Optimal configuration method of wind-storage hybrid energy storage system based on quadratic empirical mode decomposition and multi-feature fusion

Abstract

The invention discloses a wind-storage hybrid energy storage system optimization configuration method based on quadratic empirical mode decomposition and multi-feature fusion, and the method comprises the following steps: S1, carrying out the prediction of wind power output through employing a discrete wavelet decomposition-Transformer-bidirectional LSTM multi-feature composite prediction wind power output method, and generating a wind power prediction output result; s2, performing k-means clustering analysis on the wind power prediction output result and historical wind power output data to obtain wind power output typical day data, and performing empirical mode decomposition on the wind power output typical day data twice; s3, performing empirical mode decomposition twice to obtain a low-frequency component and a high-frequency component; obtaining rated power of a storage battery and a super capacitor in the hybrid energy storage system; and S4, carrying out optimal configuration on the capacities of the storage battery and the supercapacitor. And secondary empirical mode decomposition and reconstruction are carried out on the wind-storage hybrid energy storage system after primary reconstruction, so that the difficulty of hybrid energy storage capacity configuration is reduced, the economy and the operation stability in the face of extreme conditions are improved, and the wind power configuration cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of coordinated optimization of renewable energy power generation and energy storage systems, and in particular to a coordinated configuration optimization method for a wind-storage hybrid energy storage system based on wind power output prediction and quadratic empirical mode decomposition. Background Art

[0002] At present, the mainstream methods in the field of wind speed prediction mainly use deviation correction or averaging the prediction set to obtain the final prediction results. However, these traditional methods have obvious limitations in practical applications: on the one hand, the prediction results are easily affected by outliers in the prediction set; on the other hand, it is difficult to effectively capture the local characteristic changes during the rapid fluctuation of wind power output. In terms of wind-storage system capacity configuration, the conventional practice is usually to decompose and reconstruct the wind power signal once, and only meet the grid connection requirements through the low-frequency components, while the hybrid energy storage system absorbs the high-frequency components. However, this configuration scheme based on a single set of empirical mode decomposition faces many challenges in actual engineering applications, including high complexity of the configuration process, insufficient operational stability of the system under extreme working conditions, and poor overall economic efficiency.

[0003] Patent CN112668807A proposes a wind speed prediction method based on deviation data calculation, which obtains the final predicted value of wind speed through a basic model of wind speed prediction deviation value and a deviation correction model of wind speed prediction deviation value.

[0004] Patent CN118982447B proposes a wind speed prediction method based on secondary decomposition and reconstruction. By quantifying the signal's complexity, the method groups and decomposes the signal to produce a wind speed prediction set. The final prediction value is then obtained by averaging the wind power output within the prediction set.

[0005] Patent CN118353073B proposes a method and device for optimizing the configuration of new energy storage based on empirical mode decomposition, which uses an improved cuckoo algorithm to obtain an optimized configuration scheme for energy storage of new energy.

[0006] Patent CN112581172A proposes a multi-model fusion method for electricity sales forecasting based on empirical mode decomposition, combining a random forest model to predict high-, medium-, and low-frequency series. By using the fusion model, the accuracy and stability of the forecast are improved.

[0007] Patent CN109146186A proposes a short-term wind power forecasting method based on double decomposition, which obtains the final wind power forecast result by superimposing the forecast models.

[0008] Patent CN108521133A proposes a frequency regulation power allocation method for joint dispatching of thermal storage based on ensemble empirical mode decomposition. By establishing an optimization model for the segmentation points of the frequency regulation signal of the thermal-storage joint system, the frequency regulation power value allocated to the thermal storage is determined, thereby improving the frequency regulation performance of the power system while increasing economic benefits.

