A short-term wind speed prediction method based on quadratic decomposition and hybrid network

Abstract

The present application relates to a kind of short-term wind speed prediction method based on secondary decomposition and hybrid network, first using the integrated mode decomposition of adaptive noise technique, wind speed historical data is decomposed into multiple components, using sample entropy and heat map for analysis to different components, it is divided into high-frequency component, mid-frequency component, low-frequency component, high correlation component.And then using singular spectrum analysis, high-frequency component is further decomposed into subcomponent.Prediction of subcomponent, mid-frequency component, key component uses the hybrid network proposed, is predicted, to ensure overall prediction accuracy;Low-frequency component is predicted using extreme learning machine, to further improve operation efficiency, finally, all prediction results are superimposed to obtain wind speed prediction result.

Description

Technical Field

[0001] This invention belongs to the field of wind speed prediction technology, specifically relating to a short-term wind speed prediction method based on quadratic decomposition and hybrid networks. Background Technology

[0002] With the development of technology, human demand for energy is increasing. However, the use of traditional fossil fuels, such as coal, oil, and natural gas, releases pollution, damages the environment, and contributes to global warming. Therefore, one of the major challenges facing humanity in the future is how to effectively utilize renewable energy. The development and utilization of renewable energy plays a crucial role in energy security, solving ecological problems, and achieving sustainable human development.

[0003] Wind energy is a type of renewable energy, and its rational utilization can have a positive impact on the economy and daily life. Dispatch, management, and optimization are major challenges for the high penetration of renewable energy, including wind and solar power, in power grids of different sizes. At the same time, power dispatching departments are frequently troubled by wind speed variations, which to some extent restricts the development of wind power. Accurate wind speed forecasting is a good way to solve these problems, and such forecasts will also play a positive role in reducing operating costs and enhancing the competitiveness of wind power.

[0004] Based on different forecast time periods, wind speed forecasts are mainly divided into four categories: ultra-short-term, short-term, medium-term, and long-term. Ultra-short-term forecasts refer to forecasts a few seconds to tens of seconds in advance, helping to track real-time load; short-term forecasts refer to forecasts 30 minutes to 6 hours in advance, helping to plan loads; medium-term forecasts refer to forecasts 6 hours to 1 day in advance, helping to manage the power system; and long-term forecasts refer to forecasts of several days, several weeks, or even longer, helping to plan wind farm construction.

[0005] Accurate wind speed forecasting is crucial for maintaining the normal operation and positive development of power systems. However, wind energy exhibits significant uncertainty and randomness, making wind speed forecasting extremely challenging. Current research utilizes signal decomposition algorithms to break down the original wind speed sequence into multiple components, then forecasts each component, and finally reconstructs the sequence by summing them. However, different components have different frequencies and varying importance. Processing them using the same method may lead to reduced accuracy and wasted computational resources. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a short-term wind speed prediction method based on quadratic decomposition and hybrid networks. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] This invention provides a short-term wind speed prediction method based on quadratic decomposition and hybrid networks, comprising:

[0008] Step 1: Obtain the wind speed sequence, decompose the wind speed sequence to obtain multiple components, and divide the multiple components into high-frequency components, mid-frequency components, low-frequency components and highly correlated components.

[0009] Step 2: Perform secondary decomposition on the high-frequency components to obtain multiple high-frequency sub-components;

[0010] Step 3: Input the multiple high-frequency sub-components, the mid-frequency component, the low-frequency component, and the highly correlated component into the trained prediction model to make predictions and obtain the wind speed prediction results;

[0011] The prediction model includes an extreme learning machine, a hybrid network, and an overlay module. The extreme learning machine is used to predict the low-frequency components to obtain corresponding prediction results. The hybrid network is used to predict the multiple high-frequency sub-components, the mid-frequency components, and the highly correlated components to obtain corresponding prediction results. The overlay module is used to overlay all prediction results to obtain the wind speed prediction result.

[0012] The hybrid network comprises an input layer, a convolutional neural network, an efficient channel attention module, a bidirectional gated recurrent network, a fully connected layer, and an output layer, which are cascaded in sequence.

[0013] In one embodiment of the present invention, step 1 includes:

[0014] Step 1.1: The wind speed sequence is decomposed into multiple components using the CEEMDAN decomposition algorithm;

[0015] Step 1.2: Perform sample entropy analysis on the multiple components to obtain the sample entropy values ​​corresponding to the multiple components. According to the preset sample entropy threshold, divide the multiple components into high-frequency components, mid-frequency components and low-frequency components.

[0016] Step 1.3: Perform correlation analysis on the multiple components to determine the component with the highest correlation to the wind speed sequence. If the component is a low-frequency component, classify it as a highly correlated component; otherwise, keep the component classification result unchanged.

