Short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and clustering reconstruction

Abstract

The invention discloses a short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and clustering reconstruction. The short-term photovoltaic power deep learning prediction method mainly comprises the following steps: step 1, decomposing a photovoltaic power time sequence by using successive variational mode decomposition; 2, calculating the complexity of each decomposition component by using a fuzzy entropy algorithm; 3, dividing the components with different complexities by using Gaussian hybrid clustering, and reconstructing the components into high-frequency, intermediate-frequency and low-frequency components; step 4, using a chaos evolution optimization algorithm to optimize hyper-parameters of the bidirectional gating cycle unit (BIGRU) prediction model of each component; 5, performing short-term photovoltaic power prediction on the reconstructed high-frequency component, the reconstructed intermediate-frequency component and the reconstructed low-frequency component by adopting a BIGRU model; and step 6, superposing the predicted values of the components to obtain a photovoltaic power point prediction result. And step 7, using kernel density estimation to form a confidence interval based on a point prediction result, and finally obtaining a photovoltaic power probability interval prediction result. The method has the beneficial effects that the complexity of each decomposition component is evaluated and reconstructed by utilizing the fuzzy entropy algorithm, and comprehensive optimization of each link including model training is realized on the basis of powerful capabilities of capturing context information and processing long sequence data based on the BIGRU.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power prediction, and relates to a short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and clustering reconstruction. Background Art

[0002] In the context of the global low-carbon energy transition, photovoltaic power generation is highly volatile and significantly affected by meteorological factors. Therefore, high-precision photovoltaic power forecasting has become the key to grid absorption. Traditional methods rely on physical models or statistical learning, and have shortcomings such as delayed meteorological response, insufficient multi-source data fusion, and weak non-stationary sequence modeling capabilities. Although existing deep learning models have improved nonlinear fitting capabilities, they still face bottlenecks such as modal aliasing, subjectivity in parameter tuning, and lack of probabilistic prediction. This paper proposes a data-driven multi-algorithm collaborative prediction framework that takes into account both point prediction accuracy and probabilistic risk assessment to provide decision support for grid dispatch. Summary of the Invention

[0003] The purpose of this invention is to provide a data-driven multi-algorithm collaborative prediction framework to solve the problems of modal mixing, subjectivity in parameter tuning, and lack of probability prediction in existing photovoltaic power prediction technologies.

[0004] The technology used in this invention is a short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction, which is specifically implemented in the following steps:

[0005] Step 1: Decompose the photovoltaic power time series using the Successive Variational Mode Decomposition (SVMD) algorithm to obtain several components;

[0006] Step 2: Calculate the complexity of each decomposition component using the Fuzzy Entropy (FE) algorithm;

[0007] Step 3: Use Gaussian Mixture Model (GMM) to divide the components of different complexity and reconstruct them into high-frequency, medium-frequency, and low-frequency components;

[0008] Step 4: Optimize the hyperparameters of the bidirectional gated recurrent unit (BIGRU) prediction model of each component using Chaotic Evolution Optimization (CEO) algorithm;

[0009] Step 5: Use the BIGRU model to perform short-term photovoltaic power prediction on the reconstructed high-frequency, medium-frequency, and low-frequency components respectively;

[0010] Step 6: Superimpose the predicted values ​​of each component to obtain the photovoltaic power point prediction result;

[0011] Step 7: Use Kernel Density Estimation (KDE) to form a confidence interval based on the point prediction results, and finally obtain the photovoltaic power probability interval prediction result.

[0012] The present invention is also characterized in that:

[0013] Step 1 is implemented as follows:

[0014] Decompose the original signal f(t) into multiple modal components and a residual signal f r (t), satisfying:

[0015] f(t)=u L (t)+f r (t) (1)

[0016] Each modality u L (t) Specific frequency constraints must be met. The L-order mode should be subject to minimization constraints and has the following constraints:

[0017]

[0018]

[0019] In the equation, * represents the convolution operation, and w L Indicates the center frequency of the L-order mode.

[0020]

[0021] This filter is used to suppress the residual signal in w L The energy of the place.

