Short-term wind power prediction method based on secondary decomposition and neural network

Abstract

The invention discloses a short-term wind power prediction method based on secondary decomposition and a neural network, and the method comprises the following steps: obtaining an original wind power signal, carrying out the noise reduction processing, and decomposing the original wind power signal into a plurality of initial mode components; preprocessing is carried out, and corresponding high-resolution characteristic signals are extracted; inputting the high-resolution characteristic signal into a pre-trained wind power prediction model; outputting a corresponding wind power prediction result; the wind power prediction model comprises two parallel branch networks; the first branch network is a two-way time domain convolution network and a two-way gating circulation unit; the second branch network is a two-dimensional convolutional neural network and a two-way long-short-term memory network; the two parallel branch networks are optimized on the basis of an improved salmons algorithm. According to the method, comprehensive optimization among different models and algorithms is considered, deep correlation and dynamic change characteristics among data are fully mined and utilized, the global search capability is enhanced, a local optimal solution is avoided, and short-term wind power prediction errors are remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and more specifically to short-term wind power prediction based on secondary decomposition and improved sandcat algorithm to optimize neural network. Background Art

[0002] In the process of achieving the national dual carbon goals, renewable energy plays a vital role, especially wind energy. Wind energy has the characteristics of easy access, no pollution, and abundant resources, and its importance in the power system is becoming increasingly prominent. However, due to the uncertainty of wind power, if wind power is directly input into the power grid, it may cause unstable operation of the power grid. Therefore, in the face of wind power with complex factors, it is particularly important to establish a prediction model with higher prediction accuracy.

[0003] At present, the research on wind power prediction model has been quite extensive. According to different power prediction methods, wind power prediction methods can be divided into three categories: physical methods, statistical methods and machine learning methods. Physical methods are mainly based on current meteorological parameters and geographical parameters, such as temperature, terrain information, pressure and obstacles. This type of method predicts by solving complex high-dimensional nonlinear equations. However, due to the complex modeling mechanism of physical methods and the difficulty in collecting the required information, its effect in short-term wind power prediction is not ideal. In contrast, statistical methods predict wind power by establishing statistical models. Common models include autoregressive moving average, autoregressive integrated moving average and gray model. Although the establishment process of statistical models is relatively simple compared with physical models, due to its weak nonlinear fitting ability, its prediction results in short-term prediction are not good. Compared with the two methods mentioned above, the prediction method of machine learning can obtain faster calculation speed and process more complex time series data. In particular, specialized structural designs such as long short-term memory (LSTM) and gated recurrent unit (GRU) effectively avoid the gradient explosion or disappearance problem in model training, making it one of the popular methods in the field of time series analysis such as wind power prediction. However, in actual application scenarios with high frequency and high noise, such as short-term wind power forecasting, the prediction accuracy of a single model is often unable to meet actual needs.

[0004] Therefore, establishing a prediction model with higher prediction accuracy is an urgent problem that technicians in this field need to solve. Summary of the invention

[0005] In view of this, the present invention provides a short-term wind power prediction method based on quadratic decomposition and neural network, which combines multiple models and intelligent optimization algorithms, and conducts in-depth analysis and optimization on the characteristics of data in specific application scenarios, thereby reducing the short-term wind power prediction error.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A short-term wind power prediction method based on quadratic decomposition and neural network includes the following steps:

[0008] The original wind power signal is obtained for noise reduction and decomposed into multiple initial modal components;

[0009] Preprocessing the multiple initial modal components to extract corresponding high-resolution feature signals;

[0010] Inputting the high-resolution feature signal into a pre-trained wind power prediction model;

[0011] Output the corresponding wind power prediction result;

[0012] Among them, the wind power prediction model is two parallel branch networks; among them, the first branch network is a bidirectional time domain convolutional network and a bidirectional gated recurrent unit; the second branch network is a two-dimensional convolutional neural network and a bidirectional long short-term memory network; the two parallel branch networks are optimized based on the improved sand cat algorithm.

[0013] Furthermore, the preprocessing specifically includes:

[0014] Applying ensemble empirical mode decomposition to decompose the initial modal components once to obtain multiple subsequences;

[0015] The subsequence is decomposed twice by using an improved ensemble empirical mode decomposition algorithm to extract high-resolution feature signals.

