Figure 7 As shown. Data diversity is explained from the following aspects: Since the generative adversarial network model is prone to pattern collapse, that is, for the original time series with multiple change trends, the generator can only capture the change trend of one of the time series, resulting in the repeated generation of a certain type of sequence that is very similar to the original data in the generation stage. This phenomenon is also reflected in the principal component analysis (PCA, t-SNE) results and the cumulative probability density results. The principal component analysis results and probability density results in the similarity evaluation show that the frequency of occurrence of each value of the generated attenuation coefficient within its interval is similar to the original data. This can explain to a certain extent that there is no pattern collapse phenomenon.

[0094] Refer to the research on other time series generation models for the evaluation method of the diversity of generated data: that is, visualize the generated results.

[0095] In the model generation stage, this embodiment generates the attenuation coefficient of the same number of samples in the original data set at one time, randomly generates a random integer of [1, number of samples], and draws the generation curve corresponding to the index to evaluate whether the mode collapse phenomenon occurs. Figure 7It is shown in the figure that for the original weakening coefficient with multiple changing trends (different mutation values ​​and different mutation times), the generation model can better capture its changing trends without repeatedly generating a certain type of weakening coefficient sequence.

[0096] 3) Evaluation of the validity of conditional information

[0097] By statistically analyzing the original data, the conditional information that controls the generated model data and the combination ratio of the attenuation coefficients under different weather conditions were determined. However, whether the conditional information and the ratio can effectively guide the output of the generated model needs further verification. Therefore, the generative model was used to generate 200 sets of attenuation coefficients under three types on the day of the summer solstice, that is, when the date index is 172, and the direct radiation value was further calculated. A box plot of the distribution of the maximum direct radiation day is drawn, as shown in the figure below: Figure 8 As shown, it can be seen that the maximum direct radiation values ​​of the three types are different.

[0098] S3, based on the enhanced direct radiation data, uses the black box model to predict the scattered radiation and total radiation to construct an enhanced solar radiation dataset.

[0099] After obtaining the normal direct radiation under the random change of cloud cover, the direct radiation DHI of the horizontal plane can be calculated according to DHI=DNI*sin(h), (h is the solar altitude angle). Furthermore, the total radiation GHI of the horizontal plane can be calculated according to DHI+DIF. It can be seen that GHI, DIF, and DNI are not independent of each other, but are closely related. If DIF and GHI are modeled independently, only their respective time dependencies can be captured. Therefore, the generation of GHI and DIF data should be based on the generated DNI data. The white box model that describes the correlation between direct radiation, scattered radiation, and total radiation mainly revolves around establishing a series of direct-scattered classification models based on total radiation data to obtain direct data and scattered data to solve the problem of missing direct radiation and scattered radiation in actual meteorological observations. Therefore, this embodiment uses a black box model to capture the mutual influence and effect between the above radiation factors, more specifically, the value of GHI or DIF under a certain DNI change.

[0100] When the Pearson correlation coefficient and Spearman correlation coefficient were analyzed for meteorological factors, it was found that the Pearson correlation coefficients between the same factors obtained by selecting different time spans were very different. After analysis, the analysis range of the Pearson correlation coefficient and the Spearman correlation coefficient is mainly linear data and simple monotonous nonlinear data. However, the radiation data sequence has periodic changes and random fluctuations, and the use of the Pearson correlation coefficient and the Spearman correlation coefficient cannot accurately reflect the correlation. After investigation, the maximum information coefficient was used to analyze the correlation between radiation data.

[0101] The maximum information coefficient (MIC) is a measure used to measure the degree of association, linearity or nonlinearity between two variables X and Y. It can be regarded as the amount of information contained in a random variable about another random variable, or the uncertainty of a random variable reduced by knowing another random variable. As an improvement of mutual information, the maximum information coefficient is often used in feature selection for machine learning. Its calculation formula is shown in formula (3).

[0102]

[0103] Before establishing the data-driven model, the maximum mutual information analysis was performed on the hourly radiation data with a daily period. The calculation results are shown in Table 3.