[0022] With the above two constraints, it is impossible to distinguish between the L-order mode and the L-1st-order mode. We can introduce the idea of ​​J2 constraint and select an appropriate filter with the following frequency response, which is:

[0023]

[0024] Finally, the reconstruction conditions must be met:

[0025]

[0026] Step 2 is implemented as follows:

[0027] The steps to calculate the complexity of each decomposition component using the fuzzy entropy algorithm are as follows:

[0028] Step 2.1: u after SVMD decomposition L (t) component to reconstruct the phase space and generate the vector set X k (i)=[u k (i),u k (i+τ),...,u k (i+(m-1)τ)]

[0029] Step 2.2 For each modal component vector X k (i) and X k (j), calculate the Chebyshev distance:

[0030]

[0031] Step 2.3 uses Gaussian fuzzy function to calculate similarity:

[0032]

[0033] Step 2.4 For the embedding dimension m, calculate the similarity probability:

[0034]

[0035] Step 2.5 Repeat steps 2.2-2.4 to calculate the similarity rate when the embedding dimension is m+1 Final modal component u L The fuzzy entropy of (t) is defined as:

[0036]

[0037] Step 3 is implemented as follows:

[0038] The probability density of the entire data is the weighted sum of three Gaussian distributions:

[0039]

[0040] Where: π k is the mixing weight of the kth distribution; μ k is the mean of the kth distribution; is the variance of the kth distribution; is a one-dimensional Gaussian density function.

[0041] One-dimensional Gaussian density function formula:

[0042]

[0043] Expectation maximization is used to estimate the parameters in GMM. The iterative process includes:

[0044]

[0045]

[0046]

[0047]

[0048] After the expectation maximization algorithm converges, a classification label is assigned to each fuzzy element. The components corresponding to the same label are superimposed and reconstructed using the following formula:

[0049] u L (t) = u i (t)+u j (t) (17)

[0050] Where u i (t) and u j (t) the i-th component and the j-th component respectively;

[0051] Step 4 is implemented as follows:

[0052] The steps for optimizing model hyperparameters using the chaotic evolutionary optimization algorithm are as follows:

[0053] Step 4.1 Chaos initialization: The algorithm first uses a two-dimensional discrete memristor hyperchaotic map to generate chaotic candidate individuals, and uses the chaotic mapping formula (18) to generate a chaotic sequence of length N {z1, z2, z3....., z N}, map the chaotic sequence to the solution space using formula (19) to obtain chaotic candidate individuals, and then map the chaotic candidate individuals back to the actual position.

[0054] z t+1 =μz t (1-z t ) (18)

[0055] x j =Var min+z j ·(Var max-Var min) (19)

[0056] Where j = 1, 2, ..., Np, generating Np initial candidate solutions.

[0057] Step 4.2 Mutation operation: Generate evolution direction using formula (20)

[0058]

[0059] in, is the value of the chaotic candidate individual mapped back to the actual position, x tis a randomly selected individual from the population.

[0060] Update the individual through formula (21) and perform the mutation operation:

[0061]

[0062] Where a is the search step size. It is worth noting that Formula (21) has a strong global exploration capability, but this may lead to a slow convergence of the algorithm. Therefore, to further improve the local development capability of the algorithm, Formula (22) is used to search for the best solution in the current population to accelerate the convergence of the algorithm.

[0063]

[0064] Step 4.3 Crossover operation: Generate the test vector using formula (23):

[0065]

[0066] Where CR is the crossover probability, which is between 0 and 1, and j rand is a randomly selected dimension. And for each dimension j, a random number r is generated j , if r j <CR or j=j rand , then the test vector For mutant individuals Otherwise, the current individual

[0067] Step 4.4 Select the operation and use formula (24) to calculate the fitness value of the current individual and the experimental vector. If Then the mutant individual is replaced by the value of the experimental phasor, otherwise it remains unchanged.

[0068]

[0069] Step 4.5 repeats fitness evaluation, elite selection, and chaotic perturbation until the stopping condition is met, the change in optimal fitness is less than the threshold, as shown in formula (25), or the maximum number of iterations is reached, and finally the optimal hyperparameter combination is obtained.