[0016] Furthermore, the first branch network optimizes the number of residual blocks and the spatial loss factor through the improved sand cat optimization algorithm to capture the complexity of time series data;

[0017] The bidirectional time-domain convolutional network uses multiple residual blocks and dilated convolutions to expand the receptive field of the model, perceive data dependencies in a longer range, and predict dynamic changes in wind power. The output of each time-domain convolutional block is processed by a residual block, and the processed outputs are merged through a deep connection layer, and the merged feature outputs are transmitted to the bidirectional gated recurrent unit layer.

[0018] The output of the bidirectional gated recurrent unit and the output of each time-domain convolutional block are again subjected to feature fusion through a deep connection layer.

[0019] Furthermore, the bidirectional gated recurrent unit uses two gated recurrent unit layers, the forward gated recurrent unit processes the input sequence from beginning to end, and the backward gated recurrent unit processes in the opposite direction; each gated recurrent unit layer maintains its own hidden state and updates it according to the current input and the hidden state of the previous time step.

[0020] Furthermore, the second branch network adjusts the convolution kernel stride, the maximum pooling layer and the number of hidden units of the bidirectional long short-term memory network through the improved sand cat optimization algorithm to capture spatial features;

[0021] The spatial features of the original data are extracted through a two-dimensional convolutional neural network, and then the reduced-dimensional sequence information is sent to a bidirectional long short-term memory network for further feature extraction.

[0022] Furthermore, the improved sand cat optimization algorithm specifically includes:

[0023] 1) Lens imaging reverse learning method:

[0024] By adjusting the reverse solution, the individual can jump out of the local optimal solution;

[0025] In two-dimensional space, [a, b] is the range of the solution, where the y-axis is a convex lens, the height of the object is h, its projection on the x-axis is x, and the height of the inverted real image of the object formed by the convex lens is h * , which is x on the x-axis * , the formula is:

[0026]

[0027] Let k = h / h*, and get the dynamic new position by adjusting the scaling factor n;

[0028] Through transformation, the reverse point position in the D-dimensional space is obtained as the latest position of the sand cat individual:

[0029]

[0030]

[0031] Among them, k is the ratio of the object height to the height of the inverted real image of the object, X best(t) is the current location of the sand cat individual, is the latest position of the sand cat individual, t is the current iteration number, and T is the maximum iteration number;

[0032] 2) Levi's flying and wandering method:

[0033] In the attack phase of the sand cat optimization algorithm, the sand cat individuals are enabled to switch between short-distance local search and long-distance jumping global search;

[0034] The update formula for the individual position of the sand cat is:

[0035] L=X BEST -X best(t)

[0036]

[0037] Among them, X BEST is the current optimal solution, RL is the random step coefficient, and L is the offset between the current position and the optimal position;

[0038] 3) Triangle walking method:

[0039] While gradually approaching the prey in one direction, the sand cat individuals perform a triangular random walk around the prey, which enables the individuals to move more flexibly during the search process and find a better solution.

[0040] The update formula for the individual position of the sand cat is:

[0041] A=L 2 +LP 2 -2*LP*L*cos(2*pi*rand)

[0042]

[0043] Among them, LP represents the random step size, represents the A step size when updating the position, cos(2*pi*rand) represents the generated random angle, rand represents a function that generates random numbers, and pi=π.

[0044] Furthermore, the two parallel branch networks are optimized based on the improved sand cat algorithm, including: the two parallel branch networks dynamically adjust the subsequence prediction weights using the improved sand cat algorithm, and merge the two parallel branch networks through dynamic optimization;

[0045] Specifically include:

[0046] 1) Set the weight vector w = [w 1 ,w 2 ],

[0047] Among them, w 1 and w 2 The weights corresponding to the first branch network and the second branch network respectively;

[0048] 2) Construct the objective function as:

[0049]

[0050] 3) Use the Lagrange multiplier method to introduce a multiplier λ to handle the weight constraints, w1 +w 2 =1 and

[0051] 4) Introduce the quadratic penalty function ψ(w) of the constraint condition:

[0052]

[0053] 5) Combine the objective function and the penalty function to form a new loss function:

[0054]

[0055] 6) Use dynamic optimization to adjust the weight vector w so that the final loss function is minimized, thereby achieving the best fusion of the two branch networks;

[0056] Where N is the number of samples in the validation set, i is the index of the sample, indicating the i-th sample, and y 1,i is the prediction output of the validation set of the first branch network for the i-th sample, y 2,i is the predicted output of the validation set of the second branch network for the i-th sample; M is 2 to ensure that there are at most two variables, j is the element index in the weight vector, and w j is the element in the weight vector; γ = 0.2 is the penalty parameter, which adjusts the constraint condition w 1 +w 2 =1 is the penalty intensity when it is violated; λ = 1.1 is the Lagrange multiplier.

