A wind power prediction method based on parallel quadratic adaptive decomposition and bidirectional long short-term memory network

Abstract

This invention discloses a wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism. It constructs a meteorological feature screening model based on the maximum information coefficient, and uses an improved adaptive noise complete set empirical mode decomposition to perform an initial decomposition of the original wind power sequence. Based on sample entropy and fuzzy C-means clustering, the obtained modal components are divided into high-frequency, mid-frequency, and low-frequency classes. For the high-frequency components, a golden sine-cosine optimized particle swarm optimization algorithm is introduced with minimum envelope entropy as the fitness function to adaptively optimize the number of decomposed modes and the penalty factor of variational mode decomposition, and then performs a second decomposition. Finally, each decomposed component is input into a bidirectional long short-term memory network for parallel prediction, and the final wind power prediction value is obtained by superposition. The quadratic adaptive decomposition effectively suppresses mode aliasing and noise residue, and the bidirectional time series modeling fully captures the forward and backward dependencies of the sequence, significantly improving the accuracy and stability of wind power prediction.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, and in particular relates to a wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network in parallel. Background Technology

[0002] Currently, common technical approaches in wind power forecasting primarily focus on improving forecast accuracy to support the grid's effective absorption and stable operation of wind power fluctuations. With the continuous growth in installed wind power capacity and penetration rate, the inherent strong randomness and non-stationarity of wind power output place higher demands on forecasting models. To address this challenge, the industry commonly employs statistical models, physical models, and machine learning methods for power forecasting. In recent years, hybrid forecasting frameworks based on data decomposition have gradually become a research hotspot. This method reduces non-stationarity by decomposing the original power sequence into several sub-components and models each component separately, thereby improving overall forecasting performance.

[0003] However, existing decomposition-based prediction methods still have several limitations. First, most studies employ a single decomposition technique, whose key parameters often rely on empirical settings and lack adaptive adjustment mechanisms for different wind power scenarios, leading to unstable decomposition results. Second, traditional decomposition methods are prone to mode aliasing and residual noise during application, affecting the feature purity of the obtained components and thus limiting the accuracy of subsequent prediction models. Furthermore, existing studies often use unidirectional time-series modeling methods when constructing prediction models, failing to fully integrate the forward and backward contextual information of the sequence, thus limiting the model's ability to fully characterize complex wind power fluctuation patterns and its dynamic response capability. In summary, the problems with existing technologies are: (1) The parameters of the decomposition method used in the prediction study depend on manual setting, which is not adaptive enough and is difficult to adapt to the fluctuation characteristics of wind power sequences under different scenarios.

[0004] (2) Traditional decomposition is prone to introducing mode aliasing and noise residue, which reduces component quality and affects prediction accuracy.

[0005] (3) Time series modeling is mostly limited to one-way information extraction and fails to fully explore the two-way dependency relationship of the sequence, which restricts the model's ability to fully capture fluctuation features.

[0006] Addressing the challenges of manually setting parameters for wind power sequence decomposition, modal aliasing and residual noise, and the unidirectional limitations of time series modeling requires constructing an adaptive parameter optimization mechanism to overcome empirical dependence, designing a hybrid decomposition strategy to suppress modal aliasing and improve component purity, and introducing a bidirectional modeling framework to fully explore forward and backward dependencies in the time series. This involves the deep integration of multiple technical fields, including adaptive signal decomposition algorithms, intelligent optimization algorithms, and deep learning-based time series modeling. Simultaneously, it is crucial to ensure the model's generalization ability and stability under different wind resource characteristics and climate scenarios. The entire research process demands not only a solid foundation in time series analysis, optimization theory, and deep learning, but also consideration of computational efficiency and reliability verification in engineering practice. It represents a systematic and interdisciplinary challenge involving algorithm innovation, model construction, and empirical analysis.

[0007] Effectively addressing the aforementioned issues can significantly improve the accuracy and robustness of wind power forecasting, thereby enhancing the grid's ability to perceive and regulate wind power fluctuations, reducing system operating reserve costs, and improving the integration of new energy sources. High-precision forecasting can provide a reliable basis for day-ahead dispatching, real-time balancing, and ancillary services markets in the power system, supporting the safe grid connection of a high proportion of wind power and promoting clean energy substitution. Furthermore, the proposed quadratic adaptive decomposition and bidirectional parallel forecasting framework also provides a referable technical path for other time-series forecasting problems with non-stationary and strongly stochastic characteristics, possessing significant theoretical value and promising engineering applications.

