Super-short-term photovoltaic power generation power prediction method based on secondary decomposition and BiGRU

Abstract

The invention relates to an ultra-short-term photovoltaic power generation power prediction method based on secondary decomposition and BiGRU, and the method comprises the following steps: S1, carrying out the primary decomposition of an original signal based on CEEMDAN, and obtaining a plurality of components with different complexities; s2, calculating an entropy value of each component by using a sample entropy, and reconstructing a signal by using a K-means clustering algorithm; s3, carrying out secondary decomposition on the complex signal by adopting VMD, and taking the decomposed signal as input; s4, carrying out normalization processing on each component, respectively carrying out modeling prediction, and carrying out linear summation to form a prediction result of Clu-IMF0; s5, evaluating a prediction result of the model; according to the method, in the process of processing time series data by the multi-layer stacked BiGRU neural network, rules and features in the data can be well mined; the complexity of original data is greatly reduced through data processing; the method improves the prediction precision of the photovoltaic power, and has a good application prospect in solving the problem of power prediction of a photovoltaic power station with less data.

Description

Technical Field

[0001] The present invention relates to an ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU, and in particular to a prediction model based on complete empirical mode decomposition with adaptive noise (CEEMDAN)-variational mode decomposition (VMD)-bidirectional gated recurrent unit (BiGRU). Background Art

[0002] Solar energy is the most abundant renewable energy source, and photovoltaic power generation technology has developed rapidly, but its volatility affects the security of the power grid. Existing statistical, physical, deep learning and hybrid prediction methods still have shortcomings in prediction accuracy and dealing with the uncertainty of photovoltaic power generation. Statistical models based on historical data, such as time series analysis and regression analysis, are relatively effective for short-term predictions, but are less effective for long-term predictions and predictions under extreme weather conditions. Prediction methods based on physical principles, such as numerical weather prediction (NWP) and radiation transfer models, can provide more accurate predictions, but are computationally complex and costly. Deep learning methods based on neural networks perform well in processing complex data and pattern recognition, but require a large amount of training data and may be sensitive to overfitting. Hybrid prediction methods that combine multiple prediction methods, such as the combination of deep learning and statistical models, can improve the accuracy and robustness of predictions, but designing a suitable hybrid strategy remains a challenge. There is an urgent need for a method that can significantly improve the accuracy of ultra-short-term photovoltaic power generation prediction. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide an ultra-short-term photovoltaic power prediction method based on quadratic decomposition and BiGRU, which reduces the complexity of photovoltaic power data and improves the prediction accuracy of photovoltaic power.

[0004] The present invention adopts the following technical solution:

[0005] The present invention provides an ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU, which adopts the following steps:

[0006] S1. Based on CEEMDAN, the original signal is decomposed once to obtain several components of different complexity;

[0007] S2, using sample entropy to calculate the entropy value of each component, and reconstructing the signal using K-means clustering algorithm;

[0008] S3, use VMD to perform secondary decomposition on the complex signal and use the decomposed signal as input;

[0009] S4, normalize each component, perform modeling and prediction respectively, and linearly sum them to form the prediction result of Clu-IMF0;

[0010] S 5. Evaluate the prediction results of the model.

[0011] Furthermore, the decomposition process of the original signal in S1 is specifically as follows:

[0012] By adding white noise of different amplitudes to the original signal multiple times, and performing empirical mode decomposition (EMD) on each noise signal, the ensemble average of the first intrinsic mode function (IMF) obtained is taken as the first mode of CEEMDAN; this mode is subtracted from the original signal to obtain a new residual signal, and the above process is repeated for the residual signal until the residual does not exceed two extreme points and cannot be further decomposed, the calculation is terminated, and the two extreme points are the maximum point and the minimum point; a series of intrinsic mode functions arranged from high frequency to low frequency are output, and the intrinsic mode functions are different frequency components of the original signal.