[0009] While patents CN118353073B, CN112581172A, CN109146186A, and CN108521133A all address wind speed prediction and energy storage configuration in different areas, they all suffer from limitations such as high wind power configuration costs and insufficient system stability under extreme operating conditions. Therefore, there is an urgent need for an optimized configuration method for wind-storage hybrid energy storage systems that can effectively address these issues, enabling accurate assessment of wind power output and efficient configuration of energy storage systems. Summary of the Invention

[0010] The present invention proposes a wind-storage hybrid energy storage system optimization configuration method based on quadratic empirical mode decomposition and multi-feature fusion to solve the problems existing in the prior art, such as inaccurate wind speed prediction, high complexity of wind-storage system capacity configuration, insufficient operational stability, and poor economy.

[0011] The technical solution of the present invention is as follows: a method for optimizing the configuration of a wind-storage hybrid energy storage system based on quadratic empirical mode decomposition and multi-feature fusion, wherein the energy storage system includes batteries and supercapacitors, and the method comprises the following steps:

[0012] S1: Based on historical wind power output data, a multi-feature composite prediction method based on discrete wavelet decomposition, Transformer, and bidirectional LSTM is used to predict wind power output and generate wind power forecast results.

[0013] S2: Perform k-means cluster analysis on the wind power forecast results and historical wind power output data to obtain typical daily wind power output data, perform a primary empirical mode decomposition and low-frequency reconstruction of multiple intrinsic mode components on the typical daily wind power output data; calculate the signal fluctuation of the wind power low-frequency reconstructed component within a preset time period, and determine whether the maximum signal difference within the preset time after the low-frequency reconstruction meets the fluctuation limit; if so, it is determined to be a low-frequency component; if not, it is determined to be a high-frequency component, and the high-frequency component obtained by the primary empirical mode decomposition is subjected to a secondary empirical mode decomposition;

[0014] S3: Performing a second empirical mode decomposition on the high-frequency component obtained by the first empirical mode decomposition in step S2 to obtain a low-frequency component and a high-frequency component; absorbing the obtained low-frequency component by the lithium battery and absorbing the high-frequency component by the supercapacitor, thereby obtaining the rated power of the battery and the supercapacitor in the hybrid energy storage system;

[0015] S4: Optimize the capacity configuration of batteries and supercapacitors, with the goal of minimizing the total cost of the battery and supercapacitor hybrid energy storage system and the coupling relationship of each device as a constraint, and optimize the power / capacity configuration results of lithium batteries and supercapacitors based on Matlab / PSO.

[0016] Preferably, the multi-feature composite wind power output prediction method using discrete wavelet decomposition-Transformer-bidirectional LSTM in step S1 includes the following steps:

[0017] S11: Use MinMaxScaler to normalize historical wind power output;

[0018] The normalization formula is as follows:

[0019]

[0020] Where, X norm_ys is the original data before normalization; X min is the minimum value in the original data; X max is the maximum value in the original data; X norm is the original data after normalization; through normalization, the dimension of wind power output is eliminated and converted into a number between 0 and 1, and the rapid fluctuation points in wind power output are highlighted, which is convenient for subsequent discrete wavelet decomposition to extract frequency domain information; through normalization, high-frequency coefficients can more accurately reflect the relative fluctuation intensity; when the original data X after normalization is norm and high frequency coefficient X Wavelet When splicing, the contributions of both to the attention mechanism are equally weighted; the stability of BiLSTM is improved, so that activation functions such as sigmoid (control gating mechanism) and tanh (nonlinear features of learning output ramping) have the maximum gradient in the input range [0,1], avoiding gradient explosion;

[0021] S12: Normalized original data X norm Discrete wavelet decomposition is performed to obtain low-frequency coefficients and high-frequency coefficients;

[0022] The low-frequency coefficient decomposition formula is shown in formula (2):

[0023]

[0024] The high-frequency coefficient decomposition formula is shown in formula (3):

[0025]

[0026] Where j is the number of decomposition layers, k is the coefficient index; a j [k] is the low-frequency coefficient of the jth layer, d j[k] is the high-frequency coefficient of the jth layer; x[n] is the input signal; n is the index of the input signal x[n]; h is a low-pass filter used to extract low-frequency components; g is a high-pass filter used to extract high-frequency components;