[0057] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides a short-term wind power prediction method based on secondary decomposition and neural network, which has the following beneficial effects:

[0058] 1) The non-stationarity and nonlinearity of the original wind power series are eliminated, and the secondary decomposition architecture extracts the correlation between the data, further improving the prediction accuracy of the prediction model.

[0059] 2) By improving the sand cat optimization algorithm and flexibly applying the lens imaging reverse learning method, Levy flight walk method and triangle walk method, the problems of local optimality and insufficient exploration ability in the basic sand cat algorithm are overcome.

[0060] 3) The first branch network is constructed by combining a bidirectional temporal convolutional network with a bidirectional gated recurrent unit, which makes full use of the forward and backward time series information in the wind power data. The bidirectional temporal convolutional network layer expands the receptive field of the model by using multiple residual blocks and dilated convolutions, enabling it to perceive data dependencies in a longer range, thereby more accurately predicting the dynamic changes in wind power; the bidirectional gated recurrent unit layer acts as a deep feature extractor of the network to further process the time series data output by the bidirectional temporal convolutional network, which not only improves the accuracy of time series prediction, but also enhances the robustness and adaptability of the model when facing wind power data; at the same time, the outputs of the bidirectional temporal convolutional network and the bidirectional gated recurrent unit are deeply fused to ensure that the model can capture as much information as possible from all directions and levels.

[0061] 4) Construct the second branch network. The model first extracts the spatial features of the original data through a two-dimensional wide convolutional network, and then sends the reduced-dimensional sequence information to a bidirectional long short-term memory network for further feature extraction.

[0062] 5) Construct a branch network weight distribution model, and improve the accuracy and stability of the overall prediction model by dynamically optimizing and fusing two parallel branch networks; search for the optimal weight combination based on the verification set errors of the two branch networks to minimize the prediction error.

[0063] The present invention considers the comprehensive optimization among different models and algorithms, fully explores and utilizes the deep correlation and dynamic change characteristics among data, strengthens the global search capability and avoids the local optimal solution, and significantly reduces the short-term wind power prediction error. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0065] Figure 1 A framework diagram of the short-term wind power prediction method based on quadratic decomposition and improved sandcat algorithm to optimize neural network provided by the present invention.

[0066] Figure 2 This is a structural diagram of the first branch network model provided by the present invention.

[0067] Figure 3 This is a structural diagram of the second branch network model provided by the present invention. DETAILED DESCRIPTION

[0068] 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.

[0069] The embodiment of the present invention discloses a short-term wind power prediction method based on secondary decomposition and neural network. Figure 1 As shown, the following steps are included:

[0070] The original wind power signal is obtained for noise reduction and decomposed into multiple initial modal components;

[0071] Preprocessing the multiple initial modal components to extract corresponding high-resolution feature signals;

[0072] Inputting the high-resolution feature signal into a pre-trained wind power prediction model;

[0073] Output the corresponding wind power prediction result;

[0074] Among them, the wind power prediction model is two parallel branch networks; among them, the first branch network is a bidirectional time domain convolutional network and a bidirectional gated recurrent unit; the second branch network is a two-dimensional convolutional neural network and a bidirectional long short-term memory network; the two parallel branch networks are optimized based on the improved sand cat algorithm.

[0075] In this embodiment, in order to fully extract the correlation between the data, the data is preprocessed and a secondary decomposition of the data is constructed: the ensemble empirical mode decomposition (EEMD) decomposes the wind power after denoising by the variational mode decomposition (VMD) into three sub-sequences; secondly, the complete ensemble empirical mode decomposition (CEEMDAN) with adaptive noise is used to perform a secondary decomposition on the three components to reduce the non-stationarity and nonlinearity of the sequence and further improve the prediction accuracy.

[0076] 1) EEMD solves the problem that the classic empirical mode decomposition (EMD) is susceptible to modal aliasing by adding a white noise series to the original data. The calculation formula of EEMD is:

[0077]

[0078] Where x(t) is the data sequence after VMD denoising, ε k (t) is a k-th white noise sequence, NE is the number of decompositions, and EMD(·) is the classic empirical mode decomposition process.

[0079] 2) In order to further refine the feature extraction, CEEMDAN decomposes the three sequences obtained from EEMD. For each IMF obtained from EEMD, i , the CEEMDAN decomposition is defined as:

[0080] IMF C,i =CEEMDAN(IMF i ,N std ,NR,M) (2)

[0081] Where N std is the noise standard deviation, NR is the number of noise repetitions, and M is the maximum number of iterations.

