PCVMD-CNN-GRU-Attention load prediction method based on micro-grid

Abstract

The invention relates to the technical field of power system load prediction, and discloses a microgrid-based PCVMD-CNN-GRU-Attention load prediction method, which comprises the following steps: S1, collecting microgrid power load data, and preprocessing the data; s2, constructing a fitness function for decomposing VMD by a sparrow algorithm; s3, for a sparrow algorithm optimization VMD parameter, after the parameter is determined, dividing an original load into a plurality of sub-modals and a residual quantity; s4, performing frequency analysis on the sub-modal data and the residual data; s5, inputting the normalized component data into a CNN-GRU-Attention model for training and prediction, and storing the number of the models which is the same as the number of the components; and S6, respectively predicting each component. According to the method, feature extraction is carried out on the multi-dimensional historical data by using the CNN, modeling is carried out on the time sequence front-back relation of the extracted features by using an Attention mechanism in combination with the GRU model, and finally short-term load prediction is completed, so that the beneficial effect of preprocessing the original data according to the historical data of the microgrid load and the generated model is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of load forecasting for power systems, and in particular to a load forecasting method based on PCVMD-CNN-GRU-Attention for microgrids. Background Art

[0002] With the development of distributed energy storage and distributed photovoltaics, a large amount of energy resources lack unified management, resulting in huge waste. With the continuous innovation of information technology, microgrid systems can effectively improve the efficiency of new energy use, and the development of microgrids is becoming more and more rapid. Accurate load forecasting of user-side microgrids can help microgrid decision makers make better decisions based on actual needs, thereby improving energy utilization. Therefore, it is of great significance to study energy consumption prediction methods in microgrid systems. Nowadays, many different energy load forecasting methods have been proposed, but they also have shortcomings. The method that considers the temporal characteristics of data makes full use of the time series of load data, but the disadvantage is that the model cannot fully learn complex temporal characteristics; the traditional decomposition-integration prediction algorithm, although the complexity of the load is reduced, the law of the original load characteristics destroys the original load characteristics during the decomposition process. The decomposition effect of the traditional VMD algorithm depends greatly on the number of decompositions and the penalty factor, so how to select appropriate parameters is of great significance. Summary of the invention

[0003] 1. Technical issues to be resolved

[0004] In view of the shortcomings of the prior art, the present invention provides a load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid, which has the advantages of generating a model to pre-process the original data according to the historical data of the microgrid load; providing a model data basis for predicting future load data; and not requiring too many additional parameters, which is very suitable for actual engineering applications, thereby solving the problems of the above-mentioned technology.

[0005] (II) Technical solution

[0006] To achieve the above object, the present invention provides the following technical solution: a load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid, comprising the following steps:

[0007] S1, collect microgrid power load data and pre-process the data;

[0008] S2, construct the fitness function of the sparrow algorithm to decompose VMD; select the weighted combination of the aggregation algebra of the components after decomposition, the permutation entropy value, and the mutual information between the components and the original load as the fitness function;

[0009] S3, for the sparrow algorithm to optimize VMD parameters, after determining the parameters, the original load is divided into several sub-modes and a residual;

[0010] S4. Perform frequency analysis on the submodal data and residual data. For components with frequencies higher than 1500 Hz, standard deviation normalization is used, and for components with frequencies lower than 1500 Hz, maximum and minimum value normalization is used.

[0011] S5. Input the normalized component data into the CNN-GRU-Attention model for training and prediction, and save the same number of models as the number of components;

[0012] S6. Predict each component separately, denormalize the prediction results, and finally superimpose the normalized prediction results of each sub-modal to obtain the final calculation result and complete the short-term load forecast.