[0027] S13: Use Transformer to combine long-term dependencies with frequency domain local features to generate multiple prediction sets;

[0028] Specifically, we build a Transformer model based on Python / torch.nn, and implement the Transformer model configuration by optimizing the key parameters of time step, hidden layer dimension, number of attention heads, and number of Transformer layers. The enhanced matrix X is obtained by concatenating the normalized raw data with the high-frequency coefficients obtained by discrete wavelet decomposition. sr , X sr The model is fed into the input layer for training, leveraging the Transformer self-attention mechanism and its unique advantage in processing the long-term dependency of wind power output to enhance the model's sensitivity to short-term fluctuations, improve prediction accuracy, and obtain the long-term dependency of wind power output and generate a prediction set.

[0029] Transformer input layer fusion input formula (4):

[0030] X sr =[X norm ;X Wavelet ](4)

[0031] Among them, X norm is the original data after normalization; X Wavelet is the high-frequency coefficient obtained by discrete wavelet decomposition; X sr is the fused input layer sequence;

[0032] S14: Use bidirectional LSTM to enhance the local features of the prediction set;

[0033] Build a BiLSTMModel model based on Python / torch.nn, and optimize the configuration of the BiLSTMModel model by optimizing the key parameters of the number of hidden dimensions and the number of layers; enable bidirectional processing, and use the prediction set obtained in step S13 as the training set for training to obtain the final prediction result X YC-GL ;

[0034] S15: Denormalize the final prediction result to obtain the final predicted wind power output;

[0035] X wt =X YC-GL (X max -X min )+X min (5)

[0036] Where, X wt is the wind power output after denormalization, unit: MW;

[0037] Preferably, the first empirical mode decomposition and the second empirical mode decomposition in step S2 include the following steps:

[0038] S21: Perform an empirical mode decomposition on the wind power generation power;

[0039] The specific decomposition formula is shown in formula (6):

[0040]

[0041] Where W wt (t) is the wind power generation power at time t, MW; imf j (t) is the eigenmode component of each order after decomposition at time t; r i (t) is the residual component remaining at time t; j is the number of decomposition levels; i represents the total number of decomposition levels;

[0042] S22: Reconstruct each eigenmode component, superimpose the eigenmode components from low to high and perform low-frequency reconstruction; select the maximum signal difference within a preset time after the low-frequency reconstruction for judgment; if the difference meets the fluctuation amount judgment formula, it is judged as a low-frequency component and directly connected to the grid; if it does not meet the fluctuation amount judgment formula, it is judged as a high-frequency component and a secondary empirical mode decomposition is performed;

[0043] The specific fluctuation determination formula is shown in formula (7):

[0044]

[0045] ΔW max ≤ΔW limit (8)

[0046] Where, F 2f(i)t is the signal fluctuation of the low-frequency reconstruction component of order i in time period t, ΔW max ΔW is the maximum difference of the signal in the time period t, MW; limit is the set fluctuation limit value, MW; Δo is the time interval.

[0047] S23: performing a secondary empirical mode decomposition on the high frequency components obtained in step S22;

[0048] The specific decomposition formula is as follows:

[0049]

[0050] Where U wt(t) is the high frequency component within the time period t obtained in step S22, MW; imf m (t) is the eigenmode component of each order after decomposition within the time period t; s k (t) is the residual component remaining in time period t; m is the number of decomposition levels; k is the total number of decomposition levels.

[0051] S24: Reconstructing the eigenmode components obtained in step S23, superimposing the eigenmode components obtained in step S23 from low to high and performing low-frequency reconstruction; selecting the maximum signal difference within a preset time after the low-frequency reconstruction for judgment; if the difference meets the fluctuation amount judgment formula, it is determined to be a low-frequency component and is delivered to the lithium battery system for absorption; if it does not meet the fluctuation amount judgment formula, it is determined to be a high-frequency component and is delivered to the supercapacitor for absorption;

[0052] The specific fluctuation determination formula is shown in formulas (10)-(11):

[0053]

[0054] ΔW max ≤ΔW limit (11)

[0055] Where, I 2f(k)t is the signal fluctuation of the k-order low-frequency reconstruction component in the t time period, ΔW max ΔW is the maximum difference of the signal in the time period t, MW; limit is the set fluctuation limit value, MW; Δo is the time interval.