[0008] Therefore, a wind power prediction method and system based on quadratic adaptive decomposition and bidirectional long short-term memory network is needed to solve the above problems. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network in parallel, which is used to solve the problem of insufficient prediction accuracy caused by the non-stationarity and strong randomness of wind power sequence in the prior art, and improves prediction accuracy and stability.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism includes the following steps: S1. Construct a meteorological feature screening model based on the maximum information coefficient, quantify the linear and nonlinear correlation between meteorological variables and wind power, and screen out the key meteorological feature set; S2, the original wind power sequence is initially decomposed using the improved adaptive noise complete set empirical mode decomposition to obtain several intrinsic mode components and residuals; S3, calculate the sample entropy of each intrinsic mode component, and divide each intrinsic mode component into high-frequency, mid-frequency and low-frequency components based on the sample entropy value using the fuzzy C-means clustering algorithm. S4. The golden sine and cosine optimization particle swarm algorithm is introduced. The minimum envelope entropy is used as the fitness function to adaptively optimize the number of decomposition modes and the penalty factor of the variational mode decomposition. The optimized variational mode decomposition is then used to perform secondary decomposition on the high-frequency class components to obtain multiple sub-mode components. S5, input the mid-frequency class component, low-frequency class component, residual and sub-mode component obtained in step S3 into the bidirectional long short-term memory network for prediction, and superimpose the prediction results of each bidirectional long short-term memory network to obtain the wind power prediction value. S6, based on a preset multi-scenario comparison framework, uses a solver to train and optimize the model and outputs the final wind power prediction results.

[0011] Preferably, the specific method of step S1 is as follows: Micrometric quantification (MIC) is used to quantify the correlation strength between candidate meteorological features and wind power in a wind farm in Xinjiang, and a key input feature set is constructed based on this. ; ; In the formula, Given a two-dimensional sequence data set, it is divided into several parts in a plane. List, The grid area of ​​the row; Defined as , The maximum information coefficient between them; The mutual information of the dataset after normalization; For search scope The grid within, where The total number of samples, hyperparameters Take 0.6; for , Based on dataset The maximum mutual information.

[0012] Preferably, step S2 includes performing a first ICEEMDAN decomposition on the wind power, specifically as follows: Define the signal to be decomposed as ; This is a local mean operator for the signal; This represents the first [unit / item] obtained through EMD decomposition. First-order modal components; introduced into the original sequence Gaussian white noise with zero group mean and adjustable variance Construct a noisy sequence: ; ; In the formula, The initial noise amplitude coefficient is set to 0.2; Calculate the local mean of all noisy signals and take the average to obtain the first-order residual. : ; Obtaining the first-order residual Then, the highest frequency component and the first-order intrinsic mode component in the original signal can be separated. Defined as the original wind power sequence and Difference: ; Continue adding Gaussian white noise and recursively calculate higher-order residuals and modal components: ; Achieve complete decomposition of the target signal, obtaining all modal components and their corresponding residuals.

[0013] Preferably, step S3 includes establishing a mathematical model for calculating sample entropy: For length of Modal component time series ,structure The sample entropy of the dimensional vector is calculated as follows: ; ; ; ; Repeat the above formula to increase the dimension to Calculate The final sample entropy value is: ; In the formula, Representing vectors and The distance between them; Indicates the embedding dimension as Time vectors Compared with other vectors in the sequence The distance between them is less than the tolerance threshold. The proportion; express The average value.

[0014] Preferably, step S3 further includes establishing a fuzzy C-means clustering mathematical model: Let the sample set to be clustered be... Each sample The sample entropy value corresponding to an IMF component constitutes 3D feature vectors; the algorithm's objective is to set this... Each sample is optimized and divided into preset categories. In each cluster; (1) The fuzzy C-means algorithm is optimized by minimizing the following objective function: ; In the formula, It is The fuzzy membership matrix; Indicates the first The sample belongs to the first The membership degree of each cluster satisfies And for any sample All ; yes A set of cluster centers; The fuzzy weighting index is used to control the degree of fuzziness in the clustering results; in this method, a fixed value of 3 is used. Indicates sample To the cluster center The Euclidean distance; (2) The fuzzy C-means algorithm optimizes the membership degree alternately. and cluster center Until the objective function converges, given the current cluster centers, recalculate the membership degree of each sample to each cluster based on the inverse relationship between the distances between the samples and each center: ; (3) With the current membership matrix fixed, recalculate the center position of each cluster based on the weighted average of membership degrees: ; The above update steps are iterated alternately until the objective function... The iteration process terminates when the decrease in membership matrix is ​​less than the preset convergence tolerance. With cluster center Once stability is achieved, the membership matrix of each IMF component to high, medium, and low frequency clusters and the corresponding cluster centers are finally obtained.