[0104] Table 3 MIC analysis of radiation factors

[0105] MIC[x;y] DIF DNI GHI DNI 0.72650344 0.99994257 0.76193006

[0106] Since the maximum information coefficients of DNI and DIF, GHI are relatively similar, the prediction models of NDI→DIF and DNI→GHI are trained respectively, and the one with better performance on the test set is selected as the generation model of DNI data.

[0107] This example improves the GAN model structure used in the paper "TTS-CGAN: A Transformer Time-Series Conditional GAN ​​for Biosignal Data Augmentation" to obtain a transformer-based generative adversarial network prediction model. The model architecture is as follows: Fig. 9As shown, the encoder module that constitutes the main body of its generator and discriminator is composed of multiple self-attention layers, feedforward layers and random inactivation layers (dropout) layers. Since the input of the discriminator is two time series when generating sequences, the encoder module using transformer helps to capture the autocorrelation of time series. This embodiment intends to use a transformer-based generative adversarial network for time series regression prediction, and the output of the generator is a single value at the prediction time, not a time series. At the same time, according to the model training experience, the setting of the discriminator is simplified to a fully connected layer to increase the training speed.

[0108] Based on the transformer-based generative adversarial network architecture proposed above, the DNI data in the original dataset is used for training. The model-related hyperparameter information is shown in Table 4.

[0109] Table 4 Related information of Transformer-based generative adversarial network model

[0110]

[0111]

[0112] The purpose of establishing the DNI→GHI and DNI→DIF prediction models is to use neural networks to capture the relationship between DNI changes and DIF / GHI over a period of time. Therefore, the dataset construction process is as follows: Fig.10 shown.

[0113] Due to the huge amount of data and the complexity of the network architecture, the training time of a single prediction model is more than 2 days. The model is saved every 10 epochs and tested on the test set. After running 110 epochs, the performance of the test set did not change significantly and the training was stopped.

[0114] The model prediction error on the test set is shown in Table 5:

[0115] Table 5. DNI→DIF, DNI→DIF prediction model test set error

[0116] Evaluation indicators DNI → GHI DNI→DIF MAE 0.136328 0.107220 MSE 0.034277 0.019086 RMSE 0.185139 0.138152

[0117] according to Fig.11 The addition of the Transformer mechanism significantly improves the time series prediction accuracy based on the generative adversarial network model. After comparison, the DNI→DIF prediction model performs better, so this prediction model is used to generate DNI data.

[0118] S4, calculates photovoltaic power generation based on the enhanced solar radiation dataset and constructs a time-series photovoltaic power generation dataset.

[0119] After the training of the generative model was completed and the statistical characteristics of the generated data and the original data were evaluated by the above method, the generative model was applied to generate 10 sets of attenuation coefficients for each of the three types for each day of the year, and further calculated the direct radiation value and the corresponding scattered radiation value. Based on the developed radiation-photovoltaic power generation calculation program, a total of 24 daily and hourly photovoltaic power generation scenarios were obtained.

[0120] The average of the 24 data points at each time point of the daily photovoltaic power generation sequence is taken and the date corresponding to the date in the original data set is intercepted. The generated data is spliced ​​with the original true photovoltaic power generation sequence, and the data points of photovoltaic non-power generation at night are eliminated to form an enhanced training data set for the photovoltaic power generation prediction model.

[0121] When using the radiation data output by the generation model to calculate photovoltaic power generation and generate photovoltaic power generation scenes, and further enhance photovoltaic power generation data, it is necessary to know the relevant parameter information of the photovoltaic power generation system. When the target data to be enhanced and the corresponding solar radiation data are known, in order to avoid the error of scene generation caused by the photovoltaic calculation model, the photovoltaic power generation calculation program needs to be calibrated to ensure that when the real solar radiation is used as the input of the calculation program, the difference between the output solar photovoltaic power generation and the real power generation is within an acceptable range.