[0070] |fBEST (t) -fBEST (t-1) |<∈ (25)

[0071] Step 5 is implemented as follows:

[0072] Use BIGRU gated recurrent unit to train the model for high-frequency, medium-frequency and low-frequency components. The mathematical model of BIGRU is:

[0073] r t =σ(W r ·[h t-1 ,x t ]) (26)

[0074] z t =σ(W z ·[h t-1 ,x t ]) (27)

[0075]

[0076]

[0077] Where: r t 、z t are reset gate and update gate respectively; σ is Sigmoid activation function; W is weight matrix; h t is the intermediate memory state; x t is the input at time t; h t is the hidden state at time t.

[0078] The output information at time t is the sum of the forward and reverse hidden layer outputs, and its calculation formula is:

[0079]

[0080] Where: G(·) is the GRU hidden layer state corresponding to the vector encoding; h′ t 、h″ t are the forward and backward hidden layer output states respectively; α t and β t are the output weights b of the corresponding hidden layers t is the corresponding bias of the hidden layer state at time t.

[0081] BIGRU is a bidirectional neural network composed of two GRU hidden layers, which can connect the output of the current moment in the network with the state of the previous moment and the state of the next moment, which is more conducive to the extraction of deep features.

[0082] Step 6 is implemented as follows:

[0083] The predicted values ​​of each subsequence are superimposed to obtain the photovoltaic power prediction value and complete the prediction.

[0084] Step 7 is implemented as follows:

[0085] Select the Gaussian kernel function and use the Silverman rule to control the bandwidth, as shown in formulas (31) and (32).

[0086]

[0087]

[0088] For any point, the density is estimated as:

[0089]

[0090] The cumulative distribution function is calculated by discretization and the prediction interval is determined without confidence level:

[0091]

[0092] Interval 1-α =[q α / 2 ,q 1-α / 2 ] (35)

[0093] Assuming that the given confidence probability is m, the quantile q = 1-m is selected and recorded as the lower quantile and upper quantile, and then the lower quantile and upper quantile of the photovoltaic power are calculated by substituting them into the above formula to obtain the lower limit and upper limit of the prediction interval.

[0094] The beneficial effects of the present invention are: first, multi-level feature extraction of photovoltaic power time series is realized through SVMD and FE, and the components of each frequency in the photovoltaic power series are separated to improve the adaptability to complex environmental scenarios. Then, GMM is used to divide and reconstruct the components into high-frequency, medium-frequency and low-frequency subsequences. Secondly, the chaotic evolution optimization algorithm CEO is used to optimize the parameters of the bidirectional gated recurrent unit BIGRU. CEO simulates the chaotic evolution process of the two-dimensional discrete memristor mapping to perform global optimization of the hyperparameters of the BIGRU network, avoiding the subjectivity and local optimality problems of traditional parameter adjustment methods. The BIGRU network synchronously mines the forward and backward time dependencies through bidirectional time series modeling, and combines the gating mechanism to adaptively adjust the information flow, effectively capturing the spatiotemporal correlation of photovoltaic power in complex scenarios. In addition, the present invention realizes uncertainty quantification through KDE, avoids the limitations of traditional point prediction, and provides risk-controlled decision support for power grid scheduling. In summary, the present invention solves the problems of multi-scale dynamic characteristic modeling, hyperparameter adaptive matching and uncertainty quantification in photovoltaic power prediction through the deep integration of signal decomposition, intelligent optimization and bidirectional time series modeling. It combines accuracy, robustness and engineering practicality, and provides reliable technical guarantee for the safe and stable operation of new energy high-penetration power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 This is a BIGRU neural network structure diagram in the short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and clustering reconstruction of the present invention.

[0096] Figure 2 This is the overall flow chart of the system in the short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction of the present invention.

[0097] Figure 3 The present invention is a fuzzy entropy size graph of a short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction.

[0098] Figure 4 It is the SVMD decomposition diagram of the short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction in the present invention.

[0099] Figure 5 This is a reconstruction diagram of the SVMD-FE short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and clustering reconstruction of the present invention.

[0100] Figure 6 It is a CEO iteration graph in a short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction in the present invention.

[0101] Figure 7 This is a point prediction result diagram of a short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction in the present invention.