[0017] In one embodiment of the present invention, step 1.1 includes:

[0018] Step 1.1.1: Take the wind speed sequence as the sequence to be decomposed, and add Gaussian white noise to the sequence to be decomposed multiple times to construct multiple corresponding new sequences.

[0019] z i (t)=z(t)+w ρ o i (t), i = 1, 2, ..., I;

[0020] In the formula, z i z(t) represents the i-th new sequence, z(t) represents the wind speed sequence, and o i (t) represents independent Gaussian white noise, w ρ The noise figure is represented by I, which indicates the number of times Gaussian white noise is added.

[0021] Step 1.1.2: Perform EMD decomposition on each new sequence, average the decomposition results of all new sequences to obtain a component, remove the component from the current sequence to be decomposed to obtain the corresponding residual component, and use the residual component as the new sequence to be decomposed;

[0022] Step 1.1.3: After adding Gaussian white noise to the new sequence to be decomposed multiple times, multiple new sequences are obtained. Steps 1.1.2 and 1.1.3 are repeated until the residual components no longer meet the decomposition conditions, resulting in multiple components and residual components.

[0023] In one embodiment of the present invention, step 2 includes: performing a secondary decomposition of the high-frequency components using singular spectrum analysis to obtain multiple high-frequency sub-components.

[0024] In one embodiment of the present invention, the extreme learning machine is a feedforward neural network based on a single hidden layer, including a cascaded input layer, a hidden layer with multiple hidden nodes, and an output layer.

[0025] The output function of the extreme learning machine is expressed as:

[0026]

[0027] In the formula, x represents the input of the Extreme Learning Machine, h represents the activation function, g represents the number of hidden nodes, and ω represents the number of hidden nodes. i Represents the node weight b i β represents the hidden layer bias term of a node. i This represents the output weights between the hidden layer and the output layer.

[0028] In one embodiment of the invention, in the hybrid network,

[0029] The multiple high-frequency sub-components, the mid-frequency components, and the highly correlated components are respectively input to the trained hybrid network through the input layer;

[0030] The convolutional neural network performs deep feature extraction on the input components;

[0031] The high-efficiency channel attention module is used to further extract the deep features to obtain key features;

[0032] The bidirectional gated recurrent network is used to extract time-related features from the features;

[0033] The fully connected layer is used to integrate the deep features, the key features, and the time-related features to obtain the prediction result;

[0034] The output layer outputs the prediction result.

[0035] In one embodiment of the present invention, the activation function of the convolutional neural network is the Selu activation function.

[0036] In one embodiment of the present invention, the output of the bidirectional gated recurrent network is represented as:

[0037]

[0038]

[0039]

[0040] Among them, h t W represents the output of the bidirectional gated recurrent network. t α W represents the forward propagation weights. t β P represents the backpropagation weights. t Indicates the bias amount. This represents the output of the hidden layer during forward propagation at time t. This represents the output of the backpropagation hidden layer at time t, where GRU stands for Gated Recurrent Unit, and x t This represents the input of a bidirectional gated recurrent network. This represents the output of the hidden layer during backpropagation at time t-1. This represents the output of the hidden layer during forward propagation at time t-1.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] The short-term wind speed prediction method based on quadratic decomposition and hybrid networks of the present invention preprocesses data using a combination of CEEMDAN and SSA. The processed components are then predicted using hybrid networks CEB and ELM. The present invention rationally classifies the components generated by decomposition and uses different methods to predict different categories of components. ELM reduces the prediction time for simple components, while CEB improves the prediction accuracy of complex components. This method helps to reduce the consumption of computing resources while improving prediction accuracy.

[0043] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of a short-term wind speed prediction method based on quadratic decomposition and hybrid network provided in an embodiment of the present invention;

[0045] Figure 2 This is a flowchart of a short-term wind speed prediction method based on quadratic decomposition and hybrid network provided by an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of an extreme learning machine model provided in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of a hybrid network structure provided in an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram of the framework structure of a high-efficiency channel attention module provided in an embodiment of the present invention;

[0049] Figure 6 This is a schematic diagram of a GRU structure provided in an embodiment of the present invention;

[0050] Figure 7 This is a schematic diagram of the structure of a bidirectional gated loop network provided in an embodiment of the present invention;

[0051] Figure 8 This is a turbine distribution diagram of a wind farm provided in an embodiment of the present invention;

[0052] Figure 9 This is a schematic diagram of wind speed data provided in an embodiment of the present invention;

[0053] Figure 10 This is a schematic diagram of the CEEMDAN decomposition results of wind speed data provided in an embodiment of the present invention;

[0054] Figure 11 This is a visual diagram of the sample entropy of all components of the wind speed data provided in this embodiment of the invention after CEEMDAN decomposition.

