Ultra-short-term photovoltaic output prediction method based on adaptive feed-forward generative network

Abstract

The invention discloses an ultra-short-term photovoltaic output prediction method based on an adaptive feed-forward generative network, and the method comprises the steps: recognizing the internal parameters of a photovoltaic module through an improved raccoon optimization algorithm according to the collected electrical and meteorological data, and predicting the output through a feed-forward generative adversarial network ultra-short-term photovoltaic output prediction model according to the internal parameters and meteorological factors. According to the method, hybrid optimization identification is carried out by adopting a Raccoon coaccoon optimization algorithm (COA) in combination with chaotic initialization, a token ring and a sequence least square programming (SLSQP) technology so as to predict internal parameters of a component, and iterative training is carried out by combining a generative adversarial network with a long short-term memory network. And the prediction precision of the photovoltaic power data with the scale of 3-4 hours is well improved.

Description

Technical Field

[0001] The invention relates to a photovoltaic output prediction method, and in particular to an ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network. Background Art

[0002] The existing ultra-short-term prediction methods for photovoltaic output can be divided into physical methods and statistical methods according to different prediction models. These methods can provide ultra-short-term predictions of photovoltaic power generation to a certain extent, but both physical methods based on the principle of photovoltaic power generation and statistical methods based on historical data are easily affected by external environmental factors and internal parameters of photovoltaic modules, resulting in inaccurate prediction results. This is mainly reflected in the processing of meteorological data and the influence of internal parameters of photovoltaic modules on output power. Weather changes, climate conditions and other external factors can significantly affect the performance of photovoltaic modules, and these changes are directly reflected in the output power. In addition, the internal parameters of photovoltaic modules, such as photocurrent, reverse saturation current, equivalent series resistance, equivalent parallel resistance and ideal factor, also have a key impact on their power generation capacity. In summary, the interaction between meteorological data and internal parameters will lead to deviations in prediction results, thereby reducing the accuracy of ultra-short-term predictions.

[0003] Through literature search of the prior art, it is found that the direction of this type of patents is mainly focused on prediction based on the photoelectric conversion equation and photovoltaic output prediction using the mapping rules in historical data. In the Chinese patent "Distributed photovoltaic ultra-short-term power prediction method based on photovoltaic clear sky power decomposition and dynamic and static correlation joint modeling", Sun Yiwei et al. proposed a photovoltaic ultra-short-term power prediction method based on photovoltaic clear sky power decomposition and dynamic and static correlation joint modeling. The method integrates historical output data, constructs a clear sky power model based on a preset low-rank matrix, and constructs a graph convolution prediction model based on an intelligent driving method. However, the invention fails to fully consider the nonlinear response characteristics of photovoltaic modules under different seasonal weather conditions, resulting in insufficient prediction accuracy under complex climate conditions. In addition, in the Chinese patent "EMD-BiGRU ultra-short-term photovoltaic power prediction method based on EMD decomposition", Wang Xiao et al. established BiGRU models for each component after the original power decomposition, and obtained the final power generation prediction value by accumulating each prediction result, which has good prediction accuracy for complex time series. However, this method lacks dynamic update of real-time meteorological data during model training, which limits its applicability and accuracy in rapidly changing environments.

[0004] The traditional detection method of photovoltaic output prediction mainly uses sensors to monitor signals. On the one hand, the traditional neural network photovoltaic power prediction model is restricted by the source data of the power plant, ignores the impact of some environmental factors on photovoltaic power, and lacks effective use of multivariate environmental sequence information; on the other hand, due to the nonlinear changes between photovoltaic power and multivariate environmental sequence information, as the network input variables increase, the model convergence speed will slow down and overfitting problems will occur; at the same time, due to the lack of consideration of the characteristics of photovoltaic power changing over time, the improvement of prediction accuracy is limited.

[0005] In summary, the existing ultra-short-term prediction methods for photovoltaic output have some problems, such as being greatly influenced by external environmental factors, not considering the internal parameters of components, and insufficient model adaptability. This provides room for development to further improve the accuracy and reliability of photovoltaic power generation prediction. Summary of the invention

[0006] Purpose of the invention: The purpose of the invention is to provide an ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network.