[0015] Preferably, in step S4, the particle swarm optimization algorithm specifically includes: Defined in In the 3D search space, there are A population of particles , of which The position of each particle is represented as Its speed is expressed as The best position in each particle's individual history is denoted as ; The optimal position experienced by all particles in the entire population is recorded as the global optimal position. During each iteration, the particle's velocity and position are updated, with the velocity update as follows: ; In the formula, , They represent particles respectively In the In the nth iteration Dimensional velocity and position; Inertia weighting factor; , The learning factor is set to 2. A random number between [0, 1]; , They represent particles respectively In the In the nth iteration The individual's historical optimal position and global optimal position; For inertia weighting factor A linear decreasing weight strategy is used for calculation: ; In the formula, The initial inertia weight is set to 0.9; The inertial weight is set to 0.4 when selecting the maximum number of generations. and These are the maximum number of generations to be evolved in the population and the current number of generations to be evolved in the population, respectively. The particle positions are updated as follows: .

[0016] Preferably, in step S4, when using the golden sine and cosine optimized particle swarm optimization algorithm, the traditional linear superposition cognitive term is replaced with a nonlinear guidance mechanism based on sine and cosine functions, and a dynamic coefficient related to the golden ratio is used for adjustment; specifically, its velocity update formula is modified as follows: ; In the formula, and These are the values ​​of the sine and cosine functions, respectively. The internal random phase angle; This is a dynamic adjustment coefficient. The value is adaptively adjusted according to the golden ratio as the iteration process progresses: ; In the formula, and These are constants related to the golden ratio, taken as 0.618 and 1.618 respectively; The particle position update in the Golden Sine / Cosine Optimized Particle Swarm Optimization (PSO) algorithm is the same as in the PSO algorithm: ; A golden sine and cosine optimization particle swarm optimization algorithm is introduced, using the minimum envelope entropy as the fitness function to optimize the parameter combination (K, Adaptive optimization is performed.

[0017] Preferably, the specific method of step S5 is as follows: Bidirectional Long Short-Term Memory (BSSM) networks introduce cellular states as the main thread of continuous-time memory and dynamically regulate them using three gating mechanisms: forget gate, input gate, and output gate. The mathematical model of the BSSM network is expressed as follows: ; ; ; ; ; ; In the formula, , , These represent the forget gate, input gate, and output gate respectively. Output vector at time step; and These represent the current cell state and the hidden state, respectively. , , and Here is the weight matrix corresponding to each gate unit. , , and Then it is the corresponding bias vector; and These represent the Sigmoid activation function and the hyperbolic tangent activation function, respectively.

[0018] Preferably, the multi-scenario comparison framework in step S6 is as follows: Scenario 1: The original wind power sequence is subjected to variational mode decomposition and then input into the BiLSTM model as the predicted value for wind power prediction, i.e., VMD-BiLSTM; Scenario II: The original wind power sequence is optimized by particle swarm optimization and then input into the BiLSTM model as the prediction value for wind power prediction, i.e., PSO-VMD-BiLSTM; Scenario III: The original wind power sequence is decomposed in two stages. The first stage uses CEEMDAN decomposition, and the second stage uses particle swarm optimization to optimize variational mode decomposition. The decomposed sequence is then input into the BiLSTM model as the predicted value for wind power prediction, i.e., CEEMDAN-PSO-VMD-BiLSTM. Scenario IV: The original wind power sequence is decomposed in two stages. The first stage uses ICEEMDAN decomposition, and the second stage uses the Golden Sine Cosine Optimization Particle Swarm Optimization Algorithm to optimize the variational mode decomposition. The decomposed sequence is then input into the LSTM model as the prediction value for wind power prediction, i.e., ICEEMDAN-GDPSO-VMD-LSTM. Scenario V: The original wind power sequence is decomposed in two stages. The first stage uses ICEEMDAN decomposition, and the second stage uses the Golden Sine Cosine Optimized Particle Swarm Optimization (GSO) algorithm to optimize the variational mode decomposition. The decomposed sequence is then input into the BiLSTM model as the predicted value for wind power prediction, i.e., ICEEMDAN-GDPSO-VMD-BiLSTM.