[0055] Figure 12 This is a schematic diagram of the correlation matrix provided in an embodiment of the present invention. Detailed Implementation

[0056] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail a short-term wind speed prediction method based on quadratic decomposition and hybrid networks proposed according to the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0057] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0058] Example 1

[0059] Please refer to the above. Figure 1 and Figure 2 , Figure 1 This is a schematic diagram of a short-term wind speed prediction method based on quadratic decomposition and hybrid network provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a short-term wind speed prediction method based on quadratic decomposition and hybrid networks, provided by an embodiment of the invention. As shown in the figure, the short-term wind speed prediction method based on quadratic decomposition and hybrid networks in this embodiment includes:

[0060] Step 1: Obtain the wind speed sequence, decompose the wind speed sequence to obtain multiple components, and divide the multiple components into high-frequency components, mid-frequency components, low-frequency components and highly correlated components.

[0061] In one optional implementation, step 1 includes:

[0062] Step 1.1: The wind speed sequence is decomposed into multiple components using the CEEMDAN decomposition algorithm;

[0063] CEEMDAN is an improvement upon the EMD signal decomposition algorithm. The EMD algorithm decomposes the original sequence into several IMF (Intrinsic Mode Function) components. Different frequency IMF components represent corresponding local features of the signal. The signal decomposition is based on the time axis and has a certain degree of adaptability. However, EMD also has drawbacks, such as mode aliasing and endpoint effects caused by dataset limitations. The subsequent EEMD algorithm addresses this by adding Gaussian white noise to influence extreme point characteristics, but some white noise cannot be removed, potentially leading to clutter and hindering the acquisition of useful feature information. To minimize residual white noise, EEMD requires increasing the number of processing iterations, which is computationally wasteful. CEEMD was then proposed to solve this problem. Its idea is to add a pair of opposite positive and negative white noises as auxiliary noise to the source signal. While one sequence receives positive noise, the other receives negative noise. The reconstructed signals cancel each other out, eliminating redundant white noise and reducing computational cost. To further improve CEEMD, CEEMDAN was proposed, which changed the original white noise addition method and decomposition process. CEEMDAN has better decomposition effect, is more conducive to model learning, and is also faster.

[0064] In an optional implementation, step 1.1 includes:

[0065] Step 1.1.1: Use the wind speed sequence as the sequence to be decomposed, and add Gaussian white noise to the sequence multiple times to construct multiple corresponding new sequences.

[0066] z i (t)=z(t)+w ρ o i (t), i = 1, 2, ..., I(1);

[0067] In the formula, z i z(t) represents the i-th new sequence, z(t) represents the wind speed sequence, and o i (t) represents independent Gaussian white noise, w ρ The noise figure is represented by I, which indicates the number of times Gaussian white noise is added.

[0068] Step 1.1.2: Perform EMD decomposition on each new sequence, average the decomposition results of all new sequences to obtain a component, remove the component from the current sequence to be decomposed to obtain the corresponding residual component, and use the residual component as the new sequence to be decomposed;

[0069] For example, taking a wind speed sequence as the sequence to be decomposed, multiple new sequences z are constructed by adding Gaussian white noise to it multiple times. iFor each new sequence, perform EMD decomposition, average the decomposition results for all new sequences to obtain the first component, and remove the first component from the wind speed sequence to obtain the corresponding first residual component:

[0070]

[0071] res1(t)=z(t)-IMF1(t)(3);

[0072] In the formula, Represents the new sequence z i (t) is the decomposition result obtained after EMD decomposition, where IMF1(t) represents the first component and res1(t) represents the first residual component.

[0073] Step 1.1.3: After adding Gaussian white noise to the new sequence to be decomposed multiple times, multiple new sequences are obtained. Repeat steps 1.1.2 and 1.1.3 until the residual component no longer meets the decomposition conditions, and multiple components and residual components are obtained.

[0074] In this embodiment, noise is added to the residual components of the p-th stage (p∈{2,3,…I}) obtained after decomposition, and EMD decomposition is continued. The resulting decomposition and residual components are represented as follows:

[0075]

[0076] res p (t)=res p-1 (t)-IMF p (t) (5);

[0077] In the formula, E1 represents EMD decomposition, E p-1 This represents the (p-1)th component obtained from EMD decomposition.

[0078] In this embodiment, the residual component does not meet the decomposition condition, that is, the number of extreme points of the residual component signal does not exceed 2. Finally, the result of the wind speed sequence after CEEMDAN decomposition can be expressed as:

[0079]

[0080] In the formula, IMF n This represents the multiple components obtained from the decomposition, and RES(t) represents the residual term.

[0081] Step 1.2: Perform sample entropy analysis on multiple components to obtain the sample entropy values ​​corresponding to multiple components. Based on the preset sample entropy threshold, divide the multiple components into high-frequency components, mid-frequency components, and low-frequency components.