[0007] Technical solution: The present invention provides an ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network. The collected electrical and meteorological data are used to identify the internal parameters of the photovoltaic components through an improved raccoon optimization algorithm. The internal parameters are combined with meteorological factors to predict the output through a feedforward generation adversarial network ultra-short-term photovoltaic output prediction model.

[0008] Combining the internal parameters and external environmental factors of the photovoltaic system, this invention applies deep learning to photovoltaic output prediction, aiming to further improve the accuracy and efficiency of the prediction. This method is expected to consider the impact of internal and external factors of photovoltaic components and historical output on photovoltaic power generation, use generative adversarial networks to provide diversified samples, avoid overfitting due to too single data, combine LSTM and attention mechanism to capture temporal dependencies, focus on key information, and provide more reliable technical support for the operation and maintenance of photovoltaic systems, thereby optimizing energy scheduling and management, and improving the economy and sustainability of photovoltaic power generation.

[0009] Further, the method of identifying the internal parameters is as follows: setting the internal parameters of the photocurrent Iph, the diode saturation current I o , equivalent series resistance R s , equivalent parallel resistance R sh, the initial value of the ideal factor n is 0, 0, 0, 100, 1; the interval ranges are 0-10A, 0-1A, 0-1Ω, 100-200Ω, 1-3, respectively. For the initial data, it does not fluctuate much in a small range. The improved raccoon optimization algorithm is applied to the data in groups of four for internal parameter identification. The maximum number of iterations is set to 100, the number of populations is 30, and it is vectorized and chaotically initialized to generate individuals throughout the search space. Then, the hunting and attack of the coati on the iguana is simulated through exploration; the process of escaping from the predator is simulated through development. In view of the lack of population information guidance in the development process, only searching around the current individual leads to difficulty. With the problem of convergence, the idea of ​​token ring update is introduced to improve the formula by introducing global information. Formula (3) is used to select the next solution to be updated according to the current token position. At the end of each iteration, the token position is adjusted by formula (4) so ​​that the search can be started from a different starting point in the next iteration. At the same time, the applicability of the entire problem space is limited by adding global constraints. At the end of each development, the SLSQP algorithm is used for precise convergence to obtain a higher quality solution. The ablation experiment confirms the indispensability of chaotic initialization, token ring improvement and sequential least squares programming optimization in the algorithm; at the same time, comparative experiments are introduced to confirm the superiority of this improved algorithm.

[0010] Furthermore, the method of chaos initialization: Equations (1)-(2) are used to make a chaotic sequence and map each component separately,

[0011] Z ij =μZ ij (1-Z ij )(j=1,2,...,d) (1)

[0012] X ij =b lj +(b uj -b lj )Z ij (i=1, 2,...,M; j=1, 2,...,d) (2)

[0013] Among them, μ represents the control parameter, Z ij is a d-dimensional vector Z i The weight in b lj and b uj Represents the lower and upper bounds of the mapping interval.

[0014] Further, the token ring update method: Formula (3) is used to update the current token index, and at the end of each iteration, the token position is adjusted using formula (4) to ensure that the search starts from a different point in the next iteration,

[0015]

[0016] token=(token+1)%SearchAgents_no (4)

[0017] Among them, token is the token ring, starting from 0, i is the individual index in the population, starting from 0, and SearchAgents_no represents the number of populations.

[0018] Furthermore, the method of generating adversarial networks is as follows: the input cascade data is first preliminarily encoded through a fully connected layer and a ReLU activation function; then the local subsequence information is focused on through a self-attention layer, the data dimension is adjusted through a linear layer, and then different contextual information in the input data is captured through a two-head attention mechanism and mapped to a potential vector space, and layer normalization and residual connection are introduced after each type of attention mechanism to alleviate the gradient disappearance in the deep network; two LSTM hidden layers and a fully connected layer are set in the decoder to convert it into a target sequence, and for the discriminator, a one-dimensional convolutional neural network is designed for judgment. In addition, the model also uses a pre-trained SVM to construct a hinge loss, and combines the discriminator loss and the generator loss to generate a more accurate PV power sequence.