[0013] Furthermore, CEEMDAN uses the following steps to decompose the original signal:

[0014] S11, the signal to be decomposed f i (t) Add n Gaussian white noise sequences with a mean of 0 to construct the sequence to be decomposed f for n experiments i (t), using the following formula:

[0015] f i (t)=f(t)+ε0w i (t),i=1,2,3,...,n (1)

[0016] Where: f i (t) is the signal to be decomposed; f(t) is the original signal sequence; ε0 is the signal-to-noise ratio; w i (t) is the white noise sequence added for the i-th time;

[0017] S12. For each f i (t) is decomposed by EMD algorithm to obtain the first modal component (IMF) and the first residual component r1(t), using the following formula

[0018]

[0019] r1(t)=f(t)-IMF1(t) (3)

[0020] Where: EMD1(·) is the first component obtained by EMD decomposition;

[0021] S13, adding noise to the residual component obtained after decomposition and continuing to decompose using EMD, the formula is as follows:

[0022]

[0023] r k (t) = r k-1 (t)-IMF k (t) (5)

[0024] In the formula, ε k is the signal-to-noise ratio of the kth white noise; EMD k (·) is the kth component obtained by EMD decomposition; IMF k (·) is the kth modal component;

[0025] S14. When the residual component does not exceed two extreme points, it means that the residual component can no longer be further decomposed into components that meet the inherent mode function (IMF) conditions, and the calculation is terminated; at this time, the residual trend is obvious and direct, and the original signal sequence f(t) is decomposed into n modal components and the residual term R(t), and the formula is as follows

[0026]

[0027] Furthermore, S2 specifically adopts the following steps:

[0028] S21, calculating the entropy value of each intrinsic mode function (IMFs) using a sample entropy algorithm;

[0029] S22, use K-means clustering method to classify IMFs;

[0030] S23. According to the clustering results, all IMFs belonging to the same class are merged and reconstructed into complex signal Clu-IMF0, simple signal Clu-IMF1 and trend signal Clu-IMF2.

[0031] Furthermore, S3 specifically adopts the following steps:

[0032] S31. Set the number of modes, center frequency, penalty term and number of iterations, then construct a variational problem to find a set of intrinsic mode functions. The process is as follows:

[0033]

[0034] Where f(t) is the original signal sequence; u k is a limited bandwidth; w k is the center frequency; K is the number of IMFs after decomposition; δ(t) is the unit pulse function; is the partial derivative of t; * is the convolution symbol;

[0035] S32. In the variational problem, in order to solve the constrained optimization problem, the Lagrange multiplier λ is introduced as an auxiliary term to deal with the equality constraint, and the second-order penalty factor α is added as an auxiliary term to deal with the inequality constraint or enhance the stability of the solution. Then, the constrained variational problem is transformed into an unconstrained variational problem, and the augmented Lagrangian expression is obtained as follows:

[0036]

[0037] S33, use the alternating direction multiplier strategy to calculate the saddle point, that is, the optimal solution to the original problem, and solve u k 、w k and λ, which are calculated as follows:

[0038]

[0039] Where: τ is the noise tolerance; yes u i where λ(t) is the Fourier transform of f(t) and λ(t).

[0040] Furthermore, S4 is specifically:

[0041] First, each component is normalized; then the linear summation method is used for the complex signal Clu-IMF0, the prediction results of each component are added together to form a whole, and the prediction is performed using BiGRU; the specific process is as follows:

[0042] r t =σ(W r *[h t-1 ,x t ]) (10)

[0043] z t =σ(W z *[h t-1 ,x t ]) (11)

[0044]

[0045]

[0046] In the formula, r t 、z t are reset gate and update gate respectively; σ is Sigmoid activation function; W is weight matrix; is the intermediate memory state; x t is the input at time t; h t is the output gate at time t; * is matrix multiplication;

[0047] The output information at time t is the sum of the forward and reverse hidden layer outputs, and its calculation formula is:

[0048]

[0049] Where: GRU(·) is the GRU hidden layer state corresponding to the vector encoding; are the forward and backward hidden layer output states respectively; α t and β t are the output weights of the corresponding hidden layers; b t is the corresponding bias of the hidden layer state at time t;

[0050] For the simple signal Clu-IMF1 and the trend signal Clu-IMF2, BiGRU is directly used to predict them; the prediction results of the complex signal Clu-IMF0, the simple signal Clu-IMF1 and the trend signal Clu-IMF2 are added together and integrated into a final prediction result.

[0051] Furthermore, S5 specifically selects the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R-square) to evaluate the prediction accuracy of various prediction models;

[0052] The calculation formulas for the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R-square) are:

[0053]

[0054] Furthermore, BiGRU is a bidirectional neural network composed of two GRU hidden layers, which can connect the output of the network at the current moment with the state of the previous moment and the state of the next moment, which is more conducive to the extraction of deep features.