[0082] For the multiple components obtained from the decomposition, the higher the frequency and the higher the complexity, the weaker the smoothness of the corresponding component graph. To more intuitively represent the complexity of different IMF components, sample entropy is used as a measure. The smaller the sample entropy value, the stronger the self-similarity of the sequence; the larger the sample entropy value, the more complex the sequence. Moreover, sample entropy does not depend on the data length.

[0083] For example, in the embodiment, the reconstruction dimension of the sample entropy is m=2, the threshold is r=0.1, and it is defined that when m=2 and r=0.1, the sample entropy value of the IMF component is >=1, which is a high frequency component; the sample entropy value is greater than 0.05 and less than 1, which is a mid frequency component; and the sample entropy value is less than 0.05, which is a low frequency component.

[0084] Step 1.3: Perform correlation analysis on multiple components to determine the component with the highest correlation to the wind speed sequence. If the component is a low-frequency component, classify it as a highly correlated component; otherwise, keep the component classification result unchanged.

[0085] Step 2: Perform secondary decomposition on the high-frequency components to obtain multiple high-frequency sub-components;

[0086] In an optional implementation, step 2 includes: performing a secondary decomposition of the high-frequency components using singular spectral analysis (SSA) to obtain multiple high-frequency sub-components.

[0087] The specific implementation steps of SSA are as follows:

[0088] Step a: For a high-frequency component x = {x1, x2, ..., x...} N The high-frequency components are then arranged in a hysteresis manner using a window of length L to obtain the corresponding trajectory matrix X.

[0089]

[0090] Where K = N - L + 1, K represents the number of lag vectors, and N represents the sequence length of the high-frequency components;

[0091] Step b: Perform singular value decomposition (SVD) on the trajectory matrix X;

[0092] The trajectory matrix X can be written as X = X1 + X2 + ... + X d Singular value decomposition is performed on each submatrix, and the result is expressed as: U i and V i These are the eigenvectors corresponding to the left and right matrices. It is a singular value.

[0093] Step c: Equate X = X1 + X2 + ... + X dThe subscript set {1,..., d} is partitioned into m non - overlapping subsets I1,..., I m , and let I = {i1,..., i p} corresponding matrix Then further, the trajectory matrix X is expressed as

[0094] Step d: The grouped matrix pairs are averaged through the diagonal to transform and obtain a new sequence representation as follows:

[0095]

[0096] where, x ij represents the element in the matrix, 1 ≤ i ≤ L, 1 ≤ j ≤ K. Assume L * = min{L, K} and K * = max{L, K}. When L < K, Otherwise

[0097] Step 3: Input multiple high - frequency sub - components, intermediate - frequency components, low - frequency components and high - correlation components into the trained prediction model pair for prediction to obtain the wind speed prediction result;

[0098] In an optional implementation manner, the prediction model includes an extreme learning machine, a hybrid network and a stacking module.

[0099] Among them, in this embodiment, the extreme learning machine is used to predict the low - frequency component to obtain the corresponding prediction result;

[0100] The extreme learning machine (ELM) reduces complex iterations compared with deep learning, has a very fast training speed, and at the same time has good generalization. Please refer to the schematic diagram of an extreme learning machine model provided in the embodiment of the present invention as shown in Figure 3 . In this embodiment, the extreme learning machine is a single - hidden - layer feed - forward neural network (SLFN), including a cascaded input layer, a hidden layer with multiple hidden nodes and an output layer.

[0101] Among them, the hidden layer is expressed as h i (x)= h(ω i x + b i ), where, ω i represents the node weight, b i represents the node hidden - layer bias term, h represents the activation function, then the output function of the extreme learning machine is expressed as:

[0102]

[0103] In the formula, x represents the input of the extreme learning machine, g represents the number of hidden nodes, β iThis represents the output weights between the hidden layer and the output layer.

[0104] Where, β i The solution method involves minimizing the approximate squared difference, and the objective function is:

[0105] argmin||YH(x)β||2 (10);

[0106] Where H(x) is the output matrix of the hidden layer, and Y is the target matrix of the training data. Further derivation yields β. * =H T (HH T ) -1 ,β * This represents the output weights, but finding the inverse of a high-dimensional matrix is ​​quite complex, so we use a generalized inverse matrix to represent it, let... The generalized inverse matrix of H is ultimately calculated as follows: T represents the network computation output.

[0107] It should be noted that since low-frequency component prediction is relatively easy, ELM is used for prediction in this embodiment, which does not require iteration and is faster, thus further improving computational efficiency and reducing computation time.

[0108] In this embodiment, a hybrid network is used to predict multiple high-frequency sub-components, mid-frequency components, and highly correlated components to obtain corresponding prediction results.