[0019] The 188-dimensional real data composed of the measured power and the support data selected by grayscale correlation analysis (GRA) is denoted as x, the data concatenated by the noise and support data is denoted as z, and the fake data simulated by the generator is recorded as G(z). Then, the prediction results are adaptively optimized through the three joint losses listed in equations (5)-(7). The excavator loss is used to reconstruct and optimize the generator, and x and G(z) are sent to the discriminator to calculate the discriminator loss. The discriminator loss is combined with the trained SVM to calculate the hinge loss.

[0020]

[0021] l h =max(0,1-y·f(x)) (7)

[0022] Where z is the data of noise and support data spliced ​​in accordance with pz(z) distribution, x is the real data of measured power and support data spliced ​​in accordance with Pr distribution, G(·) is the generator, D(·) is the discriminator, E represents the expectation, y is the label, indicating whether it is real data, and its value is {-1, 1}, f(x) is the prediction output of the support vector machine model, l G is the generator loss, l D is the discriminator loss, l h is the hinge loss.

[0023] Furthermore, for the generator, the network is mainly composed of a self-attention layer, a multi-head attention layer, an LSTM layer and a fully connected layer. The cascaded data z is first transferred to a new feature space through a 188*128 weight matrix through a fully connected layer for preliminary feature extraction and dimensionality reduction; then, nonlinearity is introduced through the ReLU activation function to enable the model to learn more complex features. The selected ReLU activation function can effectively alleviate the gradient vanishing problem, and then the self-attention layer is used to generate a query matrix, a key matrix and a value matrix with a shape of 128*128 for the input 128-dimensional features, and important information is retained by focusing on local subsequences while reducing computational complexity; the linear layer is used for feature conversion to adjust the output dimension to 64; with the help of a 64*32 head per head The dual-headed attention mechanism performs parallel calculations to capture different contextual information in the input data to enhance the expressiveness of the model and enable it to focus on richer features. At the same time, layer normalization and residual connections are added after each attention layer. Layer normalization improves the stability and speed of training and alleviates internal covariate shift. The residual connection adds the input directly to the output of a certain layer, which helps to alleviate the gradient vanishing problem in deep networks and makes information more easily propagated. That is, for the dual-headed attention layer, the output is the sum of the input and output values, which is still 64 dimensions. The features extracted by the encoder are input into the decoder composed of LSTM layers. The hidden layer is set to two layers with dimensions of 64 and 128 respectively. Then the output of the LSTM layer is converted to the target dimension of 188 through the fully connected layer.

[0024] Furthermore, for the discriminator, it is set as a one-dimensional convolutional neural network with an input layer size of 188. Three one-dimensional convolutional layers are added to extract local features in the time series and capture the relationship between asynchronous ones. The maximum pooling is introduced to reduce the feature dimension. A fully connected layer is set to extract high-level features. Finally, the tanh activation function is used to set the output range between [-1, 1], which is regarded as the confidence score for the real samples and the generated samples. At the same time, a SVM model with a linear kernel is constructed for training and prediction to achieve joint loss.

[0025] Detailed description:

[0026] 1) In a photovoltaic power station, the output current and output voltage of the photovoltaic array are mainly measured. The photovoltaic array is composed of photovoltaic modules connected in series and parallel, and each photovoltaic module is composed of photovoltaic cells connected in series and parallel. Photovoltaic cells are essentially PN junctions that use the photoelectric effect of semiconductors to convert light energy into electrical energy. Their mathematical model is also established based on the equivalent circuit of photovoltaic cells. The principle of obtaining internal parameters is as follows: Figure 2As shown. Considering the complexity of the algorithm, in this invention, the photovoltaic array is equivalent to a large photovoltaic cell and an improved intelligent optimization algorithm COA is used to identify the corresponding internal parameters. In order to solve the problems of slow convergence speed, insufficient global exploration ability, and easy to fall into local optimum in the identification of internal parameters of photovoltaic modules, an improvement scheme such as chaotic initialization individual, token ring update and sequential least squares planning is proposed. The overall process of the improved COA is as follows: Figure 3 shown.