[0055] The positive effects of the present invention are as follows: the multi-layer stacked BiGRU neural network of the present invention can well mine the laws and features in the data during the process of processing time series data; the prediction performance of the complete empirical mode decomposition with adaptive noise (CEEMDAN)-variational mode decomposition (VMD)-bidirectional gated recurrent unit (BiGRU) model is significantly improved, and the comparison of the prediction effect in rainy weather is particularly obvious. The processing of data greatly reduces the complexity of the original data; the prediction accuracy of photovoltaic power is improved, and it has good application prospects in solving the power prediction problem of photovoltaic power stations with little data. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Attached Figure 1 A schematic diagram of the network structure of BiGRU of the present invention;

[0057] Attached Figure 2Schematic diagram of the network structure of the GRU of the present invention;

[0058] Attached Figure 3 It is a flowchart of the power prediction process of the present invention;

[0059] Attached Figure 4 This is a comparison chart of the prediction effects of the CEEMDAN-VMD-BiGRU model. DETAILED DESCRIPTION

[0060] The BiGRU network structure is as follows Figure 1 As shown in the figure, BiGRU is a neural network architecture that combines a bidirectional structure with a GRU; it processes the forward and reverse sequence data separately through two GRU networks, and then merges the outputs of the two networks to capture the bidirectional long-term dependencies in the sequence data. The BIGRU structure contains two independent GRU networks: one processes the data from the beginning to the end of the sequence (forward), and the other processes the data from the end to the beginning of the sequence (reverse). Each GRU network contains an update gate and a reset gate to control the flow of information in the forward and reverse directions respectively.

[0061] The GRU structure is as follows Figure 2 As shown in the figure, from left to right, it consists of an input layer, a hidden layer, and an output layer. In the hidden layer, the tanh activation function is used. The network adopts a forward propagation method, that is, starting from the input layer, passing through the hidden layer, and finally reaching the output layer. At each time step, the network updates its state to capture the long-term dependencies in the sequence data.

[0062] As attached Figure 3 As shown, the ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU of the present invention adopts the following steps:

[0063] S1. Based on CEEMDAN, the original signal is decomposed once to obtain several components of different complexity;

[0064] The original signal decomposition process is as follows:

[0065] By adding white noise of different amplitudes to the original signal multiple times, and performing empirical mode decomposition (EMD) on each noise signal, the ensemble average of the first intrinsic mode function (IMF) obtained is taken as the first mode of CEEMDAN; this mode is subtracted from the original signal to obtain a new residual signal, and the above process is repeated for the residual signal until the residual does not exceed two extreme points and cannot be further decomposed, the calculation is terminated, and the two extreme points are a maximum point and a minimum point; a series of intrinsic mode functions arranged from high frequency to low frequency are output, and the intrinsic mode functions are different frequency components of the original signal.

[0066] CEEMDAN uses the following steps to decompose the original signal:

[0067] S11, the signal to be decomposed f i (t) Add n Gaussian white noise sequences with a mean of 0 to construct the sequence to be decomposed f for n experiments i (t), using the following formula:

[0068] f i (t)=f(t)+ε0w i (t),i=1,2,3,...,n (1)

[0069] Where: f i (t) is the signal to be decomposed; f(t) is the original signal sequence; ε0 is the signal-to-noise ratio; w i (t) is the white noise sequence added for the i-th time;

[0070] S12. For each f i (t) is decomposed by EMD algorithm to obtain the first modal component (IMF) and the first unique residual component r1(t), using the following formula

[0071]

[0072] r1(t)=f(t)-IMF1(t) (3)

[0073] Where, EMD1(·) is the first component obtained by EMD decomposition;

[0074] S13, adding noise to the residual component obtained after decomposition and continuing to decompose using EMD, the formula is as follows:

[0075]

[0076] r k (t) = r k-1 (t)-IMF k (t) (5)

[0077] In the formula, ε k is the signal-to-noise ratio of the kth white noise; EMD k (·) is the kth component obtained by EMD decomposition; IMF k (·) is the kth modal component;