[0109] In this embodiment, the hybrid network (CEB) includes a cascaded input layer, a convolutional neural network (CNN), an efficient channel attention module (ECA), a bidirectional gated recurrent network (BIGRU), a fully connected layer, and an output layer, as follows: Figure 4 The diagram shown is a schematic representation of a hybrid network provided by an embodiment of the present invention.

[0110] In one optional implementation, multiple high-frequency sub-components, mid-frequency components, and highly correlated components are respectively input into the trained hybrid network through the input layer; the convolutional neural network performs deep feature extraction on the input components; the efficient channel attention module is used to further extract the deep features to obtain key features; the bidirectional gated recurrent network is used to extract the temporal correlation features from the features; the fully connected layer is used to integrate the deep features, key features, and temporal correlation features to obtain the prediction result; and the output layer outputs the prediction result.

[0111] In this embodiment, the activation function of the convolutional neural network (CNN) is the Selu activation function. Choosing the extended exponential linear unit activation function, i.e., the Selu activation function, avoids the neuron death problem present in the ReLU activation function, while also giving the network self-normalization properties.

[0112] The Efficient Channel Attention (ECA) module is designed to optimize the SENet (Squeeze and Activate Network) model. ECA has fewer parameters than SENet, making it more lightweight. The absence of dimensionality-reducing cross-channel interactions makes ECA more efficient and performs better. The framework structure of ECA is as follows: Figure 5 The diagram shown is a framework structure diagram of an efficient channel attention module provided by an embodiment of the present invention, wherein GAP is global average pooling and ψ is a function for adaptively selecting the kernel size.

[0113] In time series forecasting, Recurrent Neural Networks (RNNs) are frequently used due to their good performance in handling nonlinear problems. However, RNNs perform poorly when dealing with long-range information and suffer from gradient vanishing. The introduction of Long Short-Term Memory (LSTM) networks addressed these issues to some extent. The Gated Recurrent Network (GRU) model was then proposed to improve upon some of the problems of LSTM. This model reduces the three gates of the original LSTM (forget gate, input gate, and output gate) to two gates (update gate and reset gate), resulting in a simpler structure, fewer training parameters, faster fitting speed, and shorter model training time. A schematic diagram of GRU is shown below. Figure 6 As shown, the formula is as follows:

[0114] z t =σ(W z ·[h t-1 ,x t ]) (11);

[0115] r t =σ(W r ·[h t-1 ,x t ]) (12);

[0116]

[0117] Among them, h t-1 It is the previous state. It is a candidate state, z t It's an update gate; the function of an update gate is to control the current h. t From h t-1 How much information to retain, and from How much new information is received, r t The decision to reset the door was made Do I need to receive h? t-1 Information, where σ is the sigmoid function.

[0118] A conventional GRU operates on a one-way propagation path. However, both past and future wind speed data influence prediction results. Comprehensive information extraction further improves prediction accuracy. This embodiment employs a bidirectional gated recurrent network (BIGRU), with the BIGRU framework structure as follows: Figure 7 As shown, the output of the bidirectional gated recurrent network is expressed as:

[0119]

[0120]

[0121]

[0122] Among them, h t W represents the output of the bidirectional gated recurrent network. t α W represents the forward propagation weights. t β P represents the backpropagation weights. t Indicates the bias amount. This represents the output of the hidden layer during forward propagation at time t. This represents the output of the backpropagation hidden layer at time t, where GRU stands for Gated Recurrent Unit, and x t This represents the input of a bidirectional gated recurrent network. This represents the output of the hidden layer during backpropagation at time t-1. This represents the output of the hidden layer during forward propagation at time t-1.

[0123] In this embodiment, the CNN in CEB is used to extract deep features and can better capture local features, enriching the feature set. An efficient channel attention module (ECA) is introduced to extract key features from the existing features, reducing historical information loss, improving overall performance, and maintaining low complexity. BIGRU is combined to further extract temporal correlations from the features. The Selu activation function is chosen to avoid the neuron death problem associated with ReLU activation functions, while also giving the network self-normalization properties.

[0124] It should be noted that the mid-frequency components, highly correlated components, and sub-components generated by the secondary decomposition of high-frequency components are predicted using the proposed hybrid network (CEB), which can ensure the overall prediction accuracy.

[0125] In this embodiment, the training of the prediction model adopts existing training methods, which will not be described in detail here.

[0126] In this embodiment, an overlay module is used to overlay all the prediction results of the Extreme Learning Machine and the Hybrid Network to obtain the wind speed prediction result.

[0127] This embodiment of the short-term wind speed prediction method based on quadratic decomposition and hybrid networks preprocesses data using a combination of CEEMDAN and SSA. The processed components are then predicted using hybrid networks CEB and ELM. This invention rationally classifies the components generated by decomposition and uses different methods to predict different categories of components. ELM reduces the prediction time for simple components, while CEB improves the prediction accuracy for complex components. This approach helps reduce computational resource consumption while improving prediction accuracy.