[0082] In this embodiment, the high-resolution feature signal obtained after preprocessing is input into a pre-trained wind power prediction model;

[0083] The wind power prediction model consists of two parallel branch networks; the first branch network is a bidirectional time domain convolutional network and a bidirectional gated recurrent unit, namely, a BiTCN-BiGRU branch network; the second branch network is a two-dimensional convolutional neural network and a bidirectional long short-term memory network, namely, a CNN-BiLSTM branch network.

[0084] The two parallel branch networks improve the balance between the local optimum and global exploration capabilities of the sand cat optimization algorithm by improving it. Specifically, they adopt the lens imaging reverse learning strategy to improve search diversity and reduce the possibility of falling into the local optimum; apply the Levy flight walk strategy in the attack phase to enable individuals to flexibly switch between local and global searches, thereby improving convergence speed and accuracy; and introduce the triangle walk strategy to help sand cat individuals more efficiently approach the optimal solution in complex search spaces.

[0085] Specifically, the improved SandCat optimization algorithm includes:

[0086] 1) Lens imaging reverse learning method: Lens imaging reverse learning effectively overcomes the situation where sand cat individuals fall into local optimal solutions and have poor reverse solutions by flexibly adjusting the reverse solutions, thereby making it more likely for individuals to jump out of the local optimal solution and improving the overall efficiency and effectiveness of the algorithm.

[0087] Assume that in two-dimensional space, [a,b] is the possible range of the solution, where the y-axis is a convex lens. The height of the object is h, and its projection on the x-axis is x. The height of the inverted real image of the object formed by the convex lens is h* , which is x on the x-axis * The formula is as follows:

[0088]

[0089] Let k = h / h*, this strategy can get a dynamic new position by adjusting the scaling factor n. When n = 1, this strategy is a traditional reverse learning strategy.

[0090] The position of the reverse point in the D-dimensional space is obtained by transformation:

[0091]

[0092] Where k is the ratio of the object height to the height of the inverted real image of the object, X best(t) is the current location of the sand cat individual, is the latest position of the sand cat individual, t is the current iteration number, and T is the maximum iteration number.

[0093] In each iteration, the improved Sand Cat optimization algorithm generates the reverse solution of the current solution. The dynamic adjustment of the reverse solution can help the Sand Cat jump out of the local optimum and explore a better hyperparameter combination. Assuming that the Sand Cat is at the local optimum in the parameter space of the number of residual blocks and the spatial loss factor, the new position generated by reverse learning can enable it to find a better solution in a larger search space, thereby improving the global search ability of the algorithm.

[0094] 2) Levy flight walk method: In the attack phase of the sand cat optimization algorithm, random walks of different step lengths are used to enable sand cat individuals to flexibly switch between short-distance local searches and long-distance jumping global searches. The early Levy flight walk strategy ensures that sand cat individuals search in a diverse solution space to avoid premature convergence of the algorithm; in the later stage, the convergence speed is increased to quickly find the optimal configuration of the number of residual blocks and the spatial loss factor. The update formula for the position of the sand cat individual is:

[0095] L=X BEST -X best(t) (6)

[0096]

[0097] Where, X BEST Represents the current optimal solution, RL represents the random step coefficient, and L is the offset between the current position and the optimal position.

[0098] 3) Triangular walking method: When the sand cat gradually approaches the prey in a certain direction, it walks randomly in a triangle around it, which makes the individual walk more flexibly during the search process, helping to overcome obstacles in the search space and find a better solution. The update formula of the sand cat's individual position is:

[0099] A=L 2 +LP 2 -2*LP*L*cos(2*pi*rand) (8)

[0100]

[0101] Wherein, LP represents the random step size, A represents the step size when updating the position, cos(2*pi*rand) represents the generated random angle, rand represents a function for generating random numbers, and pi=π.

[0102] In this embodiment, the BiTCN-BiGRU branch network model structure diagram is shown in FIG. Figure 2 As shown in the figure, a combination of a bidirectional temporal convolutional network BiTCN and a bidirectional gated recurrent unit BiGRU is used to fully utilize the front-end and back-end time series information in the wind power data. The BiTCN layer expands the receptive field of the model by using multiple residual blocks and dilated convolutions, enabling it to perceive data dependencies over a longer range, thereby more accurately predicting dynamic changes in wind power. As a deep feature extractor of the network, the BiGRU layer further processes the time series data output by BiTCN, which not only improves the accuracy of time series prediction, but also enhances the robustness and adaptability of the model when facing wind power data; at the same time, the outputs of BiTCN and BiGRU are deeply fused to ensure that the model can capture as much information as possible from all directions and levels.