[0122] The total radiation that can be obtained on the slope per unit area is calculated by the photovoltaic inclination angle, and then multiplied by the total area of ​​the photovoltaic panel and the total efficiency coefficient of the system to obtain the calculated photovoltaic power generation. The actual photovoltaic power generation is divided by the calculated photovoltaic power generation to obtain the total efficiency coefficient at that moment, and the calculated efficiency coefficient frequency is statistically analyzed. Its frequency histogram is as follows: Fig.12 As shown. Since the maximum efficiency of photovoltaic power generation cells does not exceed 20%, and the total efficiency coefficient is superimposed on the system efficiency, it should not be greater than 20%. Therefore, the rows where the efficiency coefficient is greater than 20% in the matrix are removed. When the generated radiation data is used to calculate the corresponding photovoltaic power generation scene, sampling is performed from the efficiency coefficient sequence after removal to obtain the corresponding photovoltaic power generation sequence.

[0123] S5, using the time series photovoltaic power generation data set to train the photovoltaic power generation prediction model, and realizing the photovoltaic system power generation prediction based on the trained photovoltaic power generation prediction model.

[0124] For the mixed data and the original data, a data-driven model is trained to perform the autoregressive prediction task of the photovoltaic power generation sequence: that is, {PV (t=1) ,PV (t=2) ,PV (t=3) ,PV (t=4)} as model input, PV (t=5) As the model output. The two data-driven models have the same model architecture and related hyperparameters except for the different training data sets. Fig.13 The hyper parameters are shown in Table 6. After training, the models composed of two different training sets are tested separately with the test set composed of the same real data.

[0125] Table 6 Hyperparameters of photovoltaic power generation prediction model

[0126] parameter value Batch_size 32 Conv1d number of hidden units 128 Number of hidden units in the attention module 128 Number of training rounds 2500 Learning Rate 0.001

[0127] The loss function in the training process and the evaluation index of the model test are both root mean square error. In order to avoid contingency, the above process of forming an enhanced training set for the prediction model and using it for prediction model training and using the real data test set for evaluation was repeated 10 times. The flowchart is shown in Fig.14 The prediction results are shown in Table 7.

[0128] Table 7 Comparison of test errors between enhanced training set and original training set

[0129]

[0130] For scenarios in which photovoltaic power generation data has not yet been obtained in the photovoltaic system planning and design stage, this method can generate a series of photovoltaic power generation scenarios to guide the planning of power system energy types and the formulation of related scheduling strategies.

[0131] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence, characterized in that: The following steps are involved: S1, obtaining solar radiation data, wherein the solar radiation data includes solar direct radiation, scattered radiation and total radiation data; S2, using generative adversarial networks to enhance the direct solar radiation data; S3, based on the enhanced direct radiation data, the black box model is used to predict the diffuse radiation and total radiation to construct an enhanced solar radiation dataset; S4, calculates photovoltaic power generation based on the enhanced solar radiation dataset and constructs a time series photovoltaic power generation dataset; S5, using the time series photovoltaic power generation data set to train the photovoltaic power generation prediction model, and realizing the photovoltaic system power generation prediction based on the trained photovoltaic power generation prediction model.

2. A photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 1, characterized in that: The relationship between the direct solar radiation, diffuse radiation and total radiation is: GHI=DNI*sin(h0)+DIF Among them, DNI is the normal solar direct radiation, DIF is the solar diffuse radiation, h0 is the solar altitude angle, DNI*sin(h) is the direct radiation GHI on the horizontal plane, and GHI is the total radiation on the horizontal plane.

3. The photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 1 is characterized in that: The direct solar radiation includes a deterministic part and a random part, wherein the deterministic part is determined by the relative position between the earth and the sun, and the random part includes the randomness of the change of atmospheric composition described by the atmospheric transparency coefficient and the randomness of the change of cloud cover in the atmosphere described by the attenuation coefficient.