[0056] Beneficial effects

[0057] The present invention uses a composite prediction model that combines discrete wavelet decomposition with Transformer to achieve accurate assessment of wind power output, providing a reliable basis for energy storage configuration. At the same time, the wind-storage hybrid energy storage system after the first reconstruction is subjected to secondary empirical mode decomposition and reconstruction, which reduces the difficulty of hybrid energy storage capacity configuration, improves economic efficiency and operational stability in the face of extreme situations, and reduces wind power configuration costs.

[0058] The present invention uses discrete wavelet decomposition to extract low-frequency coefficients that reflect long-term trends and high-frequency coefficients that reflect short-term fluctuations, thereby simultaneously locating features in the time and frequency domains and enhancing the model's ability to capture long-term trends and short-term fluctuations. Compared with Fourier transform, wavelet decomposition can simultaneously locate features in the time and frequency domains and is more suitable for non-stationary wind power output data. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Attachment Figure 1: Flowchart of the multi-feature composite wind power output prediction method based on discrete wavelet decomposition-Transformer-bidirectional LSTM of the present invention

[0060] Attachment Figure 2 :Flowchart of the configuration optimization method of wind-storage hybrid energy storage system based on quadratic empirical mode decomposition of the present invention

[0061] Attachment Figure 3 :Flowchart of the optimization method of the second empirical mode decomposition of the present invention

[0062] Attachment Figure 4 : Schematic diagram of the hybrid energy storage system of the present invention DETAILED DESCRIPTION

[0063] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0064] A method for optimizing the configuration of a wind-storage hybrid energy storage system based on quadratic empirical mode decomposition and multi-feature fusion, wherein the energy storage system includes batteries and supercapacitors, and the method comprises the following steps:

[0065] S1: Based on historical wind power output data, a multi-feature composite prediction method based on discrete wavelet decomposition, Transformer, and bidirectional LSTM is used to predict wind power output and generate wind power forecast results.

[0066] Specifically, the multi-feature composite wind power output prediction method using discrete wavelet decomposition-Transformer-bidirectional LSTM in step S1 includes the following steps:

[0067] S11: Use MinMaxScaler to normalize historical wind power output;

[0068] The normalization formula is as follows:

[0069]

[0070] Where, X norm_ys is the original data before normalization; X min is the minimum value in the original data; X max is the maximum value in the original data; X norm is the original data after normalization; through normalization, the dimension of wind power output is eliminated and converted into a number between 0 and 1, and the rapid fluctuation points in wind power output are highlighted, which is convenient for subsequent discrete wavelet decomposition to extract frequency domain information; through normalization, high-frequency coefficients can more accurately reflect the relative fluctuation intensity; when the original data X after normalization is normand high frequency coefficient X Wavelet When splicing, the contributions of both to the attention mechanism are equally weighted; the stability of BiLSTM is improved, so that activation functions such as sigmoid (control gating mechanism) and tanh (nonlinear features of learning output ramping) have the maximum gradient in the input range [0,1], avoiding gradient explosion;

[0071] S12: Normalized original data X norm Discrete wavelet decomposition is performed to obtain low-frequency coefficients and high-frequency coefficients;

[0072] The low-frequency coefficient decomposition formula is shown in formula (2):

[0073]

[0074] The high-frequency coefficient decomposition formula is shown in formula (23):

[0075]

[0076] Where j is the number of decomposition layers, k is the coefficient index; a j [k] is the low-frequency coefficient of the jth layer, d j [k] is the high-frequency coefficient of the jth layer; x[n] is the input signal; n is the index of the input signal x[n]; h is a low-pass filter used to extract low-frequency components; g is a high-pass filter used to extract high-frequency components;

[0077] Through discrete wavelet decomposition, low-frequency coefficients reflecting long-term trends and high-frequency coefficients reflecting short-term fluctuations are extracted, enabling simultaneous location of features in both the time and frequency domains, enhancing the model's ability to capture long-term trends and short-term fluctuations. Compared to Fourier transforms, wavelet decomposition can simultaneously locate features in both the time and frequency domains, making it more suitable for non-stationary wind power output data.