[0103] In order to fully extract the time series data in the temporal convolutional network TCN block, this embodiment adopts a structured information integration strategy. First, the output of each TCN block is processed by a residual block, which not only helps to maintain the stability of information in the deep network transmission process, but also enhances the implicit temporal features in the data; then, the processed output is merged through a deep connection layer to enhance the information flow in the model feedforward path, and the merged feature output is delivered to the bidirectional gated recurrent unit BiGRU layer. At the same time, the output of the bidirectional GRU and the output of each TCN block are again feature fused through a deep connection layer. This design enables the network to integrate multi-dimensional features at different processing stages and optimizes the information utilization of the model. In this way, the model can not only capture the instantaneous features of each time point, but also integrate the long-term dependencies of the entire sequence, thereby improving the accuracy of wind power prediction and the generalization ability of the model.

[0104] In order to further improve the prediction accuracy of wind power, the output of BiTCN is passed to the BiGRU layer. BIGRU solves the propagation direction problem by using two GRU layers. The forward GRU processes the input sequence from beginning to end, while the backward GRU processes it in the opposite direction. Each GRU layer maintains its own hidden state and updates it based on the current input and the hidden state of the previous time step. Compared with GRU, BIGRU has the advantages of bidirectional modeling ability, better long-term dependency modeling, more accurate sequence labeling, and powerful feature extraction ability. The calculation formula of its internal hidden layer unit is:

[0105]

[0106] In the formula, x t is the feature vector input to the BiGRU layer at time t; and is the forward and backward state of the hidden layer at time t; and is the forward and backward weights of the hidden layer at time t; b t is the bias of the hidden layer state at time t.

[0107] In this embodiment, the CNN-BiLSTM branch network model structure diagram is shown in FIG. Figure 3 As shown in the figure, the spatial features of the original data are first extracted through a two-dimensional wide convolutional network, and then the sequence information after dimensionality reduction is sent to BiLSTM for further feature extraction.

[0108] Specifically, Convolutional Neural Network (CNN) is a neural network model widely used in image and video processing tasks. CNN models are usually composed of multiple convolutional layers, activation functions, pooling layers, and fully connected layers, and are trained through the back-propagation algorithm. During the training process, CNN minimizes the loss function by adjusting the weight parameters so that the model can better predict the target. Although CNN has the ability to automatically extract multi-dimensional spatial features from big data, it has poor processing capabilities for time series data with strong time dependence. In contrast, the long short-term memory network LSTM can effectively solve the long-term dependency problem due to the introduction of gated units. By combining the two, the ability to extract spatial features can be enhanced and the calculation time can be relatively reduced.

[0109] The convolution calculation formula is:

[0110]

[0111] In the formula, is the convolution operation; X i is the sequence of convolution calculations; W iis the weight of the convolution kernel; b b is the offset; f(·) is the activation function.

[0112] Although CNN has the ability to automatically extract multi-dimensional spatial features from big data, it has poor processing capabilities for time series data with strong time dependence. In contrast, LSTM can effectively solve the problem of long-term dependence due to the introduction of gated units. By combining the two, the ability to extract spatial features can be enhanced and the calculation time can be relatively reduced. LSTM uses input gates, output gates, and forget gates to control the selective flow of information, solving the problems of gradient disappearance and gradient explosion, and is widely used in time series. The specific calculation formula of the LSTM network is:

[0113] f t =σ(W f *[h t-1 ,x t ]+b f ) (12)

[0114] i t =σ(W i *[h t-1 ,x t ]+b i ) (13)

[0115] g t =tanh(W c *[h t-1 ,x t ]+b c ) (14)

[0116] C t =f t *C t-1 +i t *g t (15)

[0117] o t =σ(W o *[h t-1 ,x t ]+b o ) (16)

[0118] h t =o t *tanh(C t ) (17)

[0119] In the formula, f t is the activation value of the forget gate; σ is the activation function; W f is the weight vector of the forget gate; * indicates convolution operation; h t-1 is the output of the memory unit at time t-1; xt is the input of the memory unit at time t; b f is the bias of the forget gate; i t is the activation value of the input gate; W i is the weight vector of the input gate; b i g is the bias of the input gate input control; t is the candidate state of the memory unit; W c is the weight vector of the candidate state of the input gate; b c is the bias of the candidate state of the input gate; C t is the state of the memory unit at time t; o t is the activation value of the output gate; W o is the weight vector of the output gate; b o is the bias of the output gate. BiLSTM is composed of forward LSTM and backward LSTM.