[0078] S14. When the residual component does not exceed two extreme points, it means that the residual component can no longer be further decomposed into a component that satisfies the intrinsic mode function (IMF) condition, and the calculation is terminated. At this time, the residual trend is obvious and direct, and the original signal sequence is decomposed into n modal components and the residual term R(t):

[0079]

[0080] S2. Calculate the entropy value of each component using sample entropy, and reconstruct the signal using K-means clustering algorithm; specifically, the following steps are adopted:

[0081] S21. Calculate the entropy value of each intrinsic mode function (IMFs) using a sample entropy algorithm. The calculation process of the sample entropy algorithm is as follows:

[0082] ①Let X = [x1, x2, …, x N ] is a time series of length N, representing an intrinsic mode function (IMF);

[0083] ②Define a subsequence X of length m m (i) = [x i ,x i+1 ,…,x i+m+1 ], where 1≤i≤N-m+1;

[0084] ③For each X m (i), calculate its relationship with all other subsequences X m (j) The maximum distance d[x m (i),x m (j)] = max|x i+k-1 -x j+k-1 |;

[0085] ④ Set a similarity threshold r and count each X m (i) Satisfaction

[0086] d[x m (i),x m (j)]≤r X m (j) The number of B m (i);

[0087] ⑤Calculate the average number of matching vectors

[0088] ⑥ Increase the value of m to m+1, and repeat steps ② to ⑤ to calculate

[0089] ⑦ Sample entropy SampEn is defined as:

[0090] S22. Use K-means clustering method to classify IMFs. The classification process is as follows:

[0091] ① Let S = {s1, s2, …, s n} is a set of n sample entropy values, each s icorresponds to an intrinsic mode function (IMF);

[0092] ② Select k cluster centers C = {c1, c2, ..., c k};

[0093] ③For each sample s i , assigning it to the nearest cluster center

[0094] ④Update cluster center: Where S j belongs to cluster center c j A sample set of

[0095] ⑤ Repeat steps ③ and ④ until the cluster center no longer changes significantly;

[0096] S23. According to the clustering results, all IMFs belonging to the same class are merged and reconstructed into complex signal Clu-IMF0, simple signal Clu-IMF1 and trend signal Clu-IMF2. The mathematical description of merging and reconstruction is as follows:

[0097] ①Assume j belongs to cluster center c j The set of all IMFs;

[0098] ② Merge all IMFs belonging to the same category:

[0099] ③The reconstructed signal X' is the sum of all merged IMF and residual components: Where R is the residual component.

[0100] S3. Use VMD to perform secondary decomposition on the complex signal and use the decomposed signal as input. The specific steps are as follows:

[0101] S31. Set the number of modes, center frequency, penalty term and number of iterations, then construct a variational problem to find a set of intrinsic mode functions. The process is as follows:

[0102]

[0103]

[0104] Where f(t) is the original signal sequence; u k is a limited bandwidth; w k is the center frequency; K is the number of IMFs after decomposition; δ(t) is the unit pulse function; is the partial derivative of t; * is the convolution symbol;

[0105] S32. In the variational problem, in order to solve the constrained optimization problem, the Lagrange multiplier λ is introduced as an auxiliary term to deal with the equality constraint, and the second-order penalty factor α is added as an auxiliary term to deal with the inequality constraint or enhance the stability of the solution. Then, the constrained variational problem is transformed into an unconstrained variational problem, and the augmented Lagrangian expression is obtained as follows:

[0106]

[0108] S33, use the alternating direction multiplier strategy to calculate the saddle point, that is, the optimal solution to the original problem, and solve u k 、w k and λ, which are calculated as follows:

[0109]

[0110] Where: τ is the noise tolerance; yes u i where λ(t) is the Fourier transform of f(t) and λ(t).

[0111] The specific mathematical process of the alternating direction multiplier is as follows:

[0112] ① Initialization: Choose a suitable initial value x (0) and (0) .

[0113] ②Update:

[0114] ③ Repeat step ② until a certain convergence standard is reached, such as the number of iterations limit, error threshold, etc.;

[0115] In the above process, f(·) and g(·) are the smooth and non-smooth parts of the two variables, respectively, and A is a short matrix that may contain linear and nonlinear terms. The alternating direction multiplier minimizes the objective function of the original problem by alternatingly updating the two variables while ensuring that the update of each variable is in the direction of reducing the original objective function.