[0078] S13: Use Transformer to combine long-term dependencies with frequency domain local features to generate multiple prediction sets;

[0079] Specifically, the Transformer model is built based on Python / torch.nn, and the Transformer model configuration is achieved by optimizing key parameters such as time step, hidden layer dimension, number of attention heads, and number of Transformer layers. The enhanced matrix X is obtained by concatenating the normalized raw data with the high-frequency coefficients obtained by discrete wavelet decomposition. sr , X sr The model is fed into the input layer for training, utilizing the Transformer self-attention mechanism and its unique advantage in processing the long-term dependency of wind power output to enhance the model’s sensitivity to short-term fluctuations, improve prediction accuracy, obtain the long-term dependency of wind power output, and generate a prediction set.

[0080] The Transformer input layer fusion input formula is shown in formula (4):

[0081] X sr =[X norm ;X Wavelet ](4)

[0082] Among them, X norm is the original data after normalization; X Wavelet is the high-frequency coefficient obtained by discrete wavelet decomposition; X sr is the fused input layer sequence;

[0083] S14: Use bidirectional LSTM to enhance the local features of the prediction set;

[0084] Build a BiLSTMModel model based on Python / torch.nn, and optimize the configuration of the BiLSTMModel model by optimizing key hyperparameters such as the number of hidden dimensions and the number of layers. Enable bidirectional processing and use the prediction set obtained in step S13 as the training set for training to obtain the final prediction result X YC-GL .

[0085] The Transformer excels at modeling long-term dependencies but lacks sensitivity to local time series changes. Therefore, a BiLSTM is used after the Transformer model to compensate for its shortcomings in local time series modeling. This allows the model to learn long-term trends while also accurately capturing short-term fluctuations. Traditional unidirectional LSTMs can only utilize past information and are insensitive to reverse causal relationships. Using a bidirectional LSTM here can better capture wind power output relationships across time periods and is suitable for long-term wind power output forecasting.

[0086] S15: Denormalize the final prediction result to obtain the final predicted wind power output;

[0087] X wt =X YC-GL (X max -X min )+X min (5)

[0088] Where, X wt The wind power output is obtained after denormalization, in MW.

[0089] S2: Perform k-means cluster analysis on the wind power forecast results and historical wind power output data to obtain typical daily wind power output data, perform a primary empirical mode decomposition and low-frequency reconstruction of multiple intrinsic mode components on the typical daily wind power output data; calculate the signal fluctuation of the wind power low-frequency reconstructed component within a preset time period, and determine whether the maximum signal difference within the preset time after the low-frequency reconstruction meets the fluctuation limit; if so, it is determined to be a low-frequency component; if not, it is determined to be a high-frequency component, and the high-frequency component obtained by the primary empirical mode decomposition is subjected to a secondary empirical mode decomposition;

[0090] Specifically, performing secondary empirical mode decomposition on the high-frequency components obtained by the primary empirical mode decomposition in step S2 includes the following steps:

[0091] S21: Perform an empirical mode decomposition on the wind power generation power;

[0092] The first-order empirical mode decomposition formula is shown in formula (6):

[0093]

[0094] Where W wt (t) is the wind power generation power at time t, MW; imf j (t) is the eigenmode component of each order after decomposition at time t; r i (t) is the residual component remaining at time t; j is the number of decomposition levels; i is the total number of decomposition levels

[0095] The formula means that the sum of each order eigenmodal component and the remaining residual component is the original generated power of wind power.