4. A photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 3, characterized in that: Determination method of the atmospheric transparency coefficient The following steps are involved: A1, calculate the initial value of the atmospheric transparency coefficient according to the following formula: Among them, P m is the initial value of the atmospheric transparency coefficient at a certain place at a certain time, DNI clearsky It is the direct ground radiation under clear and cloudless conditions, in W / m 2 ; EDNI is the extraterrestrial normal direct radiation, in W / m 2 ; m is the atmospheric optical mass number, which is used to characterize the relative distance that sunlight travels to reach the ground, and its value varies with the solar altitude angle h; A2, the hourly solar altitude angle is corrected to reflect only the change of atmospheric state: the atmospheric transparency under clear sky condition and solar altitude angle h=30° is divided into several levels, and the change of the initial value of the atmospheric transparency coefficient of the ground sun hourly with the solar altitude angle under different levels is statistically sorted, and the power function is used for fitting to obtain the fitting curve; A3, further extrapolate the fitting curve and correct the atmospheric transparency coefficient. The specific method is as follows: 1) Assume that the solar altitude angle at a certain moment is h * , the initial value of the atmospheric transparency coefficient is 2) Calculate h=h under different typical meteorological conditions based on the fitting curve * The atmospheric transparency coefficient at time is expressed as Among them, n is the number of typical meteorological phenomena; 3) According to the fitting curve, the atmospheric transparency coefficient at h = 30° under different typical meteorological conditions is calculated and recorded as 4) With Y as the dependent variable and X as the independent variable, the regression equation is obtained by linear fitting using the least squares method: P2 = a*P m +b, where a and b are regression equation coefficients; 5) The actual atmospheric transparency coefficient Substitute the regression equation obtained in the previous step to obtain the corrected atmospheric transparency coefficient P2.

5. The photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 3 is characterized in that: The attenuation factor is defined as the ratio of the actual radiation reaching the surface to the radiation in a clear, cloudless state: Among them, DNI clearsky It is the direct ground radiation under cloudless conditions, and DNI is the direct solar radiation from the normal direction.

6. A photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 5, characterized in that: The attenuation coefficient is divided into three categories according to the weakening effect of cloud cover on direct radiation. The first category is the case where the daytime normal solar direct radiation is 0, the second category is the case where the daytime normal solar direct radiation is not weakened by clouds, and the third category is the case where the daytime normal solar direct radiation changes randomly. The first and second categories are directly represented by definition, and the third category is represented by learning using a generative adversarial network.

7. A photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 6, characterized in that: Based on the upper / lower quartiles of the mean and the median of the variance of the cloud description parameters, the diurnal variation characteristics of cloud cover are divided into 8 categories. The generative adversarial network of cloud cover and attenuation coefficient is trained based on the categories of the diurnal variation characteristics of cloud cover to generate the diurnal variation feature sequence and attenuation coefficient sequence of cloud cover.

8. The photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 1, characterized in that: The method of predicting scattered radiation and total radiation using a black box model based on the direct radiation data after data enhancement is specifically as follows: the maximum information coefficient between direct radiation and scattered radiation, and between direct radiation and total radiation are calculated respectively, and according to the comparison results of the maximum information coefficients, a black box model for predicting scattered radiation based on direct radiation and a black box model for predicting total radiation based on direct radiation are trained respectively.

9. The photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 1, characterized in that: The black box model is a generative adversarial network prediction model, and the generator of the generative adversarial network prediction model adopts a transformer structure.

10. The photovoltaic system power generation prediction method based on solar radiation data augmentation and emergence according to claim 1, characterized in that: The step S4 is specifically as follows: according to the enhanced solar radiation data set, the total radiation on the slope per unit area is calculated by the photovoltaic inclination angle, and then multiplied by the total area of ​​the photovoltaic panels and the total efficiency coefficient of the system to obtain the photovoltaic power generation calculation amount, the actual photovoltaic power generation is divided by the calculated photovoltaic power generation to obtain the total efficiency coefficient at that moment, the calculated efficiency coefficient frequency is counted to obtain the efficiency coefficient frequency matrix; the rows where the coefficients in the matrix have an efficiency coefficient greater than 20% are removed; sampling is performed from the efficiency coefficient sequence after removal to obtain the corresponding photovoltaic power generation sequence, and a time series photovoltaic power generation data set is constructed.

Citation Information

Cited By

  • Wide-area distributed photovoltaic power generation efficiency evaluation method

    CN119886582A

  • Regional horizontal irradiance inversion method and system based on satellite data and deep learning

    CN120387486A