[0019] Preferably, the specific method for training the model and optimizing the parameters using a solver in step S6 to output the final wind power prediction result is as follows: S601, Initialize the prediction model parameters, set the maximum number of iterations T to 30, the current number of iterations t=0, and the error convergence threshold to 0.0001; S602, construct five prediction model architectures corresponding to scenarios I to V respectively, and randomly initialize the weights and bias parameters to be optimized for each model; S603, t=t+1; S604: Input the training dataset into the five scenario models respectively, and calculate the wind power prediction value for each scenario through forward propagation; S605, calculate and retain the prediction error of each scenario model in the current iteration based on the predicted value and the actual value; S606 compares the prediction error of each scenario with the previous generation's best error. If the new error is better, the optimal model parameters and optimal error for that scenario are updated; otherwise, the original optimal parameters and error are retained. S607. If all scene models have reached the maximum number of iterations T or the optimal error change is less than the convergence threshold, the program ends; otherwise, jump to S608. S608, update the model parameters according to the optimization algorithm corresponding to each scenario, and repeat S603-S607.

[0020] Furthermore, it also includes a computer device, comprising a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to implement the wind power prediction method based on parallel quadratic adaptive decomposition and bidirectional long short-term memory network.

[0021] Furthermore, it also includes a computer-readable storage medium storing computer instructions for causing a computer to execute the wind power prediction method based on a parallel quadratic adaptive decomposition and bidirectional long short-term memory network.

[0022] The beneficial effects of this invention are as follows: 1. This invention quantifies the linear and nonlinear correlation between meteorological variables and wind power by constructing a meteorological feature screening model based on the maximum information coefficient, significantly improving the effectiveness of input features and laying a reliable data foundation for subsequent high-precision prediction. This method introduces a quadratic adaptive decomposition framework that combines improved adaptive noise complete set empirical mode decomposition with variational mode decomposition optimized by golden sine and cosine optimization particle swarm optimization algorithm. This effectively suppresses mode aliasing and noise residue problems, improves the purity and stability of the decomposed components, and thus enhances the model's ability to extract complex fluctuation features of wind power.

[0023] 2. This technical solution utilizes bidirectional long short-term memory networks for parallel modeling, fully exploring the forward and backward dependencies of time-series data and enhancing the model's dynamic response capability to non-stationary, strongly random wind power sequences. Simultaneously, it incorporates a golden sine / cosine optimization particle swarm optimization algorithm to adaptively optimize decomposition parameters, avoiding the blindness of manual parameter tuning and enhancing the model's generalization ability and robustness across different wind power scenarios.

[0024] 3. This invention not only significantly improves the accuracy and stability of wind power prediction, but also provides reliable technical support for power system dispatch and operation under conditions of high wind power grid connection. This method has significant engineering application value in improving the capacity for renewable energy absorption, reducing system operating reserve costs, and promoting clean energy substitution. It also provides a generalizable technical path for other non-stationary time-series prediction problems, demonstrating its broad prospects in the construction of smart grids and new power systems. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the correlation analysis of different meteorological characteristics MIC in this invention patent; Figure 2This is the overall flowchart of the ICEEMDAN-GDPSO-VMD-BiLSTM prediction model of this invention patent; Figure 3 This is a schematic diagram of the entropy and frequency classification membership of the IMF samples after ICEEMDAN decomposition in an embodiment of the present invention. Figure 4 This is a diagram showing the fuzzy C-means clustering results of an embodiment of the present invention; Figure 5 This is a graph showing the intrinsic mode function results obtained from the secondary decomposition of high-frequency components according to an embodiment of the present invention; Figure 6 These are graphs showing the wind power prediction results obtained from different prediction models in embodiments of the present invention. Figure 7 This is a graph showing the wind power prediction results of various prediction models under different data sources obtained in the embodiments of the present invention; Figure 8 This is a schematic diagram of the computer device structure in an embodiment of the present invention. Detailed Implementation

[0026] Example 1: A wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism includes the following steps: S1. Construct a meteorological feature screening model based on the maximum information coefficient, quantify the linear and nonlinear correlation between meteorological variables and wind power, and screen out the key meteorological feature set; S2, the original wind power sequence is initially decomposed using the improved adaptive noise complete set empirical mode decomposition to obtain several intrinsic mode components and residuals; S3, calculate the sample entropy of each intrinsic mode component, and divide each intrinsic mode component into high-frequency, mid-frequency and low-frequency components based on the sample entropy value using the fuzzy C-means clustering algorithm. S4. The golden sine and cosine optimization particle swarm algorithm is introduced. The minimum envelope entropy is used as the fitness function to adaptively optimize the number of decomposition modes and the penalty factor of the variational mode decomposition. The optimized variational mode decomposition is then used to perform secondary decomposition on the high-frequency class components to obtain multiple sub-mode components. S5, input the mid-frequency class component, low-frequency class component, residual and sub-mode component obtained in step S3 into the bidirectional long short-term memory network for prediction, and superimpose the prediction results of each bidirectional long short-term memory network to obtain the wind power prediction value. S6, based on a preset multi-scenario comparison framework, uses a solver to train and optimize the model and outputs the final wind power prediction results.