[0102] Figure 8 This is a graph of interval prediction results in a short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction in the present invention. DETAILED DESCRIPTION

[0103] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0104] The present invention provides a short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction, which is specifically implemented according to the following steps:

[0105] Step 1 is implemented as follows:

[0106] Decompose the original signal f(t) into multiple modal components and a residual signal f r (t), satisfying:

[0107] f(t)=u L (t)+f r (t) (4)

[0108] Each modality u L (t) Specific frequency constraints must be met. The L-order mode should be subject to minimization constraints and has the following constraints:

[0109]

[0110]

[0111] In the equation, * represents the convolution operation, and w L Indicates the center frequency of the L-order mode.

[0112]

[0113] This filter is used to suppress the residual signal in w L The energy of the place.

[0114] With the above two constraints, it is impossible to distinguish between the L-order mode and the L-1st-order mode. We can introduce the idea of ​​J2 constraint and select an appropriate filter with the following frequency response, which is:

[0115]

[0116] Finally, the reconstruction conditions must be met:

[0117]

[0118] Step 2 is implemented as follows:

[0119] Step 2.1: u after SVMD decomposition L (t) component to reconstruct the phase space and generate the vector set X k (i)=[u k (i),u k (i+τ),...,u k (i+(m-1)τ)]

[0120] Step 2.2 For each modal component vector X k (i) and X k (j), calculate the Chebyshev distance:

[0121]

[0122] Step 2.3 uses Gaussian fuzzy function to calculate similarity:

[0123]

[0124] Step 2.4 For the embedding dimension m, calculate the similarity probability:

[0125]

[0126] Step 2.5 Repeat steps 2.2-2.4 to calculate the similarity rate when the embedding dimension is m+1 Final modal component u L The fuzzy entropy of (t) is defined as:

[0127]

[0128] Step 3 is implemented as follows:

[0129] The probability density of the entire data is the weighted sum of three Gaussian distributions:

[0130]

[0131] Where: π k is the mixing weight of the kth distribution; μ k is the mean of the kth distribution; is the variance of the kth distribution; is a one-dimensional Gaussian density function.

[0132] One-dimensional Gaussian density function formula:

[0133]

[0134] Expectation maximization is used to estimate the parameters in GMM. The iterative process includes:

[0135]

[0136]

[0137]

[0138]

[0139] After the expectation maximization algorithm converges, a classification label is assigned to each fuzzy element. The components corresponding to the same label are superimposed and reconstructed using the following formula:

[0140] u L (t) = u i (t)+u j (t) (17)

[0141] Where u i (t) and u j (t) the i-th component and the j-th component respectively;

[0142] Step 4 is implemented as follows:

[0143] Step 4.1 Chaos initialization: The algorithm first uses a two-dimensional discrete memristor hyperchaotic map to generate chaotic candidate individuals, and uses the chaotic mapping formula (18) to generate a chaotic sequence of length N {z1, z2, z3....., z N}, map the chaotic sequence to the solution space using formula (19) to obtain chaotic candidate individuals, and then map the chaotic candidate individuals back to the actual position.

[0144] z t+1 =μz t (1-z t ) (18)

[0145] x j =Var min+z j ·(Var max-Var min) (19)

[0146] Where j = 1, 2, ..., Np, generating Np initial candidate solutions.

[0147] Step 4.2 Mutation operation: Generate evolution direction using formula (20)

[0148]

[0149] in, is the value of the chaotic candidate individual mapped back to the actual position, x t is a randomly selected individual from the population.

[0150] Update the individual through formula (21) and perform the mutation operation.

[0151]

[0152] Where a is the search step size. It is worth noting that Formula (21) has a strong global exploration capability, but this may lead to a slow convergence of the algorithm. Therefore, to further improve the local development capability of the algorithm, Formula (22) is used to search for the best solution in the current population to accelerate the convergence of the algorithm.

[0153]

[0154] Step 4.3 Crossover operation: Generate a test vector using formula (23).