[0120] Calculate h through forward LSTM 0 ,h 1 ,…,h t , calculate l through backward LSTM 0 ,l 1 ,…,l t , and then concatenate the features of the two directions to get the final output R t :

[0121] R t =[h t ,l t ] (18)

[0122] In the formula, h t for l t The positive output of the moment memory unit; l t is the reverse output of the memory unit at time t.

[0123] This embodiment also adopts a branch network weight distribution model, and merges two parallel branch networks through dynamic optimization to improve the accuracy and stability of the overall prediction model. The weight determination in this embodiment is achieved by an improved sand cat optimization algorithm, with the purpose of obtaining the minimum prediction error.

[0124] 1) Define a weight vector w = [w 1 ,w 2 ],

[0125] where w 1 and w 2 They correspond to the weights of the BiTCN-BiGRU and CNN-BiLSTM branch networks respectively;

[0126] 2) Construct the objective function as:

[0127]

[0128] 3) In order to deal with the constraints of weights, an additional multiplier λ is introduced using the Lagrange multiplier method to ensure that w 1 +w 2 =1 and

[0129] 4) Introduce the quadratic penalty function ψ(w) of the constraint condition to enhance the stability and convergence speed of the optimization process;

[0130]

[0131] 5) Combine the objective function and the penalty function to form a new loss function:

[0132]

[0133] 6) Use dynamic optimization to adjust the weight vector w so that the final loss function is minimized, thereby achieving the best fusion of the two branch networks;

[0134] Where N is the number of samples in the validation set, i is the index of the sample, indicating the i-th sample, and y 1,i is the prediction output of the validation set of the first branch network for the i-th sample, y 2,i is the predicted output of the validation set of the second branch network for the i-th sample; M is 2 to ensure at most two variables, j is the element index in the weight vector, and w j is the element in the weight vector; γ = 0.2 is the penalty parameter, which adjusts the constraint w 1 +w 2 =1 is the penalty intensity when it is violated; λ = 1.1 is the Lagrange multiplier.

[0135] In this embodiment, during the modeling process of the branch network model, the training data also needs to be input into the combined model for multiple trainings. The training process calculates the loss function and back-propagates to update the model parameters. When the number of trainings is greater than the total number of trainings, the model completes the training and determines the optimal hyperparameters of the branch network. During the online prediction process, the test data is input into the trained prediction network model for real-time prediction, and evaluation indicators are set to verify the prediction performance of the model. The evaluation indicators can more effectively reflect the relationship between the model prediction results and the true values. According to the data characteristics of wind power, this embodiment uses the mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) as evaluation indicators. The specific calculation formula is as follows:

[0136]

[0137] In the formula, N is the number of samples, i is the sequence number, is the predicted value at a certain moment, y i is the actual measured value at the same time.

[0138] The improved Sand Cat optimization algorithm dynamically adjusts the key parameters of the number of convolution kernels and residual blocks of the BiTCN-BIGRU model, aiming to improve the model's ability to process time series data and capture more detailed data features by enhancing the depth and complexity of the model; at the same time, the CNN-BiLSTM model carefully adjusts the number of convolution kernels, the stride of the maximum pooling layer, and the number of hidden units of BiLSTM. The purpose of hyperparameter adjustment is to optimize the model's feature extraction and long-term dependency learning capabilities, thereby improving prediction accuracy. By improving the Sand Cat optimization algorithm, the optimal weight combination can be iteratively searched on the validation set to minimize the prediction error. This approach not only significantly improves the prediction accuracy of the model, but also makes full use of the complementarity between different models.

[0139] In order to verify the effectiveness of the proposed model and take into account the efficiency issue, in one embodiment, the 2021 data provided by a wind farm in the northwest was selected for simulation experimental analysis. In this experiment, one month of data was randomly selected as the actual wind power data. The original data was divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0140] Both the BiTCN-BiGRU and CNN-BiLSTMDE models adopt multi-batch training, with the number of batch samples set to 45 and the number of training batches set to 150. The initial learning rate of the model is 0.001, the learning rate scheduling strategy is "piecewise", the verification frequency is 50 small batches, and the learning rate reduction factors of the hidden layer BiTCN-BiGRU model and CNN-BiLSTMDE model are 0.01 and 0.02 respectively. In addition, during the training process, the Adam optimization algorithm is used to optimize the gradient to minimize the mean square error of the model.