[0013] Preferably, the improved VMD implementation process steps in S2 are:

[0014] S1.1. Perform Hilbert transform on the original signal to obtain the analytical signal;

[0015] S1.2. Convert the spectrum to the fundamental frequency, where the center frequency formula estimated by multiplying the spectrum by the exponential term is:

[0016]

[0017] S1.3. Using the demodulated signal to estimate the bandwidth, the following formula expresses the constrained variational problem:

[0018]

[0019] S1.4, Lagrange multiplier λ(t) and quadratic penalty factor α are introduced to transform it into an unconstrained variational problem to be solved, expressed as:

[0020]

[0021] S1.5. Use the multiplier exchange method to solve the unconstrained variational problem, perform continuous iterations, and update h k and ω k , the updated expression is as follows:

[0022]

[0023] In the above expression: h k is the Kth modal component generated when the data is decomposed; ω kis the center frequency corresponding to the Kth modal component; g(t) is the time function, δ(t) is the impulse function; j is the unity of imaginary numbers; t is time; * is the convolution operator; α is the penalty factor; λ is the Lagrange multiplier; f is the input signal; n is the number of iterations; is the Fourier isometric transform of the discrete modal frequency domain; is the Fourier isometric transform of the input signal in frequency domain; is the Fourier isometric transform of the Lagrange multiplier in the frequency domain.

[0024] Preferably, the method steps for determining the decomposition level K of the VMD and the value of the penalty factor α are as follows:

[0025] S2.1, set an initial K and α value, and use the VMD algorithm to decompose the corresponding components;

[0026] S2.2, calculate the center frequency (omega), permutation entropy (PE) of each component, and the mutual information value (MI) between each component and the original load;

[0027] S2.3. Use f as the fitness function of the sparrow algorithm to search for the optimal value for calculation, where C1=0.75, C2=1.5, which are used to match the weights of PE and MI. The weights are: a large number of experiments have resulted in the optimal constant term. The value of omega is too large, and the fitness function is not easy to reflect the values ​​of PE and MI. Use lg(omega) instead.

[0028] Preferably, the optimal value search function expression in S2.3 is:

[0029]

[0030] Preferably, the optimization process of the sparrow algorithm is as follows:

[0031] S3.1, calculation starts, and several [k, α] combinations are randomly generated as the initial positions of the discoverer and follower;

[0032] S3.2, calculate the fitness value f and sort it during VMD decomposition;

[0033] S3.3, update the location of the discovered and followed persons according to the warning value;

[0034] S3.4, perform VMD decomposition on each finder and follower position [k,α];

[0035] S3.5, calculate the aggregation number, permutation entropy value and mutual information value with the original sequence of each IFM;

[0036] S3.6, randomly select an observer and update the position;

[0037] S3.7, determine whether it is the best position and fitness value, if not, repeat step S3.3, if yes, proceed to step S3.8;

[0038] S3.8, loop to calculate the fitness function and obtain the minimum value;

[0039] S3.9. Output the VMD parameter combination [k, α] to complete the operation.

[0040] Preferably, the normalization method is: standard deviation normalization and maximum and minimum value normalization.

[0041] Preferably, the standard deviation method normalization selection condition is: when the center frequency f of the component>1500Hz, the component is normalized using the standard deviation method.

[0042] Preferably, the maximum-minimum value method normalization selection condition is: except for the component center frequency f>1500Hz, the remaining components are normalized using the maximum-minimum value method.

[0043] Preferably, the training and prediction results of the CNN-GRU-Attention model are denormalized, and then the denormalized data are superimposed and reconstructed, that is, the load prediction results of each sub-modal are added to form the final short-term load forecast result.

[0044] Preferably, the value range of K is: [2, 8], and the optimization interval of α is: [500, 3000].

[0045] Compared with the prior art, the present invention provides a load forecasting method based on PCVMD-CNN-GRU-Attention in microgrid, which has the following beneficial effects:

[0046] 1. The present invention uses CNN to extract features from multidimensional historical data, uses the Attention mechanism combined with the GRU model to model the temporal relationship of the extracted features, and finally completes short-term load forecasting, achieving the beneficial effect of pre-processing the original data based on the historical data of the microgrid load and generating a model.