[0155]

[0156] Where CR is the crossover probability, which is between 0 and 1, and j rand is a randomly selected dimension. And for each dimension j, a random number r is generated j , if r j <CR or j=j rand , then the test vector For mutant individuals Otherwise, the current individual

[0157] Step 4.4 Select the operation and use formula (24) to calculate the fitness value of the current individual and the experimental vector. If Then the mutant individual is replaced by the value of the experimental phasor, otherwise it remains unchanged.

[0158]

[0159] Step 4.5 repeats fitness evaluation, elite selection, and chaotic perturbation until the stopping condition is met, the change in optimal fitness is less than the threshold, as shown in formula (25), or the maximum number of iterations is reached, and finally the optimal hyperparameter combination is obtained.

[0160] |fBEST (t) -fBEST (t-1) |<∈ (25)

[0161] Step 5 is implemented as follows:

[0162] Use BIGRU gated recurrent unit to train the model for high-frequency, medium-frequency and low-frequency components. The mathematical model of BIGRU is:

[0163] r t =σ(W r ·[h t-1 ,x t ]) (26)

[0164] z t =σ(W z ·[h t-1 ,x t ]) (27)

[0165]

[0166]

[0167] Where: r t 、z t are reset gate and update gate respectively; σ is Sigmoid activation function; W is weight matrix; h t is the intermediate memory state; x t is the input at time t; h t is the hidden state at time t.

[0168] The output information at time t is the sum of the forward and reverse hidden layer outputs, and its calculation formula is:

[0169]

[0170] Where: G(·) is the GRU hidden layer state corresponding to the vector encoding; ht ′、h t ″ are the forward and backward hidden layer output states respectively; α t and β t are the output weights b of the corresponding hidden layers t is the corresponding bias of the hidden layer state at time t.

[0171] BIGRU is a bidirectional neural network composed of two GRU hidden layers, which can connect the output of the current moment in the network with the state of the previous moment and the state of the next moment, which is more conducive to the extraction of deep features.

[0172] Step 6 is implemented as follows:

[0173] The predicted values ​​of each subsequence are superimposed to obtain the photovoltaic power prediction value and complete the prediction.

[0174] Step 7 is implemented as follows:

[0175] Select the Gaussian kernel function and use the Silverman rule to control the bandwidth, as shown in formulas (31) and (32).

[0176]

[0177]

[0178] For any point, the density is estimated as:

[0179]

[0180] The cumulative distribution function is calculated by discretization and the prediction interval is determined without confidence level:

[0181]

[0182] Interval 1-α =[q α / 2 ,q 1-α / 2 ] (35)

[0183] Assuming that the given confidence probability is m, the quantile q = 1-m is selected and recorded as the lower quantile and upper quantile, and then the lower quantile and upper quantile of the photovoltaic power are calculated by substituting them into the above formula to obtain the lower limit and upper limit of the prediction interval.

[0184] Example

[0185] This paper selects the historical data of a photovoltaic power station as the experimental data set, and selects the historical power data of 21 sunny days as the data set, of which the first 20 days are the training set and the 21st day is the test set, with a total of 6048 data points. First, SVMD is used to decompose the power time series, which helps to strengthen the time series signal and compare between signals. The results of the algorithm processing are shown in the figure below. Figure 4 As shown, the complexity of several components after SVMD decomposition is calculated using fuzzy entropy, and the results are as follows Figure 3 As shown, Gaussian clustering is used to divide the six components into three categories, and each category of components is superimposed and reconstructed. The algorithm processing results are shown as follows Figure 5 shown.

[0186] Then, the CEO optimization algorithm is used to optimize the BIGRU network hyperparameters for each component input. The iterative process is as follows: Figure 6 As shown, the BIGRU network structure is as follows Figure 1 As shown in the figure, the parameters and processed components are input into the network model and trained to obtain the point prediction results and interval prediction results.

[0187] This study used 5,760 sets of experimental data for modeling and analysis, of which the first 5,472 sets were used for model training, and the remaining 288 independent test sets were used for performance evaluation. Experimental results show that the hybrid model, which uses fuzzy entropy reconstruction and chaotic evolutionary optimization algorithms for parameter optimization, achieves prediction performance indicators of 174.03% root mean square error, 128.81% mean square error, and 4.80% mean absolute percentage error. Compared with a baseline model that does not use feature reconstruction and intelligent optimization algorithms, all three evaluation indicators show significant improvement trends. The data comparison analysis shown in Table 1 demonstrates that the algorithm-improved hybrid model effectively captures the characteristics of short-term photovoltaic power fluctuations. The fit between its prediction curve and the actual power curve improves by approximately 12.6% compared to the baseline model, validating the applicability of the proposed method in the field of renewable energy power forecasting.