[0141] In order to eliminate the gap between different levels of data, the data is normalized according to formula (12):

[0142]

[0143] In order to verify the accuracy of this model in terms of wind power, a comparative analysis method was used to compare this model with the CNN-XGB, TCN-LSTM, LSTM-SVM and LSTM-XGB models to evaluate its performance in wind power forecasting. In order to objectively evaluate the prediction accuracy of the model and avoid the influence of accidental errors, the data set was predicted 10 times to obtain the average value as the final result, which was compared with the true value. The test set results are shown in Table 1.

[0144] Table 1 Model prediction accuracy results

[0145]

[0146]

[0147] Table 1 clearly demonstrates the obvious advantages of the proposed model in terms of prediction accuracy. Compared with the comparative prediction models, the proposed model has achieved significant improvements in the evaluation indicators. This not only highlights the efficiency of the proposed model in the training process, but also highlights its significant ability to reduce prediction errors.

[0148] In order to further prove that the wind power prediction of this model has more accurate prediction results, under the same data environment conditions, the ablation experiment of this model is carried out. The ablation models include the prediction performance of this model without dual decomposition (A), this model without dual decomposition and no weight optimization (B), this model with dual decomposition and no weight optimization (C), this model without dual decomposition and no weight optimization (D), this model with dual decomposition and no weight optimization (E), this model with dual decomposition and no weight optimization (F), and this model with dual decomposition and no weight optimization (G).

[0149] Table 2 Model prediction accuracy results

[0150] MAE MAPE RMSE This model 1.1146 0.0273 1.4694 A 2.3193 0.0618 2.9119 B 4.7016 0.1245 6.1791 C 5.1447 0.1336 5.8787 D 3.7907 0.0957 4.3801 E 3.5791 0.0893 4.1489 F 2.2964 0.0595 3.1361 G 3.7108 0.0914 4.3106

[0151] From the data in Table 2, we can see that this model outperforms other models in all three indicators. Specifically, the MAE of this model is 1.1146, with a maximum value reduction of 4.0301; the MAPE is 0.0273, with a maximum value reduction of 0.1063; and the RMSE is 1.4694, with a maximum value reduction of 4.7097.

[0152] The present invention proposes a combined prediction model optimized by an improved sandcat optimization algorithm, which verifies the effectiveness of the wind power prediction collaborative model through actual cases and improves the accuracy of power prediction.

[0153] 1) A quadratic decomposition is constructed to process the VMD denoised wind power series to reduce its complexity and non-stationarity and improve the accuracy of short-term wind power prediction;

[0154] 2) Construct a meta-learning strategy that uses an improved sandcat optimization algorithm to dynamically adjust the weights of the two-branch network fusion, effectively utilizing the advantages of different models in processing different types of data and enhancing the robustness and accuracy of the model under diverse data conditions.

[0155] 3) In order to solve the problem that a single model cannot utilize future time features and cannot effectively learn long time series features, a parallel BiTCN-BiGRU and CNN-BiLSTM model is proposed, and then the fusion and complementarity of the two model features are achieved through a meta-learning strategy.

[0156] The short-term wind power forecasting method based on quadratic decomposition and improved sandcat algorithm optimization neural network proposed in the present invention can effectively capture the wind power change trend, thereby improving the accuracy of short-term wind power forecasting. Therefore, the method can provide more reliable wind power data for power grid optimization and dispatching.

[0157] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0158] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A short-term wind power forecasting method based on quadratic decomposition and neural network, characterized in that: The following steps are involved: The original wind power signal is obtained for noise reduction and decomposed into multiple initial modal components; Preprocessing the multiple initial modal components to extract corresponding high-resolution feature signals; Inputting the high-resolution feature signal into a pre-trained wind power prediction model; Output the corresponding wind power prediction result; Among them, the wind power prediction model is two parallel branch networks; among them, the first branch network is a bidirectional time domain convolutional network and a bidirectional gated recurrent unit; the second branch network is a two-dimensional convolutional neural network and a bidirectional long short-term memory network; the two parallel branch networks are optimized based on the improved sand cat algorithm.

2. A short-term wind power prediction method based on secondary decomposition and neural network as claimed in claim 1, characterized in that: The preprocessing specifically includes: Applying ensemble empirical mode decomposition to decompose the initial modal components once to obtain multiple subsequences; The subsequence is decomposed twice by using an improved ensemble empirical mode decomposition algorithm to extract high-resolution feature signals.