[0027] Chaos Initialization

[0028] Considering the randomness and extensiveness of coatis distribution, chaos technology is introduced to generate the initial population in the search space to increase the probability of finding the optimal solution. Equations (1)-(2) are used to make chaotic sequences and map each component separately.

[0029] Z ij =μZ ij (1-Z ij )(j=1,2,...,d) (1)

[0030] X ij =b lj +(b uj -b lj )Z ij (i=1, 2,...,M; j=1, 2,...,d) (2)

[0031] Where μ represents the control parameter, Z ij is a d-dimensional vector Z i The amount in. lj and b uj Represents the lower and upper bounds of the mapping interval.

[0032] Token Ring Update

[0033] During the development process, COA mainly focuses on the immediate environment of the current individual and does not consider the group-wide information, which is not conducive to achieving convergence in the later stage. To solve this problem, token ring transmission is introduced to update the global optimal information. Formula (3) is used to update the current token index. At the end of each iteration, the token position is adjusted using formula (4) to ensure that the search starts from a different point in the next iteration.

[0034]

[0035] token=(token+1)%SearchAgents_no(4)

[0036] Among them, token is the token ring, starting from 0, i is the individual index in the population, starting from 0, and SearchAgents_no represents the number of populations.

[0037] Sequential Least Squares Programming (SLSQP)

[0038] SLSQP is a sequential quadratic programming algorithm that breaks down a large problem into small problems and solves the small problems through linear constraints. After each iterative update, the convergence of all individual parameters is used to achieve a balance between local optimization and global search.

[0039] 2) Generative adversarial network is a deep learning model that produces better output results through mutual game learning between generator and discriminator. To address the challenge that traditional GAN ​​has difficulty capturing time-varying dynamics in sequences, an improved autoregressive feedforward generative adversarial network is proposed. This new method combines the generation ability of GAN with the prediction ability of LSTM. In addition, it also introduces self-attention mechanism and multi-head attention mechanism to further improve the feature extraction and expression ability of the network while efficiently processing long sequence information. The model also uses pre-trained SVM to construct hinge loss, combining discriminator loss and generator loss to generate more accurate PV power sequences. The structure of the model is as follows Figure 4 shown.

[0040] The present invention represents the 188-dimensional real data composed of the measured power and the support data selected by the grayscale correlation analysis (GRA) as x, the data spliced ​​by the noise and the support data as z, and the fake data simulated by the generator as G(z). Then, the prediction results are adaptively optimized by the three joint losses listed in equations (5)-(7). The generator is reconstructed and optimized using the excavator loss, and x and G(z) are sent to the discriminator to calculate the discriminator loss, which is combined with the trained SVM to calculate the hinge loss.

[0041]

[0042] l h =max(0,1-y·f(x)) (7)

[0043] Where z is the data of noise and support data spliced ​​in accordance with pz(z) distribution, x is the real data of measured power and support data spliced ​​in accordance with Pr distribution, G(·) is the generator, D(·) is the discriminator, E represents the expectation, y is the label, indicating whether it is real data, and its value is {-1, 1}, f(x) is the prediction output of the support vector machine model, l G is the generator loss, l D is the discriminator loss, l h is the hinge loss.