[0047] 2. The present invention generates a model to pre-process the original data based on the historical data of the microgrid load, and constructs the fitness function of the sparrow algorithm based on aggregation algebra, permutation entropy value, and mutual information parameters. The sparrow algorithm is used to optimize the parameters of VMD based on the fitness function, and finally the VM is decomposed to obtain the decomposition result, thereby achieving the beneficial effect of providing a model data basis for predicting future load data.

[0048] 3. The present invention can select historical load data according to the actual situation of the microgrid, or use aggregation algebra, permutation entropy, and mutual information combination as the fitness function, use the sparrow algorithm to optimize the VMD parameters, and use VMD to decompose the original power load data to reduce the complexity of the original load data, thereby achieving the beneficial effect of not requiring too many additional parameters and being very suitable for actual engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a structural diagram of the GRU model of the present invention;

[0050] Figure 2 Schematic diagram of the attention mechanism unit structure of the present invention;

[0051] Figure 3 Schematic diagram of the CNN-GRU-Attention prediction model of the present invention;

[0052] Figure 4 Schematic diagram of the VMD optimization process of the sparrow algorithm of the present invention;

[0053] Figure 5 This is a schematic diagram of the PCVMD-CNN-GRU-Attention prediction process of the present invention;

[0054] Figure 6 This is a schematic diagram of the original load data of the present invention;

[0055] Figure 7 Schematic diagram of modal components and residual components after PCVMD decomposition of the present invention;

[0056] Figure 8 This is a schematic diagram of the PCVMD-CNN-GRU-Attention prediction results of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Embodiment 1

[0059] See also Figure 1 ,S1: Obtain the historical load data of the microgrid, with the time scale accurate to minutes, generally one to three years of historical load data. ,Furthermore, the collected data is preprocessed, filling null values ​​and deleting duplicate values.

[0060] S2: Use the combination of aggregation algebra, permutation entropy, and mutual information as the fitness function of the sparrow algorithm.

[0061] S3: The VMD parameters are optimized using the sparrow algorithm, and the K and α values ​​are iteratively optimized. In order to avoid over-decomposition and under-decomposition, the value range of K is selected as [2,8], and the optimization interval of α is [500,3000]. The permutation entropy is used to measure the complexity of time series data. It is suitable for the analysis of non-stationary signals and has good robustness. Aggregation algebra reflects the frequency characteristics of each component. The smaller the value of omega, the faster the aggregation speed. Mutual Information is a measure of the degree of mutual dependence between variables, which can measure the inclusion relationship between the original load data and the decomposed components.

[0062] The formula for permutation entropy is as follows:

[0063]

[0064] The formula for mutual information is as follows:

[0065]

[0066] The aggregation algebra can reflect the frequency characteristics of each component. The smaller the value of omega is, the faster the aggregation speed is.

[0067] Permutation entropy is used to measure the complexity of time series data. It is suitable for the analysis of non-stationary signals and has good robustness. The larger the permutation entropy of a component, the stronger the random volatility of the data. The smaller the permutation entropy of a component, the smaller the degree of random volatility and the more regular and orderly the data.

[0068] Mutual Information is a measure of the degree of mutual dependence between variables, and can measure the inclusion relationship between the original load data and the decomposed components.

[0069] The fitness function is as follows:

[0070]

[0071] Step 3: Use the sparrow algorithm and the constructed fitness function to optimize the VMD parameter decomposition layer number K and penalty factor α.

[0072] It should be noted that:

[0073] Sparrow Search Algorithm (SSA) is an optimization algorithm based on the foraging behavior of sparrows in nature. It simulates the behavioral strategy of sparrows in the foraging process. In the algorithm, the sparrow population is divided into several different roles: observers, producers, discoverers, and followers.

[0074] Responsible for finding food for the crowd and providing guidance to followers. The number of discoveries is PD, and the position is updated as follows:

[0075]

[0076] Where: t represents the current iteration number; R2 represents the warning value, which is greater than 0 and less than 1; ST∈[0.5,1] is the safety value; L is a matrix with all elements equal to 1.