[0128] Example 2

[0129] This embodiment illustrates the short-term wind speed prediction method based on quadratic decomposition and hybrid networks described in Embodiment 1 through specific experiments.

[0130] 1. Selection of experimental data

[0131] The experimental data comes from the SOTAVENTO wind farm, located in southwestern Europe, in Galicia, Spain, 10 km south of the García Rodriguez Bridge. It is a 600-700 meter above-seat level. The wind farm's coordinates are N: 43,354,377°, W: 7,881,213°. The turbine distribution of the wind farm is as follows... Figure 8 As shown. The experimental data is the built-in SCADA data of the wind turbine. SCADA stands for Supervisory Control and Data Acquisition. The SOTAVENTO wind farm SCADA system consists of SCADA components for different wind turbines and the SCADA master equipment at the server site. The specific data is divided into three parts: Powergenerated (10 minutes), Wind Speed ​​(10 minutes), and Cumulative generated.

[0132] The experiment used historical wind speed data from August 2021, with a time resolution of 10 minutes. Time series prediction requires determining the stationarity of the data. If the series is non-stationary, it means that the current characteristics of the series cannot be carried over to the future, and therefore the historical series cannot be used to predict the future. Therefore, further testing of the stationarity of the original series is necessary. The Augmented Dickey-Fuller (ADF) test was used for verification. The ADF test results are shown in Table 2. Observation shows that the T-statistic value of the series is -6.1813, which is less than 1% level -3.4318, 5% level -2.8621, and 10% level -2.567115, and the p-value of 6.4442e-08 is less than 0.05 and close to 0, rigorously proving that the original series is stationary and can be used for time series prediction. Furthermore, the data described in Table 1 and... Figure 9 Visualization clearly shows that the wind speed sequence fluctuates greatly, is non-linear, contains a lot of noise, and is difficult to predict linearly using simple models.

[0133] Table 1. Wind speed sequence description

[0134]

[0135] Table 2. ADF test results for wind speed series

[0136]

[0137]

[0138] 2. Evaluation Indicators

[0139] In this embodiment, the evaluation metrics used to compare the prediction results include: root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). RMSE and MAE represent the error between the predicted and actual values, while MAPE is the percentage absolute error. The smaller the values ​​of RMSE, MAE, and MAPE, the better the model performance and the more accurate the prediction results.

[0140] 3. Experimental Results and Analysis

[0141] All data were normalized before experiments, with the training, validation, and test sets in a 7:2:1 ratio. To ensure the stability of the prediction results, all models were tested 5 times. Each test had a maximum of 100 epochs and a batch size of 64. An early stopping mechanism was employed to halt training before accuracy decreased, effectively preventing overfitting.

[0142] 3.1 Comparison of different activation functions

[0143] Different activation functions can affect the final prediction results for the same network. Four activation functions were selected to test the proposed CNN-ECA-BIGRU (CEB) hybrid network. All four activation functions were applied to the CNN stage: Tanh, ReLU, Leaky ReLU, and SeLU. The results are shown in Table 3. It can be seen that SeLU showed the best results compared to the other activation functions. Therefore, CEB uses SeLU as the activation function on the CNN for the final prediction.

[0144] Table 3. Prediction results for different activation functions

[0145]

[0146] 3.2 Ablation Experiment of Model Components

[0147] For the proposed hybrid network, an ablation study is required. The core of the ablation study is to control variables and demonstrate that each component of the hybrid network plays a role in the final result by deleting and combining different components of the proposed hybrid network.

[0148] The experimental procedure first decomposed each component of the hybrid network to demonstrate the effectiveness of CNN, BIGRU, and ECA, and then combined different components. The final comparison network models included CNN, BIGRU, CNN-BIGRU, CNN-ECA, and CNN-ECA-BIGRU. The CNNs all used two layers: the first layer had 64 convolutional kernels of size 1, and the second layer had 32 convolutional kernels of size 1. The Selu activation function was used in all layers. The BIGRU layer had 16 neurons. Specific parameters are shown in Table 4.

[0149] Table 4. Introduction to Ablation Experimental Model Parameters

[0150]

[0151] Table 5 shows the experimental results for different network models. It can be seen that the RNN variant BIGRU is better at handling time series data than CNN, and it is stronger at extracting time series features. However, CNN is more computationally efficient and iterates faster than RNN. Stacking RNN layers may improve training results, but it will significantly increase training time. Therefore, to ensure training speed, the number of stacked RNNs in the model must be reduced. Combining CNN with BIGRU can fully utilize the characteristics of both neural networks, making the network more capable of extracting and representing features. Adding ECA to individual components at the end results in a slight improvement in the final prediction results, indicating that ECA has strong applicability. The ablation experiment results show that each component plays a role, with CEB performing well.