[0188] Table 1 Comparison of prediction indicators of various models

Claims

1. A short-term photovoltaic power deep learning prediction method based on decomposition sequence complexity evaluation and cluster reconstruction is implemented in the following steps: Step 1: Decompose the photovoltaic power time series using the Successive Variational Mode Decomposition (SVMD) algorithm to obtain several components; Step 2: Calculate the complexity of each decomposition component using the Fuzzy Entropy (FE) algorithm; Step 3: Use Gaussian Mixture Model (GMM) to divide the components of different complexity and reconstruct them into high-frequency, medium-frequency, and low-frequency components; Step 4: Optimize the hyperparameters of the bidirectional gated recurrent unit (BIGRU) prediction model of each component using Chaotic Evolution Optimization (CEO) algorithm; Step 5: Use the BIGRU model to perform short-term photovoltaic power prediction on the reconstructed high-frequency, medium-frequency, and low-frequency components respectively; Step 6: Superimpose the predicted values ​​of each component to obtain the photovoltaic power point prediction result; Step 7: Use Kernel Density Estimation (KDE) to form a confidence interval based on the point prediction results, and finally obtain the photovoltaic power probability interval prediction result.

2. According to the method for short-term photovoltaic power deep learning prediction based on decomposition sequence complexity evaluation and cluster reconstruction in claim 1, step 1 is specifically implemented as follows: Decompose the original signal f(t) into multiple modal components and a residual signal f r (t), satisfying: f(t)=u L (t)+f r (t)(1) Each modality u L (t) Specific frequency constraints must be met. The L-order mode should be subject to minimization constraints and has the following constraints: In the equation, * represents the convolution operation, and w L Indicates the center frequency of the L-order mode. This filter is used to suppress the residual signal in w L The energy of the place. With the above two constraints, it is impossible to distinguish between the L-order mode and the L-1st-order mode. We can introduce the idea of ​​J2 constraint and select an appropriate filter with the following frequency response, which is: Finally, the reconstruction conditions must be met:

3. According to the method for short-term photovoltaic power deep learning prediction based on decomposition sequence complexity evaluation and cluster reconstruction in claim 1, step 2 is specifically implemented as follows: Step 2.1: u after SVMD decomposition L (t) component to reconstruct the phase space and generate the vector set X k (i)=[u k (i),u k (i+τ),...,u k (i+(m-1)τ)] Step 2.2 For each modal component vector X k (i) and X k (j), calculate the Chebyshev distance: Step 2.3 uses Gaussian fuzzy function to calculate similarity: Step 2.4 For the embedding dimension m, calculate the similarity probability: Step 2.5 Repeat steps 2.2-2.4 to calculate the similarity rate when the embedding dimension is m+1 Final modal component u L The fuzzy entropy of (t) is defined as:

4. According to the method for short-term photovoltaic power deep learning prediction based on decomposition sequence complexity evaluation and cluster reconstruction of claim 1, step 3 is specifically implemented as follows: The probability density of the entire data is the weighted sum of three Gaussian distributions: Where: π k is the mixing weight of the kth distribution; μ k is the mean of the kth distribution; is the variance of the kth distribution; is a one-dimensional Gaussian density function. One-dimensional Gaussian density function formula: Expectation maximization is used to estimate the parameters in GMM. The iterative process includes: After the expectation maximization algorithm converges, a classification label is assigned to each fuzzy element. The components corresponding to the same label are superimposed and reconstructed using the following formula: u L (t)=u i (t)+u j (t) (17) Where u i (t) and u j (t) the i-th component and the j-th component respectively.