[0044] The network of the generator consists of a self-attention layer, a multi-head attention layer, an LSTM layer, and a fully connected layer. The real data x is first transferred to a new feature space through a 188*128 weight matrix through a fully connected layer for preliminary feature extraction and dimensionality reduction; then the nonlinearity is introduced through the ReLU activation function to enable the model to learn more complex features. The selected ReLU activation function can effectively alleviate the gradient vanishing problem. Then, the self-attention layer generates a query matrix, a key matrix, and a value matrix with a shape of 128*128 for the input 128-dimensional features, retaining important information while reducing computational complexity by focusing on local subsequences; the linear layer is used for feature conversion to adjust the output dimension to 64; with the help of a dual-head attention mechanism with each head of 64*32, parallel calculation is performed to capture different contextual information in the input data to enhance the model's expressiveness and enable it to focus on richer features. At the same time, layer normalization and residual connection are added after each attention layer. Layer normalization improves the stability and speed of training and alleviates internal covariate shift; residual connection adds the input directly to the output of a certain layer, which helps to alleviate the gradient vanishing problem in deep networks and makes information propagation easier. That is, for the two-headed attention layer, the output is the sum of the input and output values, which is still 64 dimensions. The features extracted by the encoder are input into the decoder composed of LSTM layers. The hidden layer is set to two layers, and the dimensions are set to 64 and 128 respectively. Then the output of the LSTM layer is converted to the target dimension of 188 through the fully connected layer.

[0045] For the discriminator, we set the dimension of the hidden layer to 64, the activation functions of the first two layers to the Relu function, and the last layer to the tanh function, so that the output range is placed between [-1, 1] as the confidence score for the real sample and the generated sample. At the same time, a SVM model with a linear kernel is constructed for training and prediction to achieve joint loss.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1) A photovoltaic power prediction method combining the internal parameters of photovoltaic modules with external environmental factors is proposed, which greatly improves the prediction accuracy of photovoltaic power data on a 3-4h scale.

[0048] 2) An improved coati optimization algorithm (COA) is used in combination with chaos initialization, token ring and sequential least squares programming (SLSQP) techniques to perform hybrid optimization identification to predict component internal parameters.

[0049] 3) Construct a generator network with an attention mechanism module as the encoder and a long short-term memory network (LSTM) module as the decoder, combine it with the discriminator network trained with the improved raccoon optimization algorithm for recursive multi-step training, and use support vector machine (SVM) for pre-training to calculate the hinge loss, and combine the generator loss and discriminator loss to generate an accurate photovoltaic output sequence. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Basic flow chart of photovoltaic output prediction;

[0051] Figure 2 It is the internal reference schematic diagram;

[0052] Figure 3 To improve the Raccoon optimization algorithm, identify the internal reference flow chart;

[0053] Figure 4 To improve the prediction output structure diagram of the autoregressive feedforward generative adversarial network;

[0054] Figure 5 These are the prediction results of mainstream algorithms at different time steps. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be further described below.

[0056] The overall algorithm flow chart of this embodiment is as follows Figure 1 shown.

[0057] 1) Step 1, cleaning the data set. The original data set is selected from the Connellan Airport in the Yulara Solar System of the Australian Solar Center of Desert Knowledge. Its photovoltaic array has a rated power of 105.9KW, an inclination of 10 degrees, and an orientation of 40 degrees north-east; the motor panel type is SubPower SPR-327NE, the number of panels is 326, and the inverter model is SMA STP25000TL-30 & 20000TL-30. The data was selected from March 2023 to February 2024, with a sampling interval of 5 minutes. In order to deal with the problem of missing readings such as irradiance and temperature caused by equipment failure, the present invention fills the data monthly to create a complete data set spanning 366 days. Considering that the power output at night and dawn is almost zero, it is of little significance to the study, so only the power data from 07:00 to 19:00 is considered. In order to alleviate the impact of seasonal changes, the data is segmented according to the season, the window size is set to 20, and the sliding window idea is used for subsequent data processing.