[0096] S22: Reconstruct each eigenmode component, superimpose the eigenmode components from low to high and perform low-frequency reconstruction; select the maximum signal difference within a preset time after the low-frequency reconstruction for judgment; if the difference meets the fluctuation amount judgment formula, it is judged as a low-frequency component and directly connected to the grid; if it does not meet the fluctuation amount judgment formula, it is judged as a high-frequency component and a secondary empirical mode decomposition is performed;

[0097] The specific fluctuation amount judgment formula is shown in formula (7)-(8):

[0098]

[0099] ΔW max ≤ΔW limit (8)

[0100] Where, F 2f(i)t is the signal fluctuation of the low-frequency reconstruction component of order i in time period t, ΔW max ΔW is the maximum difference of the signal in the time period t, MW; limitis the set fluctuation limit value, MW; Δo is the time interval, which is set to 10 minutes here.

[0101] S23: performing a secondary empirical mode decomposition on the high frequency components obtained in step S22;

[0102] The quadratic empirical mode decomposition formula is as follows:

[0103]

[0104] Where U wt (t) is the high frequency component within the time period t obtained in step S22, MW; imf m (t) is the eigenmode component of each order after decomposition within the time period t; s k (t) is the residual component remaining in the time period t; m is the number of decomposition layers; k represents the total number of decomposition layers; the formula means that the sum of each order eigenmode component and the remaining residual component is the high-frequency component obtained in step S22.

[0105] S24: Reconstructing the eigenmode components obtained in step S23, superimposing the eigenmode components obtained in step S23 from low to high and performing low-frequency reconstruction; selecting the maximum signal difference within a preset time after the low-frequency reconstruction for judgment; if the difference meets the fluctuation amount judgment formula, it is determined to be a low-frequency component and is delivered to the lithium battery system for absorption; if it does not meet the fluctuation amount judgment formula, it is determined to be a high-frequency component and is delivered to the supercapacitor for absorption;

[0106] The specific fluctuation amount judgment formula is shown in formula (10)-(11):

[0107]

[0108] ΔW max ≤ΔW limit (11)

[0109] Where, I 2f(k)t is the signal fluctuation of the k-order low-frequency reconstruction component in the t time period, ΔW max ΔW is the maximum difference of the signal in the time period t, MW; limit is the set fluctuation limit value, MW; Δo is the time interval, which is set to 10 minutes here.

[0110] S3: Performing a second empirical mode decomposition on the high-frequency component obtained by the first empirical mode decomposition in step S2 to obtain a low-frequency component and a high-frequency component; absorbing the obtained low-frequency component by the lithium battery and absorbing the high-frequency component by the supercapacitor, thereby obtaining the rated power of the battery and the supercapacitor in the hybrid energy storage system;

[0111] S4: Optimize the capacity configuration of batteries and supercapacitors, with the goal of minimizing the total cost of the battery and supercapacitor hybrid energy storage system and the coupling relationship of each device as a constraint, and optimize the power / capacity configuration results of lithium batteries and supercapacitors based on Matlab / PSO.

[0112] Specific example analysis:

[0113] This invention is applied to a large wind farm with a total installed capacity of approximately 1700MW in Inner Mongolia. Based on the annual operation data at 5-minute intervals, it adopts the collaborative optimization method of "discrete wavelet decomposition-Transformer-bidirectional LSTM" prediction and "quadratic empirical mode decomposition" hybrid energy storage configuration to achieve high-precision prediction of wind power output and economic configuration of energy storage system. Figure 4 The specific implementation process is as follows:

[0114] S1: Data preprocessing: MinMaxScaler is used to normalize the wind power output data for the whole year to eliminate the impact of dimension on data prediction; frequency domain feature extraction: discrete wavelet decomposition is used to extract low-frequency trend components and high-frequency fluctuation components to enhance the model's sensitivity to violent fluctuation points; Transformer long-term dependency modeling: the normalized data and high-frequency coefficients are concatenated into an enhanced matrix Xsr, which is input into the Transformer model for training, and the multi-head attention mechanism is used to capture the long-term correlation of the output sequence; bidirectional LSTM optimization: using the Transformer output as the training set, the bidirectional LSTM is used to further extract the local features of the time series to generate the final prediction result; denormalization: MinMaxScaler is used to denormalize the final prediction result, and the dimension is reintroduced to generate the final predicted wind power output.