[0152] Table 5. Comparison of predictive performance of all models in ablation experiments

[0153]

[0154] 3.3. Prediction after decomposition

[0155] Based on the analysis of the experimental data, the wind speed data is nonlinear and fluctuates significantly. Therefore, data processing is crucial for improving prediction accuracy. This invention employs a signal decomposition method for preprocessing. First, the data undergoes Ceemdan decomposition, and the Ceemdan decomposition results are as follows: Figure 10 As shown.

[0156] The first data set is the original wind speed data, followed by different IMFs (Integrated Wind Factors), from IMF0 to the residuals (res). The complexity and frequency decrease from IMF0 to res, with increasing smoothness. To more intuitively represent the complexity of different IMFs, sample entropy is used. A smaller sample entropy value indicates stronger sequence self-similarity, while a larger value indicates a more complex sequence. Sample entropy is independent of data length. The reconstruction dimension m of the sample entropy is set to 2, and the threshold r is set to 0.1 and 0.2 respectively. The sample entropy results for all components are shown in Figures 6 and 7. Figure 11 As shown, Figure 11 Visualize the sample entropy of different modal components.

[0157] Table 6. Sample Entropy

[0158]

[0159] In this embodiment, when m=2 and r=0.1, the sample entropy value of the IMF is defined as >=1, which is a high-frequency component; the sample entropy value is greater than 0.05 and less than 1, which is a mid-frequency component; and the sample entropy value is less than 0.05, which is a low-frequency component. Therefore, IMF0 and IMF1 are defined as high-frequency components, IMF2-IMF6 are defined as mid-frequency components, and IMF7-IMF9 are defined as low-frequency components.

[0160] To further understand the importance of different IMFs to forecast results, a correlation analysis was conducted on the IMFs. The results showed that... Figure 12 The correlation matrix reveals that IMF 7 has the strongest correlation with the original wind speed sequence, thus classifying IMF 7 from low-frequency components to high-correlation components. For high-frequency components, due to their highest complexity, SSA is used for secondary decomposition, breaking them down into multiple high-frequency sub-components. IMF 7, along with the high-frequency and mid-frequency sub-components, is trained using CEB. Low-frequency components, with the lowest complexity, are predicted using ELM to reduce runtime.

[0161] To verify the improvement in prediction performance brought about by the proposed hybrid network, the following four models were used in the comparative experiment.

[0162] Model 1 (CEEMDAN-CEB): After decomposing the original sequence using CEEMDAN, CEB is used for prediction.

[0163] Model 2 (CEEMDAN-CEB-ELM): After CEEMDAN decomposes the original sequence, low-frequency components are predicted using ELM; high-frequency, mid-frequency, and highly correlated components are predicted using CEB.

[0164] Model 3 (CEEMDAN-SSA-CEB): After CEEMDAN decomposes the original sequence, the high-frequency components are decomposed a second time using SSA. The resulting sub-components and the remaining undecomposed components are all predicted using CEB.

[0165] Model 4 (CEEMDAN-SSA-CEB-ELM): After CEEMDAN decomposes the original sequence, low-frequency components are predicted using ELM. High-frequency components are decomposed a second time using SSA to generate sub-components. Sub-components, high-frequency components, mid-frequency components, and highly correlated components are all predicted using CEB.

[0166] The prediction results are shown in Table 7. All four models show significant improvements in accuracy compared to models without CEEMDAN decomposition. The models using ELM, without sacrificing accuracy, have reduced running time. Compared to models 1 and 2, models 3 and 4 demonstrate that secondary decomposition effectively reduces the prediction difficulty of high-frequency components, further improving prediction accuracy.

[0167] Table 7. Evaluation of Model Prediction Results

[0168]

[0169] To more objectively compare the predictive capabilities of the method proposed in this invention, it was also compared with other benchmark models, including ELM, BPNN, BILSTM, CNN-BILSTM, CEEMDAN-VMD-BIGRU, and CEEMDAN-VMD-CEB. The experimental results are shown in Table 8. It can be seen that the prediction method proposed in this invention has a higher prediction accuracy than other benchmark models.