[0027] Preferably, the specific method of step S1 is as follows: Micrometric quantification (MIC) is used to quantify the correlation strength between candidate meteorological features and wind power in a wind farm in Xinjiang, and a key input feature set is constructed based on this. ; ; In the formula, Given a two-dimensional sequence data set, it is divided into several parts in a plane. List, The grid area of ​​the row; Defined as , The maximum information coefficient between them; The mutual information of the dataset after normalization; For search scope The grid within, where The total number of samples, hyperparameters Take 0.6; for , Based on dataset The maximum mutual information.

[0028] Preferably, step S2 includes performing a first ICEEMDAN decomposition on the wind power, specifically as follows: Define the signal to be decomposed as ; This is a local mean operator for the signal; This represents the first [unit / item] obtained through EMD decomposition. First-order modal components; introduced into the original sequence Gaussian white noise with zero group mean and adjustable variance Construct a noisy sequence: ; ; In the formula, The initial noise amplitude coefficient is set to 0.2; Calculate the local mean of all noisy signals and take the average to obtain the first-order residual. : ; Obtaining the first-order residual Then, the highest frequency component and the first-order intrinsic mode component in the original signal can be separated. Defined as the original wind power sequence and Difference: ; Continue adding Gaussian white noise and recursively calculate higher-order residuals and modal components: ; Achieve complete decomposition of the target signal, obtaining all modal components and their corresponding residuals.

[0029] Preferably, step S3 includes establishing a mathematical model for calculating sample entropy: For length of Modal component time series ,structure The sample entropy of the dimensional vector is calculated as follows: ; ; ; ; Repeat the above formula to increase the dimension to Calculate The final sample entropy value is: ; In the formula, Representing vectors and The distance between them; Indicates the embedding dimension as Time vectors Compared with other vectors in the sequence The distance between them is less than the tolerance threshold. The proportion; express The average value.

[0030] Preferably, step S3 further includes establishing a fuzzy C-means clustering mathematical model: Let the sample set to be clustered be... Each sample The sample entropy value corresponding to an IMF component constitutes 3D feature vectors; the algorithm's objective is to set this... Each sample is optimized and divided into preset categories. In each cluster; (1) The fuzzy C-means algorithm is optimized by minimizing the following objective function: ; In the formula, It is The fuzzy membership matrix; Indicates the first The sample belongs to the first The membership degree of each cluster satisfies And for any sample All ; yes A set of cluster centers; The fuzzy weighting index is used to control the degree of fuzziness in the clustering results; in this method, a fixed value of 3 is used. Indicates sample To the cluster center The Euclidean distance; (2) The fuzzy C-means algorithm optimizes the membership degree alternately. and cluster center Until the objective function converges, given the current cluster centers, recalculate the membership degree of each sample to each cluster based on the inverse relationship between the distances between the samples and each center: ; (3) With the current membership matrix fixed, recalculate the center position of each cluster based on the weighted average of membership degrees: ; The above update steps are iterated alternately until the objective function... The iteration process terminates when the decrease in membership matrix is ​​less than the preset convergence tolerance. With cluster center Once stability is achieved, the membership matrix of each IMF component to high, medium, and low frequency clusters and the corresponding cluster centers are finally obtained.

[0031] Preferably, in step S4, the particle swarm optimization algorithm specifically includes: Defined in In the 3D search space, there are A population of particles , of which The position of each particle is represented as Its speed is expressed as The best position in each particle's individual history is denoted as ; The optimal position experienced by all particles in the entire population is recorded as the global optimal position. During each iteration, the particle's velocity and position are updated, with the velocity update as follows: ; In the formula, , They represent particles respectively In the In the nth iteration Dimensional velocity and position; Inertia weighting factor; , The learning factor is set to 2. A random number between [0, 1]; , They represent particles respectively In the In the nth iteration The individual's historical optimal position and global optimal position; For inertia weighting factor A linear decreasing weight strategy is used for calculation: ; In the formula, The initial inertia weight is set to 0.9; The inertial weight is set to 0.4 when selecting the maximum number of generations. and These are the maximum number of generations to be evolved in the population and the current number of generations to be evolved in the population, respectively. The particle positions are updated as follows: .