[0077]

[0078] A + Each element in the matrix is ​​either 1 or -1. Alerts are randomly generated by individuals at the group boundary, accounting for 10-20% of the total number. Their positions are updated as follows:

[0079]

[0080] Among them, in the global best is the optimal position; f g and f w Represents the best fitness function value and the worst fitness function value; ε is the smallest constant and β is a parameter.

[0081] Among them, variational mode decomposition (VMD) is an adaptive time-frequency analysis method for processing non-stationary signals. VMD decomposes the original complex signal into K intrinsic mode functions with limited bandwidth.

[0082] And extract the center frequency of the corresponding IMF, thereby effectively avoiding the end effect and mode aliasing problems of EMD. The essence of VMD is to construct and solve a variational problem. The implementation process of the above improved VMD includes the following steps:

[0083] (1) Perform Hi1bert transform on the original signal and modulate the signal spectrum to the corresponding frequency band;

[0084]

[0085] (2) where δ(t) is the impulse function h k (t) is the Kth modal component after decomposition, t is the time ω k is the corresponding IMF component h k (t) is the center frequency; j represents the imaginary unit.

[0086] (3) Estimating the bandwidth using the demodulated signal The following formula represents the constrained variational problem:

[0087]

[0088] (4)where f(t) is the original input signal. In order to transform the constraint condition into an unconstrained problem, the Lagrangian function is introduced:

[0089]

[0090] The solution is solved by using the alternating direction multiplier algorithm and it is continuously updated and iterated until K IMF components are obtained.

[0091]

[0092] (5) When using VMD decomposition, the number of decomposition layers needs to be set manually, which is often blind. Therefore, combined with the residual generated by VMD decomposition, we determine that the decomposition components include f k (t) and S res .

[0093]

[0094] After determining the parameters, the final submodal components and residuals are obtained, and standard deviation normalization and maximum and minimum value normalization are used. When the center frequency of the component is greater than 1500HZ, the component is normalized using the standard deviation, and the remaining components are normalized using the maximum and minimum values.

[0095] (6) The normalized data is modeled and predicted using the CNN-GRU-Attention model, in which the convolutional neural network (CNN) can extract multi-angle and deep-level features of the time series. Compared with the multi-dimensional input features designed manually, CNN has better discrimination and generalization capabilities: Generally, CNN includes input layer, convolution layer, pooling layer, fully connected layer and output layer, and its main parts are convolution layer and pooling layer. The convolution layer is used to extract the effective nonlinear local features of the input time series, and the pooling layer is used to compress the extracted features to generate more important feature information. The CNN-GRU-Attention model mainly includes: input layer, output layer, CNN layer, GRU layer and attention mechanism layer. The power load data is extracted through the CNN layer. The convolution operation is designed to increase its depth to compress the number of parameters, and then the feature dimension is reduced through pooling. Finally, the fully connected layer converts the features into a one-dimensional structure to complete the extraction of feature vectors. The GRU layer and the attention mechanism layer learn the inherent laws of power load from the features extracted by CNN so as to accurately predict power load data. Finally, the predicted results are output through the output layer.

[0096] (7) Among them, the gated recurrent unit (GRU) is a simplification and improvement of the long short-term memory neural network (LSTM). Among them, the gated recurrent unit (GRU) is a simplification and improvement of the long short-term memory neural network (LSTM). Compared with LSTM, GRU reduces one "gate" unit and only retains the update gate and reset gate, thereby simplifying the model and improving the operation speed. The internal structure of GRU is as follows Figure 2 , the calculation process is as follows:

[0097] z t =σ(W (z) x t +U (z) h t-1 )

[0098] r t =σ(W (r) x t +U (r) h t-1 )

[0099]

[0100] The arrows are the direction of the data; matrix multiplication is represented by ; the activation function of Sigmoid is represented by ; Tanh is the activation function Tanh; 1- represents forward propagation, and the propagated data is ; is the update gate, and the function of the update gate is to determine how much past data is transmitted to the present. is the reset gate, and its function is to determine how much past data needs to be forgotten; is the output of the hidden layer, and is the input; is the input x t and the past hidden state h t-1 The sum of U (z) ,W (z) ,U (r) ,W (r) ,U and W are parameters involved in the calculation.