[0170] Table 8. Evaluation of Model Prediction Results

[0171]

[0172]

[0173] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0174] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A short-term wind speed prediction method based on quadratic decomposition and hybrid network, characterized in that, include: Step 1: Obtain the wind speed sequence, decompose the wind speed sequence to obtain multiple components, and divide the multiple components into high-frequency components, mid-frequency components, low-frequency components and highly correlated components. Step 2: Perform secondary decomposition on the high-frequency components to obtain multiple high-frequency sub-components; Step 3: Input the multiple high-frequency sub-components, the mid-frequency component, the low-frequency component, and the highly correlated component into the trained prediction model to make predictions and obtain the wind speed prediction results; The prediction model includes an extreme learning machine, a hybrid network, and an overlay module. The extreme learning machine is used to predict the low-frequency components to obtain corresponding prediction results. The hybrid network is used to predict the multiple high-frequency sub-components, the mid-frequency components, and the highly correlated components to obtain corresponding prediction results. The overlay module is used to overlay all prediction results to obtain the wind speed prediction result. The hybrid network comprises an input layer, a convolutional neural network, an efficient channel attention module, a bidirectional gated recurrent network, a fully connected layer, and an output layer, which are cascaded in sequence.

2. The short-term wind speed prediction method based on quadratic decomposition and hybrid networks according to claim 1, characterized in that, Step 1 includes: Step 1.1: The wind speed sequence is decomposed into multiple components using the CEEMDAN decomposition algorithm; Step 1.2: Perform sample entropy analysis on the multiple components to obtain the sample entropy values ​​corresponding to the multiple components. According to the preset sample entropy threshold, divide the multiple components into high-frequency components, mid-frequency components and low-frequency components. Step 1.3: Perform correlation analysis on the multiple components to determine the component with the highest correlation to the wind speed sequence. If the component is a low-frequency component, classify it as a highly correlated component; otherwise, keep the component classification result unchanged.

3. The short-term wind speed prediction method based on quadratic decomposition and hybrid networks according to claim 2, characterized in that, Step 1.1 includes: Step 1.1.1: Take the wind speed sequence as the sequence to be decomposed, and add Gaussian white noise to the sequence to be decomposed multiple times to construct multiple corresponding new sequences. z i (t)=z(t)+w ρ o i (t),i=1,2,...,I; In the formula, z i z(t) represents the i-th new sequence, z(t) represents the wind speed sequence, and o i (t) represents independent Gaussian white noise, w ρ The noise figure is represented by I, which indicates the number of times Gaussian white noise is added. Step 1.1.2: Perform EMD decomposition on each new sequence, average the decomposition results of all new sequences to obtain a component, remove the component from the current sequence to be decomposed to obtain the corresponding residual component, and use the residual component as the new sequence to be decomposed; Step 1.1.3: After adding Gaussian white noise to the new sequence to be decomposed multiple times, multiple new sequences are obtained. Steps 1.1.2 and 1.1.3 are repeated until the residual components no longer meet the decomposition conditions, resulting in multiple components and residual components.

4. The short-term wind speed prediction method based on quadratic decomposition and hybrid networks according to claim 1, characterized in that, Step 2 includes: performing a secondary decomposition of the high-frequency components using singular spectrum analysis to obtain multiple high-frequency sub-components.

5. The short-term wind speed prediction method based on quadratic decomposition and hybrid networks according to claim 1, characterized in that, The extreme learning machine is a feedforward neural network based on a single hidden layer, including a cascaded input layer, a hidden layer with multiple hidden nodes, and an output layer; The output function of the extreme learning machine is expressed as: In the formula, x represents the input of the Extreme Learning Machine, h represents the activation function, g represents the number of hidden nodes, and ω represents the number of hidden nodes. i Represents the node weight b i β represents the hidden layer bias term of a node. i This represents the output weights between the hidden layer and the output layer.

6. The short-term wind speed prediction method based on quadratic decomposition and hybrid networks according to claim 1, characterized in that, In the hybrid network, The multiple high-frequency sub-components, the mid-frequency components, and the highly correlated components are respectively input to the trained hybrid network through the input layer; The convolutional neural network performs deep feature extraction on the input components; The high-efficiency channel attention module is used to further extract the deep features to obtain key features; The bidirectional gated recurrent network is used to extract time-related features from the features; The fully connected layer is used to integrate the deep features, the key features, and the time-related features to obtain the prediction result; The output layer outputs the prediction result.

7. The short-term wind speed prediction method based on quadratic decomposition and hybrid networks according to claim 6, characterized in that, The activation function of the convolutional neural network is the Selu activation function.

8. The short-term wind speed prediction method based on quadratic decomposition and hybrid networks according to claim 6, characterized in that, The output of the bidirectional gated cyclic network is represented as follows: Among them, h t W represents the output of the bidirectional gated recurrent network. t α W represents the forward propagation weights. t β P represents the backpropagation weights. t Indicates the bias amount. This represents the output of the hidden layer during forward propagation at time t. This represents the output of the backpropagation hidden layer at time t, where GRU stands for Gated Recurrent Unit, and x t This represents the input of a bidirectional gated recurrent network. This represents the output of the hidden layer during backpropagation at time t-1. This represents the output of the hidden layer during forward propagation at time t-1.

Images