[0058] 2) Step 2: Identify internal references. Figure 3 Identify the component internal parameters and set internal parameters I ph , Io , R s , R sh , the initial value of n is 0, 0, 0, 100, 1; the interval ranges are 0-10A, 0-1A, 0-1Ω, 100-200Ω, 1-3. For the initial data, we believe that it does not fluctuate much in a small range. We apply the improved raccoon optimization algorithm to the data in groups of four for internal parameter identification, set the maximum number of iterations to 100, and the number of populations to 30. Vectorization and chaos initialization are used to generate individuals throughout the search space, and then the hunting and attack of coatis on iguanas are simulated through exploration; the process of escaping from predators is simulated through development. In view of the lack of population information guidance in the development process, the search is only carried out around the current individual, which leads to the problem of difficulty in convergence. The idea of ​​token ring update is introduced to introduce global information to improve the formula. Formula (3) is used to select the next solution to be updated according to the current token position. At the end of each iteration, the token position is adjusted through formula (4) so ​​that the search can start from a different starting point in the next iteration. At the same time, the scope of application of the entire problem space is limited by the added global constraints. At the end of each development, the SLSQP algorithm is used to achieve accurate convergence and obtain a higher quality solution. The ablation experiment confirms the indispensability of chaos initialization, token ring improvement and sequential least squares programming optimization in the algorithm, as shown in Table 1; at the same time, a comparative experiment is introduced to confirm the superiority of this improved algorithm, and the results are shown in Table 2.

[0059] 3) Step 3: Output prediction. After using the improved COA to identify the internal parameters, the grey relational analysis (GRA) is used to select the factors that have a greater impact on the power output through formula (8):

[0060]

[0061] in and Represent the jth power output of the photovoltaic power station and the jth data of the first indicator respectively, ρ represents a constant (the present invention takes 0.5), comprehensively considers the joint influence of internal and external parameters, introduces three-phase average current, light radiation instrument parameters, solar irradiance, two temperature parameters, photocurrent Iph and ideal factor n, a total of 7 attributes as strong correlation variables. In order to predict the output power in the next four hours, the real output is combined with the supporting data composed of the above-mentioned strong correlation variables to construct real data with a dimension of 20*7+12*4, and the noise is combined with the relevant supporting data to form false data, which are respectively input into the generator and discriminator for training.

[0062] The fake data first passes through a fully connected layer for preliminary feature extraction to convert the dimension to 128, and then passes through a self-attention layer, a linear layer, and a two-headed attention mechanism layer to focus on key areas and capture contextual information. Layer normalization and residual connection are added after each layer, and the dimension conversions are 128, 64, and 64 respectively.

[0063] The extracted features are input into the decoder composed of two layers of LSTM, and then output through the fully connected layer with a dimension of 188. The real data and false data are then sent to the discriminator with a dimension set to 64 for judgment, and their confidence scores are output. During the training process, the number of training times of the generator and the discriminator are adjusted as needed to avoid mode collapse caused by overfitting of a certain component. At the same time, a linear kernel SVM is constructed for auxiliary training to enhance the stability of the generator. This improved feedforward generative adversarial network is used to predict and analyze the photovoltaic power output in different seasons within the next 4 hours. In addition, an ablation experiment is set up to prove the indispensability of the improved raccoon optimization algorithm for identifying internal parameters in the overall architecture of the design. The results are shown in Table 3; a comparative experiment is set up to prove the superiority of this algorithm compared with LSTM and the current mainstream advanced algorithms. At the same time, the prediction step is subdivided to measure its prediction accuracy in the ultra-short term. The results are shown in Tables 4 and Figure 5 .

[0064] The experimental environment of this embodiment: system Windows 11, equipped with a single NVIDIA GeForce RTX 3050 GPU, deep learning framework Pytorch2.0.1. The prediction loss of the model proposed in this embodiment is lower than the loss of any mainstream model.

[0065] Table 1: Ablation experiment of improved raccoon optimization algorithm

[0066]

[0067]

[0068] Table 2: Comparative experiment of improved raccoon optimization algorithm

[0069]

[0070] Table 3: Overall architecture ablation experiment

[0071]

[0072] Note: A: external parameters B: internal parameters + external parameters

[0073] Table 4: Comparative test of autoregressive feedforward generative adversarial network

[0074]

[0075] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.

Claims

1. A method for ultra-short-term photovoltaic output prediction based on an adaptive feedforward generation network, characterized in that: The collected electrical and meteorological data are used to identify the internal parameters of photovoltaic modules through the improved Raccoon optimization algorithm. The internal parameters are combined with meteorological factors to predict the output through the feedforward generative adversarial network ultra-short-term photovoltaic output prediction model.