[0032] Preferably, in step S4, when using the golden sine and cosine optimized particle swarm optimization algorithm, the traditional linear superposition cognitive term is replaced with a nonlinear guidance mechanism based on sine and cosine functions, and a dynamic coefficient related to the golden ratio is used for adjustment; specifically, its velocity update formula is modified as follows: ; In the formula, and These are the values ​​of the sine and cosine functions, respectively. The internal random phase angle; This is a dynamic adjustment coefficient. The value is adaptively adjusted according to the golden ratio as the iteration process progresses: ; In the formula, and These are constants related to the golden ratio, taken as 0.618 and 1.618 respectively; The particle position update in the Golden Sine / Cosine Optimized Particle Swarm Optimization (PSO) algorithm is the same as in the PSO algorithm: ; A golden sine and cosine optimization particle swarm optimization algorithm is introduced, using the minimum envelope entropy as the fitness function to optimize the parameter combination (K, Adaptive optimization is performed.

[0033] Preferably, the specific method of step S5 is as follows: Bidirectional Long Short-Term Memory (BSSM) networks introduce cellular states as the main thread of continuous-time memory and dynamically regulate them using three gating mechanisms: forget gate, input gate, and output gate. The mathematical model of the BSSM network is expressed as follows: ; ; ; ; ; ; In the formula, , , These represent the forget gate, input gate, and output gate respectively. Output vector at time step; and These represent the current cell state and the hidden state, respectively. , , and Here is the weight matrix corresponding to each gate unit. , , and Then it is the corresponding bias vector; and These represent the Sigmoid activation function and the hyperbolic tangent activation function, respectively.

[0034] Preferably, the multi-scenario comparison framework in step S6 is as follows: Scenario 1: The original wind power sequence is subjected to variational mode decomposition and then input into the BiLSTM model as the predicted value for wind power prediction, i.e., VMD-BiLSTM; Scenario II: The original wind power sequence is optimized by particle swarm optimization and then input into the BiLSTM model as the prediction value for wind power prediction, i.e., PSO-VMD-BiLSTM; Scenario III: The original wind power sequence is decomposed in two stages. The first stage uses CEEMDAN decomposition, and the second stage uses particle swarm optimization to optimize variational mode decomposition. The decomposed sequence is then input into the BiLSTM model as the predicted value for wind power prediction, i.e., CEEMDAN-PSO-VMD-BiLSTM. Scenario IV: The original wind power sequence is decomposed in two stages. The first stage uses ICEEMDAN decomposition, and the second stage uses the Golden Sine Cosine Optimization Particle Swarm Optimization Algorithm to optimize the variational mode decomposition. The decomposed sequence is then input into the LSTM model as the prediction value for wind power prediction, i.e., ICEEMDAN-GDPSO-VMD-LSTM. Scenario V: The original wind power sequence is decomposed in two stages. The first stage uses ICEEMDAN decomposition, and the second stage uses the Golden Sine Cosine Optimized Particle Swarm Optimization (GSO) algorithm to optimize the variational mode decomposition. The decomposed sequence is then input into the BiLSTM model as the predicted value for wind power prediction, i.e., ICEEMDAN-GDPSO-VMD-BiLSTM.

[0035] Preferably, the specific method for training the model and optimizing the parameters using a solver in step S6 to output the final wind power prediction result is as follows: S601, Initialize the prediction model parameters, set the maximum number of iterations T to 30, the current number of iterations t=0, and the error convergence threshold to 0.0001; S602, construct five prediction model architectures corresponding to scenarios I to V respectively, and randomly initialize the weights and bias parameters to be optimized for each model; S603, t=t+1; S604: Input the training dataset into the five scenario models respectively, and calculate the wind power prediction value for each scenario through forward propagation; S605, calculate and retain the prediction error of each scenario model in the current iteration based on the predicted value and the actual value; S606 compares the prediction error of each scenario with the previous generation's best error. If the new error is better, the optimal model parameters and optimal error for that scenario are updated; otherwise, the original optimal parameters and error are retained. S607. If all scene models have reached the maximum number of iterations T or the optimal error change is less than the convergence threshold, the program ends; otherwise, jump to S608. S608, update the model parameters according to the optimization algorithm corresponding to each scenario, and repeat S603-S607.