[0116] S4. Each component is normalized, modeled and predicted respectively, and linear summation is performed to form the prediction result of Clu-IMF0; specifically:

[0117] First, each component is normalized; then the linear summation method is used for the complex signal Clu-IMF0, the prediction results of each component are added together to form a whole, and the prediction is performed using BiGRU; the specific process is as follows:

[0118] r t =σ(W r *[h t-1,x t ]) (10)

[0119] z t =σ(W z *[h t-1 ,x t ]) (11)

[0120]

[0121]

[0122] In the formula, r t 、z t are reset gate and update gate respectively; σ is Sigmoid activation function; W is weight matrix; is the intermediate memory state; x t is the input at time t; h t is the output gate at time t; * is matrix multiplication;

[0123] The output information at time t is the sum of the forward and reverse hidden layer outputs, and its calculation formula is:

[0124]

[0125] Where: GRU(·) is the GRU hidden layer state corresponding to the vector encoding; are the forward and backward hidden layer output states respectively; α t and β t are the output weights of the corresponding hidden layers; b t is the corresponding bias of the hidden layer state at time t;

[0126] For the simple signal Clu-IMF1 and the trend signal Clu-IMF2, BiGRU prediction is directly adopted; the prediction results of the complex signal Clu-IMF0, the simple signal Clu-IMF1 and the trend signal Clu-IMF2 are added together and integrated into a final prediction result.

[0127] S5. Evaluate the prediction results of the model, specifically, select the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R-square) to evaluate the prediction accuracy of various prediction models;

[0128] The calculation formulas for the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R-square) are:

[0129]

[0130]

[0131]

[0132] Furthermore, BiGRU is a bidirectional neural network composed of two GRU hidden layers, which can connect the output of the network at the current moment with the state of the previous moment and the state of the next moment, which is more conducive to the extraction of deep features.

[0133] As attached Figure 4 As shown, three consecutive days with different weather conditions are selected as test days (the data from November 20 to December 18 are training samples, and the data from December 19 to December 21 are test samples). The photovoltaic power prediction results of the method of the present invention for three consecutive days are intuitively compared with those of the LSTM, GRU and BiLSTM methods. No matter it is sunny, cloudy or rainy, the method of the present invention is superior to the other three methods.

[0134] The multi-layer stacked BiGRU neural network of the present invention can well mine the rules and features in the data during the process of processing time series data; the prediction performance of the complete empirical mode decomposition with adaptive noise (CEEMDAN)-variational mode decomposition (VMD)-bidirectional gated recurrent unit (BiGRU) model is significantly improved, and the comparison of the prediction effect in rainy weather is particularly obvious. The processing of data greatly reduces the complexity of the original data; the prediction accuracy of photovoltaic power is improved, and it has good application prospects in solving the power prediction problem of photovoltaic power stations with little data.

Claims

1. An ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU, characterized in that: Use the following steps: S1. Based on CEEMDAN, the original signal is decomposed once to obtain several components of different complexity; S2, using sample entropy to calculate the entropy value of each component, and reconstructing the signal using K-means clustering algorithm; S3, use VMD to perform secondary decomposition on the complex signal and use the decomposed signal as input; S4, normalize each component, perform modeling and prediction respectively, and linearly sum them to form the prediction result of Clu-IMF0; S5. Evaluate the prediction results of the model.

2. The ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU according to claim 1 is characterized in that: The original signal decomposition process in S1 is as follows: By adding white noise of different amplitudes to the original signal multiple times, and performing empirical mode decomposition (EMD) on each noise signal, the ensemble average of the first intrinsic mode function (IMF) obtained is taken as the first mode of CEEMDAN; this mode is subtracted from the original signal to obtain a new residual signal, and the above process is repeated for the residual signal until the residual does not exceed two extreme points and cannot be further decomposed, the calculation is terminated, and the two extreme points are the maximum point and the minimum point; a series of intrinsic mode functions arranged from high frequency to low frequency are output, and the intrinsic mode functions are different frequency components of the original signal.