2. The ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network according to claim 1 is characterized in that: Method for identifying internal parameters: Setting internal parameters I ph , I o , R s , R sh , the initial value of n is 0, 0, 0, 100, 1; the interval ranges are 0-10A, 0-1A, 0-1Ω, 100-200Ω, 1-3, respectively. For the initial data, it does not fluctuate much in a small range. The improved raccoon optimization algorithm is applied to the data in groups of four for internal parameter identification. The maximum number of iterations is set to 100 and the number of populations is 30. It is vectorized and chaotically initialized to generate individuals throughout the search space, and then the hunting and attack of coatis on iguanas is simulated through exploration; Through the development of a process simulating its escape from predators, the idea of ​​token ring update is introduced to improve the formula by introducing global information to address the problem of difficulty in convergence caused by the lack of population information guidance and only searching around the current individual during the development process. The next solution to be updated is selected according to the current token position, and the token position is adjusted at the end of each iteration so that the search can be started from a different starting point in the next iteration. At the same time, the applicability of the entire problem space is limited by adding global constraints. At the end of each development, the SLSQP algorithm is used for precise convergence to obtain a higher quality solution. The ablation experiment confirms the indispensability of chaotic initialization, token ring improvement and sequential least squares programming optimization in the algorithm. At the same time, comparative experiments are introduced to confirm the superiority of this improved algorithm.

3. The ultra-short-term photovoltaic output prediction method based on the adaptive feedforward generation network according to claim 2 is characterized in that: The method of chaos initialization: Equations (1)-(2) are used to make a chaotic sequence and map each component separately. Z ij =μZ ij (1-Z ij )(j=1,2,...,d) (1) X ij =b lj +(b uj -b lj )Z ij (i=1,2,...,M;j=1,2,...,d) (2) Among them, μ represents the control parameter, Z ij is a d-dimensional vector Z i The weight in b lj and b uj Represents the lower and upper bounds of the mapping interval.

4. The ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network according to claim 3 is characterized in that: The token ring update method: Formula (3) is used to update the current token index. At the end of each iteration, the token position is adjusted using formula (4) to ensure that the search starts from a different point in the next iteration. token=(token+1)%SearchAgents_no(4) Among them, token is the token ring, starting from 0, i is the individual index in the population, starting from 0, SearchAgents n o represents the population size.

5. The ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network according to claim 4 is characterized in that: The method of generating adversarial networks is as follows: the input cascade data is first preliminarily encoded through a fully connected layer and a ReLU activation function; then the local subsequence information is focused on through a self-attention layer, the data dimension is adjusted through a linear layer, and then different contextual information in the input data is captured through a two-head attention mechanism and mapped to a potential vector space, and layer normalization and residual connection are introduced after each type of attention mechanism to alleviate the gradient disappearance in the deep network; two LSTM hidden layers and a fully connected layer are set in the decoder to convert it into a target sequence, and for the discriminator, a one-dimensional convolutional neural network is designed to make judgments. In addition, the model also uses a pre-trained SVM to construct a hinge loss, and combines the discriminator loss and the generator loss to generate a more accurate PV power sequence. The 188-dimensional real data composed of the measured power and the support data selected by grayscale correlation analysis (GRA) is denoted as x, the data concatenated by the noise and support data is denoted as z, and the fake data simulated by the generator is recorded as G(z). Then, the prediction results are adaptively optimized through the three joint losses listed in equations (5)-(7). The excavator loss is used to reconstruct and optimize the generator, and x and G(z) are sent to the discriminator to calculate the discriminator loss. The discriminator loss is combined with the trained SVM to calculate the hinge loss. 1 h =max(0,1-y·f(x)) (7) Among them, z is the z (z) is the data concatenated with the noise and support data of the Pr distribution, x is the real data concatenated with the measured power and support data that conforms to the Pr distribution, G(·) is the generator, D(·) is the discriminator, E represents the expectation, y is the label, indicating whether it is real data, and its value is {-1, 1}, f(x) is the predicted output of the support vector machine model, l G is the generator loss, l D is the discriminator loss, l h is the hinge loss.