[0036] Example 2: This embodiment discloses the specific implementation process of applying this method to a wind farm in Xinjiang. The correlation analysis of different meteorological characteristics obtained in this embodiment is as follows: Figure 1 As shown; the overall process of the prediction model based on ICEEMDAN-GDPSO-VMD-BiLSTM is as follows: Figure 2 As shown.

[0037] Scene description: Scenario I involves using variational mode decomposition on the original wind power sequence and then inputting it into a BiLSTM model as the predicted value for wind power prediction, i.e., VMD-BiLSTM.

[0038] Scenario II involves optimizing the variational mode decomposition of the original wind power sequence using the particle swarm optimization algorithm and then inputting the result into a BiLSTM model as the predicted value for wind power prediction, i.e., PSO-VMD-BiLSTM.

[0039] Scenario III involves a two-stage decomposition of the original wind power sequence. The first stage uses CEEMDAN decomposition, and the second stage uses particle swarm optimization to optimize variational mode decomposition. The decomposed sequence is then input into a BiLSTM model as the predicted value for wind power prediction, i.e., CEEMDAN-PSO-VMD-BiLSTM.

[0040] Scenario IV involves a two-stage decomposition of the original wind power sequence. The first stage uses ICEEMDAN decomposition, and the second stage uses the Golden Sine Cosine Optimized Particle Swarm Optimization (GSO) algorithm to optimize the variational mode decomposition. The decomposed sequence is then input into an LSTM model as the predicted value for wind power prediction, i.e., ICEEMDAN-GDPSO-VMD-LSTM.

[0041] In scenario V, the original wind power sequence is decomposed in two ways. The first decomposition uses ICEEMDAN, and the second decomposition uses the Golden Sine and Cosine Optimized Particle Swarm Optimization (GSO) algorithm to optimize the variational mode decomposition. The decomposed sequence is then input into the BiLSTM model as the predicted value for wind power prediction, i.e., ICEEMDAN-GDPSO-VMD-BiLSTM.

[0042] Scene calculation: This embodiment uses the CPLEX solver for calculation, and the specific steps are as follows: S6.1 Initialize the prediction model parameters, set the maximum number of iterations T to 30, the current number of iterations t=0, and the error convergence threshold to 0.0001.

[0043] S6.2 Construct five prediction model architectures for scenarios I to V respectively, and randomly initialize the weights and bias parameters to be optimized for each model.

[0044] S6.3, t=t+1.

[0045] S6.4 Input the training dataset into the five scenario models respectively, and calculate the wind power prediction value for each scenario through forward propagation.

[0046] S6.5 Calculate and retain the prediction error of each scenario model in the current iteration based on the predicted and actual values.

[0047] S6.6 compares the prediction error of each scenario with the previous generation's best error. If the new error is better, the optimal model parameters and optimal error for that scenario are updated; otherwise, the original optimal parameters and error are retained.

[0048] S6.7 If all scenario models reach the maximum number of iterations T or the optimal error change is less than the convergence threshold, the program ends; otherwise, jump to S6.8.

[0049] S6.8 Update the model parameters according to the optimization algorithm corresponding to each scenario, and repeat S6.3-S6.7.

[0050] Results analysis: Scenario V represents the application of the method of the present invention in an embodiment. The comparison of the prediction performance evaluation indicators of different models is shown in Table 1 below.

[0051] Table 1: Comparison of prediction performance evaluation of different models;

[0052] After ICEEMDAN decomposition, the entropy and frequency of the IMF samples and their classification membership are as follows: Figure 3 As shown, based on the fuzzy C-means clustering results... Figure 4 As shown, the eigenmode functions obtained by performing GDPSO-VMD quadratic decomposition on the high-frequency components are as follows: Figure 5 As shown, the wind power prediction results of different prediction models are as follows: Figure 6 As shown.

[0053] The prediction performance evaluation metrics of each model under different data sources are compared in Table 2 below.

[0054] Table 2: Comparison of prediction performance evaluation metrics for different models under different data sources;

[0055] Wind power prediction results of various prediction models under different data sources are as follows: Figure 7 As shown.