3. The ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU according to claim 2 is characterized in that: CEEMDAN uses the following steps to decompose the original signal: S11, the signal to be decomposed f i (t) Add n Gaussian white noise sequences with a mean of 0 to construct the sequence to be decomposed f for n experiments i (t), using the following formula: f i (t)=f(t)+ε0w i (t),i=1,2,3,...,n (1) Where: f i (t) is the signal to be decomposed; f(t) is the original signal sequence; ε0 is the signal-to-noise ratio; w i (t) is the white noise sequence added for the i-th time; S12. For each f i (t) is decomposed by EMD algorithm to obtain the first modal component (IMF) and the first residual component r1(t), using the following formula r1(t)=f(t)-IMF1(t) (3) Where: EMD1(·) is the first component obtained by EMD decomposition; S13, adding noise to the residual component obtained after decomposition and continuing to decompose using EMD, the formula is as follows: r k (t)=r k-1 (t)-IMF k (t) (5) In the formula, ε k is the signal-to-noise ratio of the kth white noise; EMD k (·) is the kth component obtained by EMD decomposition; IMF k (·) is the kth modal component; S14. When the residual component does not exceed two extreme points, it means that the residual component can no longer be further decomposed into components that meet the inherent mode function (IMF) conditions, and the calculation is terminated; at this time, the residual trend is obvious and direct, and the original signal sequence f(t) is decomposed into n modal components and the residual term R(t), and the formula is as follows 4. The ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU according to claim 1 is characterized in that: S2 specifically adopts the following steps: S21, calculating the entropy value of each intrinsic mode function (IMFs) using a sample entropy algorithm; S22, use K-means clustering method to classify IMFs; S23. According to the clustering results, all IMFs belonging to the same class are merged and reconstructed into complex signal Clu-IMF0, simple signal Clu-IMF1 and trend signal Clu-IMF2.

5. The ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU according to claim 1 is characterized in that: S3 specifically adopts the following steps: S31. Set the number of modes, center frequency, penalty term and number of iterations, then construct a variational problem to find a set of intrinsic mode functions. The process is as follows: Where f(t) is the original signal sequence; u k is a limited bandwidth; w k is the center frequency; K is the number of IMFs after decomposition; δ(t) is the unit pulse function; is the partial derivative of t; * is the convolution symbol; S32. In the variational problem, in order to solve the constrained optimization problem, the Lagrange multiplier λ is introduced as an auxiliary term to deal with the equality constraint, and the second-order penalty factor α is added as an auxiliary term to deal with the inequality constraint or enhance the stability of the solution. Then, the constrained variational problem is transformed into an unconstrained variational problem, and the augmented Lagrangian expression is obtained as follows: S33, use the alternating direction multiplier strategy to calculate the saddle point, that is, the optimal solution to the original problem, and solve u k 、w k and λ, which are calculated as follows: Where: τ is the noise tolerance; yes u i where λ(t) is the Fourier transform of f(t) and λ(t).

6. The ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU according to claim 1 is characterized in that: S4 is specifically: First, each component is normalized; then the linear summation method is used for the complex signal Clu-IMF0, the prediction results of each component are added together to form a whole, and the prediction is performed using BiGRU; the specific process is as follows: r t =σ(W r *[h t-1 ,x t ]) (10) z t =σ(W z *[h t-1 ,x t ]) (11) In the formula, r t 、z t are reset gate and update gate respectively; σ is Sigmoid activation function; W is weight matrix; is the intermediate memory state; x t is the input at time t; h t is the output gate at time t; * is matrix multiplication; The output information at time t is the sum of the forward and reverse hidden layer outputs, and its calculation formula is: Where: GRU(·) is the GRU hidden layer state corresponding to the vector encoding; are the forward and backward hidden layer output states respectively; α t and β t are the output weights of the corresponding hidden layers; b t is the corresponding bias of the hidden layer state at time t; For the simple signal Clu-IMF1 and the trend signal Clu-IMF2, BiGRU is directly used to predict them; the prediction results of the complex signal Clu-IMF0, the simple signal Clu-IMF1 and the trend signal Clu-IMF2 are added together and integrated into a final prediction result.

7. The ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU according to claim 6 is characterized in that: S5 specifically selects the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R-square) to evaluate the prediction accuracy of various prediction models; The calculation formulas for the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R-square) are:

8. The ultra-short-term photovoltaic power generation prediction method based on quadratic decomposition and BiGRU according to claim 6 is characterized in that: BiGRU is a bidirectional neural network composed of two GRU hidden layers. It can connect the output of the network at the current moment with the state of the previous moment and the state of the next moment, which is more conducive to the extraction of deep features.

Images