6. The ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network according to claim 5 is characterized in that: For the generator, the network is mainly composed of self-attention layer, multi-head attention layer, LSTM layer and fully connected layer. The cascade data z is first transferred to a new feature space through a 188*128 weight matrix through a fully connected layer for preliminary feature extraction and dimensionality reduction; then the nonlinearity is introduced through the ReLU activation function to enable the model to learn more complex features. The selected ReLU activation function can effectively alleviate the gradient disappearance problem, and then the self-attention layer is used to generate a query matrix, key matrix and value matrix with a shape of 128*128 for the input 128-dimensional features, and important information is retained by focusing on local subsequences while reducing computational complexity; the linear layer is used for feature conversion to adjust the output dimension to 64; with the help of a dual-head attention layer with each head of 64*32 The attention mechanism performs parallel calculations to capture different contextual information in the input data to enhance the expressiveness of the model and enable it to focus on richer features. At the same time, layer normalization and residual connections are added after each attention layer. Layer normalization improves the stability and speed of training and alleviates internal covariate shift. The residual connection adds the input directly to the output of a certain layer, which helps to alleviate the gradient vanishing problem in deep networks and makes information more easily propagated. That is, for the two-headed attention layer, the output is the sum of the input and output values, which is still 64 dimensions. The features extracted by the encoder are input into the decoder composed of LSTM layers. The hidden layer is set to two layers with dimensions of 64 and 128 respectively. Then the output of the LSTM layer is converted to the target dimension of 188 through the fully connected layer.

7. The ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network according to claim 6 is characterized in that: For the discriminator, it is set as a one-dimensional convolutional neural network with an input layer size of 188. Three one-dimensional convolutional layers are added to extract local features in the time series and capture the relationship between asynchronous ones. The maximum pooling is introduced to reduce the feature dimension. A fully connected layer is set to extract high-level features. Finally, the tanh activation function is used to set the output range between [-1, 1], which is regarded as the confidence score for the real samples and the generated samples. At the same time, a SVM model with a linear kernel is constructed for training and prediction to achieve joint loss.

8. The ultra-short-term photovoltaic output prediction method based on an adaptive feedforward generation network according to claim 7 is characterized in that: The method for predicting power output uses the improved COA to identify internal parameters and then uses grey relational analysis (GRA) to select internal and external factors that have a greater impact on power output through formula (8): in, and They represent the jth power output of the photovoltaic power station and the jth data of the first indicator respectively. ρ represents a constant, which is taken as 0.

5. Taking into account the joint influence of internal and external parameters, a total of 7 attributes including three-phase average current, light radiation instrument parameters, solar irradiance, two temperature parameters, photocurrent Iph and ideal factor n are introduced as strongly correlated variables. In order to predict the output power in the next four hours, the real output is combined with the supporting data composed of the above strongly correlated variables to construct real data with a dimension of 20*7+12*4. The noise is combined with the relevant supporting data to form false data, which are respectively input into the generator and discriminator for training. The false data first undergoes preliminary feature extraction through the fully connected layer to transform the dimension into 128, and then passes through the self-attention layer, linear layer and The two-head attention mechanism layer focuses on key areas and captures contextual information. Layer normalization and residual connection are added after each layer. The dimension conversions are 128, 64 and 64 respectively. The extracted features are input into the decoder composed of two layers of LSTM, and then output through the fully connected layer with a dimension of 188. The real data and false data are then sent to the discriminator with a dimension set to 64 for judgment, and their confidence scores are output. During the training process, the training times of the generator and the discriminator are adjusted as needed to avoid mode collapse caused by overfitting of a certain component. At the same time, a linear kernel SVM is constructed for auxiliary training to enhance the stability of the generator. This improved feedforward generative adversarial network is used to predict and analyze the photovoltaic power output in different seasons within the next 4 hours.

Images