[0101] The attention mechanism can determine the importance of input features for prediction results and assign different weights to various features, so that key features can be further highlighted. The structure of the attention unit is as follows: Figure 3 As shown. The attention mechanism gives greater weight to important information in a probabilistic way, that is, it highlights the role of key information in the event and improves the accuracy of prediction. Among them, X1....Xn represents the input of the model; H1....Hn represents the output of the hidden layer; α1....αn is the weight value assigned by the attention mechanism to the output of the GRU hidden layer, and Y is the output of the model.

[0102] The calculation process of Attention is as follows:

[0103] e ti =VT tanh(Wh t +Uh i +b),i=1,2,…,t-1

[0104]

[0105] h' t =f(X,h t ,x t )

[0106] In the formula, V, W, U, and b are model training parameters, and their values ​​are continuously iterated during training.

[0107] 1) In order to shorten the load forecasting time and improve the forecasting accuracy, the present invention constructs a PCVMD-CNN-GRU-Attention forecasting model. The steps of model prediction are as follows:

[0108] 2) PCVMD decomposition: Input the original microgrid power load data into the model, use VMD decomposition to obtain the modal components and residuals, construct the fitness function, and use the sparrow algorithm to determine the VMD parameters K and α.

[0109] 3) In order to solve the problem that VMD has difficulty in decomposing weak high-frequency signals, the sub-modalities with frequencies greater than 1500 Hz are normalized using the standard deviation method, and the remaining sub-modalities are normalized using the maximum and minimum method. The sequence after sub-modal normalization is input into the CNN-GRU-Attention model to perform classification training and prediction on different sub-modalities.

[0110] 4) Forecast result output: First, the sub-modal forecast data obtained in the previous step are denormalized; then, the denormalized data are superimposed and reconstructed, that is, the load forecast results of each sub-modal are added together to form the final short-term load forecast result.

[0111] 5) Prediction evaluation index: In order to quantify and compare the prediction results, the present invention uses the coefficient of determination (R2), mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) as model accuracy evaluation indicators.

[0112] The calculation formula is:

[0113]

[0114] Where: L t Represents the true value of the data; represents the predicted value of the data; M represents the total amount of data.

[0115] Embodiment 2

[0116] A microgrid power load forecasting method based on PCVMD-CNN-GRU-Attention is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation and comparative experiments.

[0117] Using the data of an Australian microgrid, the K value and α after PCVMD optimization are 4 and 1714 respectively, and the fitness function is 0.072. The optimized VMD decomposition obtains 4 IMF modal components and a residual term, such as Figure 8 It can be seen that IMF1 is a low-frequency signal, which reflects the overall change trend of power load very well. IMF2 and IMF3 are medium-frequency signals that can show fluctuations in smaller cycles, while IMF4 represents load changes in the high-frequency band.

[0118] According to the frequency law of each mode itself, different normalization methods are selected for the mode, and a model is established for the normalized mode. Different components are predicted according to the model, and the prediction results are denormalized. Finally, the prediction values ​​are superimposed together. In order to improve the accuracy of the prediction results, the collected data is preprocessed. 70% of the data set is selected as the training set and 30% as the test set. The number of iterations is 300 times, the initial learning rate is 0.01, and the prediction result is close to the true value. The prediction result shows that the PCVMD-CNN-GRU-Attention model established in this application can effectively learn the historical load data law, and the prediction effect is good.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to be limiting. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0120] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid is characterized by: The following steps are involved: S1, collect microgrid power load data and pre-process the data; S2, construct the fitness function of the sparrow algorithm to decompose VMD; select the weighted combination of the aggregation algebra of the components after decomposition, the permutation entropy value, and the mutual information between the components and the original load as the fitness function; S3, for the sparrow algorithm to optimize VMD parameters, after determining the parameters, the original load is divided into several sub-modes and a residual; S4. Perform frequency analysis on the submodal data and residual data. For components with frequencies higher than 1500 Hz, standard deviation normalization is used, and for components with frequencies lower than 1500 Hz, maximum and minimum value normalization is used. S5. Input the normalized component data into the CNN-GRU-Attention model for training and prediction, and save the same number of models as the number of components; S6. Predict each component separately, denormalize the prediction results, and finally superimpose the normalized prediction results of each sub-modal to obtain the final calculation result and complete the short-term load forecast.

2. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 1 is characterized in that: The steps of the improved VMD implementation process in S2 are: S1.

1. Perform Hilbert transform on the original signal to obtain the analytical signal; S1.

2. Convert the spectrum to the fundamental frequency, where the center frequency formula estimated by multiplying the spectrum by the exponential term is: S1.

3. Using the demodulated signal to estimate the bandwidth, the following formula expresses the constrained variational problem: S1.4, Lagrange multiplier λ(t) and quadratic penalty factor α are introduced to transform it into an unconstrained variational problem to be solved, expressed as: S1.

5. Use the multiplier exchange method to solve the unconstrained variational problem, perform continuous iterations, and update h k and ω k , the updated expression is as follows: In the above expression: h k is the Kth modal component generated when the data is decomposed; ω k is the center frequency corresponding to the Kth modal component; g(t) is the time function, δ(t) is the impulse function; j is the unity of imaginary numbers; t is time; * is the convolution operator; α is the penalty factor; λ is the Lagrange multiplier; f is the input signal; n is the number of iterations; is the Fourier isometric transform of the discrete modal frequency domain; is the Fourier isometric transform of the input signal in frequency domain; is the Fourier isometric transform of the Lagrange multiplier in the frequency domain.

3. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 2 is characterized in that: The steps of determining the decomposition level K and the penalty factor α value of the VMD are as follows: S2.1, set an initial K and α value, and use the VMD algorithm to decompose the corresponding components; S2.2, calculate the center frequency (omega), permutation entropy (PE) of each component, and the mutual information value (MI) between each component and the original load; S2.

3. Use f as the fitness function of the sparrow algorithm to search for the optimal value for calculation, where C1=0.75, C2=1.5, which are used to match the weights of PE and MI. The weights are: a large number of experiments have resulted in the optimal constant term. The value of omega is too large, and the fitness function is not easy to reflect the values ​​of PE and MI. Use lg(omega) instead.

4. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 3 is characterized in that: The optimal value search function expression in S2.3 is:

5. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 2 is characterized in that: The optimization process of the sparrow algorithm is as follows: S3.1, calculation starts, and several [k, α] combinations are randomly generated as the initial positions of the discoverer and follower; S3.2, calculate the fitness value f and sort it during VMD decomposition; S3.3, update the location of the discovered and followed persons according to the warning value; S3.4, perform VMD decomposition on each finder and follower position [k,α]; S3.5, calculate the aggregation number, permutation entropy value and mutual information value with the original sequence of each IFM; S3.6, randomly select an observer and update the position; S3.7, determine whether it is the best position and fitness value, if not, repeat step S3.3, if yes, proceed to step S3.8; S3.8, loop to calculate the fitness function and obtain the minimum value; S3.

9. Output the VMD parameter combination [k, α] to complete the operation.

6. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 2 is characterized in that: The normalization methods are: standard deviation normalization and maximum and minimum value normalization.

7. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 6 is characterized in that: The standard deviation method normalization selection condition is: when the center frequency f of the component>1500Hz, the component is normalized by the standard deviation method.

8. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 7 is characterized in that: The maximum and minimum value method normalization selection condition is: except for the component center frequency f>1500Hz, the remaining components are normalized using the maximum and minimum value method.

9. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 2 is characterized in that: The training and prediction results of the CNN-GRU-Attention model are denormalized, and then the denormalized data are superimposed and reconstructed, that is, the load prediction results of each sub-modal are added together to form the final short-term load prediction result.

10. The load forecasting method based on PCVMD-CNN-GRU-Attention of microgrid according to claim 5, characterized in that: The value range of K is: [2,8], and the optimization interval of α is: [500,3000].

Images