5. According to the method for short-term photovoltaic power deep learning prediction based on decomposition sequence complexity evaluation and cluster reconstruction of claim 1, step 4 is specifically implemented as follows: Step 4.1 Chaos initialization: The algorithm first uses a two-dimensional discrete memristor hyperchaotic map to generate chaotic candidate individuals, and uses the chaotic mapping formula (18) to generate a chaotic sequence of length N {z1, z2, z3....., z N }, map the chaotic sequence to the solution space using formula (19) to obtain chaotic candidate individuals, and then map the chaotic candidate individuals back to the actual position. z t+1 =μz t (1-z t ) (18) x j =Surely+z j ·(Varmax-Varmin) (19) Where j = 1, 2, ..., Np, generating Np initial candidate solutions. Step 4.2② Mutation operation: Generate evolution direction using formula (20) in, is the value of the chaotic candidate individual mapped back to the actual position, x t is a randomly selected individual from the population. Update the individual through formula (21) and perform mutation operation Where a is the search step size. It is worth noting that Formula (21) has a strong global exploration capability, but this may lead to a slow convergence of the algorithm. Therefore, to further improve the local development capability of the algorithm, Formula (22) is used to search for the best solution in the current population to accelerate the convergence of the algorithm. Step 4.3 Crossover operation: Generate the test vector using formula (23) Where CR is the crossover probability, which is between 0 and 1, and j rand is a randomly selected dimension. And for each dimension j, a random number r is generated j , if r j <CR or j=j rand , then the test vector For mutant individuals Otherwise, the current individual Step 4.4 Select the operation and use formula (24) to calculate the fitness value of the current individual and the experimental vector. If Then the mutant individual is replaced by the value of the experimental phasor, otherwise it remains unchanged. Step 4.5 repeats fitness evaluation, elite selection, and chaotic perturbation until the stopping condition is met, the change in optimal fitness is less than the threshold, as shown in formula (25), or the maximum number of iterations is reached, and finally the optimal hyperparameter combination is obtained. |fBEST (t) -fBEST (t-1) |<∈ (25) 6. According to the method for short-term photovoltaic power deep learning prediction based on decomposition sequence complexity evaluation and cluster reconstruction of claim 1, step 5 is specifically implemented as follows: The BIGRU gated recurrent unit is used to train the high-frequency, medium-frequency, and low-frequency components. The mathematical model of BIGRU is: r t =σ(W r ·[h t-1 ,x t ]) (26) z t =σ(W z ·[h t-1 ,x t ]) (27) Where: r t 、z t are reset gate and update gate respectively; σ is Sigmoid activation function; W is weight matrix; h t is the intermediate memory state; x t is the input at time t; h t is the hidden state at time t; The output information at time t is the sum of the forward and reverse hidden layer outputs, and its calculation formula is: Where: G(·) is the GRU hidden layer state corresponding to the vector encoding; h t ′、h t ″ are the forward and backward hidden layer output states respectively; α t and β t are the output weights b of the corresponding hidden layers t is the corresponding bias of the hidden layer state at time t. BIGRU is a bidirectional neural network composed of two GRU hidden layers, which can connect the output of the current moment in the network with the state of the previous moment and the state of the next moment, which is more conducive to the extraction of deep features.

7. The method for short-term photovoltaic power deep learning prediction based on decomposition sequence complexity evaluation and cluster reconstruction according to claim 1, wherein step 6 is specifically implemented as follows: The predicted values ​​of each subsequence are superimposed to obtain the photovoltaic power prediction value and complete the prediction.

8. The method for short-term photovoltaic power deep learning prediction based on decomposition sequence complexity evaluation and cluster reconstruction according to claim 1, wherein step 7 is specifically implemented as follows: This method selects the Gaussian kernel function and uses the Silverman rule to control the bandwidth, as shown in formulas (31) and (32). For any point, the density is estimated as: The cumulative distribution function is calculated by discretization and the prediction interval is determined without confidence level: Interval 1-α =[q α / 2 ,q 1-α / 2 ] (35) Assuming that the given confidence probability is m, the quantile q = 1-m is selected and recorded as the lower quantile and upper quantile, and then the lower quantile and upper quantile of the photovoltaic power are calculated by substituting them into the above formula to obtain the lower limit and upper limit of the prediction interval.

Images