3. A short-term wind power prediction method based on secondary decomposition and neural network as claimed in claim 1, characterized in that: The first branch network optimizes the number of residual blocks and the spatial loss factor through the improved sand cat optimization algorithm to capture the complexity of time series data; The bidirectional time-domain convolutional network uses multiple residual blocks and dilated convolutions to expand the receptive field of the model, perceive data dependencies in a longer range, and predict dynamic changes in wind power. The output of each time-domain convolutional block is processed by a residual block, and the processed outputs are merged through a deep connection layer, and the merged feature outputs are transmitted to the bidirectional gated recurrent unit layer. The output of the bidirectional gated recurrent unit and the output of each time-domain convolutional block are again subjected to feature fusion through a deep connection layer.

4. A short-term wind power prediction method based on secondary decomposition and neural network as claimed in claim 3, characterized in that: The bidirectional gated recurrent unit uses two gated recurrent unit layers, with the forward gated recurrent unit processing the input sequence from beginning to end, while the backward gated recurrent unit processes in the opposite direction; each gated recurrent unit layer maintains its own hidden state and updates it based on the current input and the hidden state of the previous time step.

5. The short-term wind power prediction method based on quadratic decomposition and neural network according to claim 1, characterized in that: The second branch network adjusts the convolution kernel stride, the maximum pooling layer and the number of hidden units of the bidirectional long short-term memory network through the improved sand cat optimization algorithm to capture spatial features; The spatial features of the original data are extracted through a two-dimensional convolutional neural network, and then the reduced-dimensional sequence information is sent to a bidirectional long short-term memory network for further feature extraction.

6. A short-term wind power prediction method based on quadratic decomposition and neural network as claimed in claim 1, characterized in that: The improved sand cat optimization algorithm specifically includes: 1) Lens imaging reverse learning method: By adjusting the reverse solution, the individual can jump out of the local optimal solution; In two-dimensional space, [a, b] is the range of the solution, where the y-axis is a convex lens, the height of the object is h, its projection on the x-axis is x, and the height of the inverted real image of the object formed by the convex lens is h * , which is x on the x-axis * , the formula is: Let k = h / h*, and get the dynamic new position by adjusting the scaling factor n; Through transformation, the reverse point position in the D-dimensional space is obtained as the latest position of the sand cat individual: Among them, k is the ratio of the object height to the height of the inverted real image of the object, X best(t) is the current location of the sand cat individual, is the latest position of the sand cat individual, t is the current iteration number, and T is the maximum iteration number; 2) Levi's flying and wandering method: In the attack phase of the sand cat optimization algorithm, the sand cat individuals are enabled to switch between short-distance local search and long-distance jumping global search; The update formula for the individual position of the sand cat is: L=X BEST -X best(t) Among them, X BEST is the current optimal solution, RL is the random step coefficient, and L is the offset between the current position and the optimal position; 3) Triangle walking method: While gradually approaching the prey in one direction, the sand cat individuals perform a triangular random walk around the prey, which enables the individuals to move more flexibly during the search process and find a better solution. The update formula for the individual position of the sand cat is: A=L 2 +LP 2 -2*LP*L*cos(2*pi*rand) Among them, LP represents the random step size, represents the A step size when updating the position, cos(2*pi*rand) represents the generated random angle, rand represents a function that generates random numbers, and pi=π.

7. The short-term wind power prediction method based on quadratic decomposition and neural network according to claim 1, characterized in that: The two parallel branch networks are optimized based on the improved sand cat algorithm, including: the two parallel branch networks dynamically adjust the subsequence prediction weights using the improved sand cat algorithm, and merge the two parallel branch networks through dynamic optimization; Specifically include: 1) Set the weight vector w = [w1, w2], Wherein, w1 and w2 correspond to the weights of the first branch network and the second branch network respectively; 2) Construct the objective function as: 3) Use the Lagrange multiplier method to introduce a multiplier λ to handle the weight constraints, w1+w2=1 and 4) Introduce the quadratic penalty function ψ(w) of the constraint condition: 5) Combine the objective function and the penalty function to form a new loss function: 6) Use dynamic optimization to adjust the weight vector w so that the final loss function is minimized, thereby achieving the best fusion of the two branch networks; Where N is the number of samples in the validation set, i is the index of the sample, indicating the i-th sample, and y 1,i is the prediction output of the validation set of the first branch network for the i-th sample, y 2,i is the predicted output of the validation set of the second branch network for the i-th sample; M is 2 to ensure at most two variables, j is the element index in the weight vector, and w j is an element in the weight vector; γ = 0.2 is a penalty parameter that adjusts the penalty intensity when the constraint w1 + w2 = 1 is violated; λ = 1.1 is the Lagrange multiplier.

Images