[0115] S2: Perform k-means cluster analysis on the wind power forecast results and historical wind power output data to obtain typical daily wind power output data, perform a primary empirical mode decomposition and low-frequency reconstruction of multiple intrinsic mode components on the typical daily wind power output data; decompose the forecast results into multi-order IMF components, and screen the low-frequency components that can be directly connected to the grid through low-frequency reconstruction; perform a secondary reconstruction on the high-frequency components obtained from the primary empirical mode decomposition, and absorb the low-frequency fluctuations by the lithium battery and the high-frequency fluctuations by the supercapacitor, thereby obtaining the rated power of the lithium battery and supercapacitor in the hybrid energy storage system;

[0116] S3: Capacity optimization solution: With the goal of minimizing the total cost of the energy storage system and the coupling relationship of each device as a constraint, the power / capacity configuration results of lithium batteries and supercapacitors are optimized based on Matlab / PSO (particle swarm optimization)

[0117] The results show that compared with the traditional solution, the capacity configuration cost under the solution of the present invention is reduced by about 5.3%.

[0118] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the inventive concept and technical solution of the present invention, or the inventive concept and technical solution are directly applied to other occasions without improvement, they are all within the scope of protection of the present invention.

Claims

1. A method for optimizing the configuration of a wind-storage hybrid energy storage system based on quadratic empirical mode decomposition and multi-feature fusion, wherein the energy storage system includes batteries and supercapacitors, and is characterized in that: The method comprises the following steps: S1: Based on historical wind power output data, a multi-feature composite prediction method based on discrete wavelet decomposition, Transformer, and bidirectional LSTM is used to predict wind power output and generate wind power forecast results. S2: Perform k-means cluster analysis on the wind power forecast results and historical wind power output data to obtain typical daily wind power output data, and perform an empirical mode decomposition and multi-eigenmode component low-frequency reconstruction on the typical daily wind power output data; Calculate the signal fluctuation of the low-frequency reconstruction component of wind power within a preset time period, and determine whether the maximum signal difference within the preset time after low-frequency reconstruction meets the fluctuation limit; if so, it is determined to be a low-frequency component; if not, it is determined to be a high-frequency component, and the high-frequency component obtained by the first empirical mode decomposition is subjected to a second empirical mode decomposition; S3: Performing a second empirical mode decomposition on the high-frequency component obtained by the first empirical mode decomposition in step S2 to obtain a low-frequency component and a high-frequency component; absorbing the obtained low-frequency component by the lithium battery and absorbing the high-frequency component by the supercapacitor, thereby obtaining the rated power of the battery and the supercapacitor in the hybrid energy storage system; S4: Optimize the capacity configuration of batteries and supercapacitors, with the goal of minimizing the total cost of the battery and supercapacitor hybrid energy storage system and the coupling relationship of each device as a constraint, and optimize the power / capacity configuration results of lithium batteries and supercapacitors based on Matlab / PSO.

2. The method for optimizing the configuration of a wind-storage hybrid energy storage system based on quadratic empirical mode decomposition and multi-feature fusion according to claim 1, characterized in that: The multi-feature composite wind power output prediction method using discrete wavelet decomposition-Transformer-bidirectional LSTM in step S1 includes the following steps: S11: Use MinMaxScaler to normalize historical wind power output; The normalization formula is shown in formula (1): Where, X norm_ys is the original data before normalization; X min is the minimum value in the original data; X max is the maximum value in the original data; X norm is the original data after normalization; S12: Normalized original data X norm Discrete wavelet decomposition is performed to obtain low-frequency coefficients and high-frequency coefficients; The low-frequency coefficient decomposition formula is shown in formula (2): The high-frequency coefficient decomposition formula is shown in formula (3): Where j is the number of decomposition layers, k is the coefficient index; a j [k] is the low-frequency coefficient of the jth layer, d j [k] is the high-frequency coefficient of the jth layer; x[n] is the input signal; n is the index of the input signal x[n]; h is a low-pass filter used to extract low-frequency components; g is a high-pass filter used to extract high-frequency components; S13: Use Transformer to combine long-term dependencies with frequency domain local features to generate multiple prediction sets; The enhanced matrix X is obtained by concatenating the normalized original data with the high-frequency coefficients obtained by discrete wavelet decomposition. sr , X sr The data is sent to the input layer for training to obtain the long-term dependency of wind power output and generate a prediction set; The Transformer input layer fusion input formula (4) is as follows: X sr =[X norm ;X Wavelet ](4) Among them, X norm is the original data after normalization; X Wavelet is the high-frequency coefficient obtained by discrete wavelet decomposition; X sr is the fused input layer sequence; S14: Use bidirectional LSTM to enhance the local features of the prediction set; The BiLSTMModel model is constructed based on Python / torch.nn. The optimal configuration of the BiLSTMModel model is achieved by optimizing the key parameters of the number of hidden dimensions and the number of layers. The prediction set obtained in step S13 is used as the training set for training to obtain the final prediction result X YC-GL ; S15: Denormalize the final prediction result to obtain the final predicted wind power output; The denormalization formula is shown in formula (5): X wt =X YC-GL (X max -X min )+X min (5) Where, X wt The wind power output is obtained after denormalization, in MW.

3. The method for optimizing the configuration of a wind-storage hybrid energy storage system based on quadratic empirical mode decomposition and multi-feature fusion according to claim 1, characterized in that: The first empirical mode decomposition and the second empirical mode decomposition in step S2 include the following steps: S21: Perform an empirical mode decomposition on the wind power generation power; The first-order empirical mode decomposition formula is shown in formula (6): Where W wt (t) is the wind power generation power at time t, MW; imf j (t) is the eigenmode component of each order after decomposition at time t; r i (t) is the residual component remaining at time t; j is the number of decomposition levels; i represents the total number of decomposition levels; S22: Reconstruct each eigenmode component, superimpose the eigenmode components from low to high and perform low-frequency reconstruction; select the maximum signal difference within a preset time after the low-frequency reconstruction for judgment; if the difference meets the fluctuation amount judgment formula, it is judged as a low-frequency component and directly connected to the grid; if it does not meet the fluctuation amount judgment formula, it is judged as a high-frequency component and a secondary empirical mode decomposition is performed; The specific fluctuation amount judgment formula is shown in formula (7)-(8): ΔW max ≤ΔW limit (8) Where, F 2f(i)t is the signal fluctuation of the low-frequency reconstruction component of order i in time period t, ΔW max ΔW is the maximum difference of the signal in the time period t, MW; limit is the set fluctuation limit value, MW; Δo is the time interval; S23: performing a secondary empirical mode decomposition on the high frequency components obtained in step S22; The quadratic empirical mode decomposition formula is shown in formula (9): Where U wt (t) is the high frequency component within the time period t obtained in step S22, MW; imf m (t) is the eigenmode component of each order after decomposition within the time period t; s k (t) is the residual component remaining in time period t; m is the number of decomposition levels; k is the total number of decomposition levels; S24: Reconstructing the eigenmode components obtained in step S23, superimposing the eigenmode components obtained in step S23 from low to high and performing low-frequency reconstruction; selecting the maximum signal difference within a preset time after the low-frequency reconstruction for judgment; if the difference meets the fluctuation amount judgment formula, it is determined to be a low-frequency component and is delivered to the lithium battery system for absorption; if it does not meet the fluctuation amount judgment formula, it is determined to be a high-frequency component and is delivered to the supercapacitor for absorption; The fluctuation amount judgment formula is shown in formula (10)-(11): ΔW max ≤ΔW limit (11) Where, I 2f(k)t is the signal fluctuation of the k-order low-frequency reconstruction component in the t time period, ΔW max ΔW is the maximum difference of the signal in the time period t, MW; limit is the set fluctuation limit value, MW; Δo is the time interval.

Images