[0056] Table 1 shows that the VMD-BiLSTM model has the highest error indices and the lowest prediction accuracy. After introducing the particle swarm optimization algorithm, the MAE, MSE, MAPE, and RMSE of the PSO-VMD-BiLSTM model are reduced by 10.47%, 6.10%, 7.88%, and 2.51% respectively compared to VMD-BiLSTM, indicating that adaptive parameter optimization can effectively improve prediction accuracy. Further combining the CEEMDAN-PSO-VMD-BiLSTM model with CEEMDAN decomposition, all error indices continue to decrease, with RMSE further reduced by 10.44%, indicating that the secondary decomposition strategy helps extract more effective temporal features. The proposed ICEEMDAN-GDPSO-VMD-BiLSTM model in this study, based on the CEEMDAN-PSO-VMD-BiLSTM model, introduces an improved adaptive noise injection mechanism and a golden sine / cosine strategy-optimized particle swarm optimization algorithm, significantly improving prediction performance: MAE is reduced by 22.18%, RMSE by 30.53%, MAPE by 16.20%, and R² is increased to 0.9963, demonstrating superior fitting ability and generalization performance. Furthermore, compared to ICEEMDAN-GDPSO-VMD-LSTM, this study employs a bidirectional long short-term memory network to simultaneously capture the forward and backward dependencies of time-series data, thereby extracting feature information more comprehensively and further improving prediction performance. This verifies the effectiveness and advancement of the proposed model in wind power prediction.

[0057] At the same time Figure 6It can be seen that, overall, all models can roughly track the changing trend of actual wind power, but there are significant differences in tracking accuracy and detail characterization ability at different fluctuation stages. Among them, the prediction curve of the VMD-BiLSTM model deviates significantly during periods of rapid power change, especially at power peaks and troughs, where its tracking ability is weak. For example, in the interval of sample points 10 to 35 in the test set, the model's prediction of peak power is generally low, and it shows obvious response lag during sharp power fluctuations. The PSO-VMD-BiLSTM model has improved in terms of overall trend fitting through parameter optimization, but the prediction of steep power change intervals still exhibits over-smoothing, failing to accurately reproduce the high-frequency fluctuation details in actual power. The CEEMDAN-PSO-VMD-BiLSTM model, which further incorporates CEEMDAN for secondary decomposition, enhances the ability to analyze and model high-frequency components. During periods of rapid power fluctuation, the prediction curve is closer to the actual value, and the accuracy of capturing short-term peaks and troughs is improved. However, from the perspective of overall fitting effect, there is still some phase lag and amplitude deviation. Compared to the models mentioned above, the proposed ICEEMDAN-GDPSO-VMD-BiLSTM model demonstrates superior tracking performance across all time periods. This model not only more accurately identifies power peak and valley locations but also closely follows actual power fluctuations, significantly reducing the average deviation and lag of the prediction curve and exhibiting better dynamic response characteristics. To further validate the effectiveness of the model structure, BiLSTM was replaced with LSTM within the same framework for comparative analysis. The results show that LSTM exhibits a more pronounced smoothing effect and tracking delay in predicting the rising and falling edges of rapidly changing power, while BiLSTM, with its bidirectional information processing capabilities, performs better in capturing complex temporal dependencies and improving dynamic response accuracy, thus comprehensively enhancing the robustness and accuracy of the wind power prediction model.

[0058] To verify the generalization ability of the proposed model, wind power data from a wind farm in Ningxia for the entire year of 2017 were selected for testing, with a data sampling interval of 15 minutes. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The model training results are shown in Table 2. Figure 7As shown in Table 2, data analysis reveals that compared to the VMD-BiLSTM, PSO-VMD-BiLSTM, CEEMDAN-PSO-VMD-BiLSTM, and ICEEMDAN-GDPSO-VMD-LSTM models, the model constructed in this paper performs better in all evaluation indicators: its MAE decreases by 62.41%, 36.11%, 28.70%, and 22.76%, respectively; MSE decreases by 85.56%, 68.70%, 61.17%, and 40.81%, respectively; MAPE decreases by 77.68%, 50.06%, 38.83%, and 24.28%, respectively; RMSE decreases by 61.99%, 44.06%, 37.69%, and 23.07%, respectively; and R² increases by 2.68%, 0.85%, 0.62%, and 0.28%, respectively. Meanwhile, Figure 7 The power prediction curves shown indicate that the power curves predicted by the model proposed in this study are closest to the actual wind power curves, further verifying that the proposed model has good generalization performance and prediction stability.

[0059] Example 3: like Figure 8 As shown, this embodiment of the invention also provides a computer device, which includes one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory for displaying graphical information of a GUI on external input / output devices, such as display devices coupled to the interfaces. In some alternative embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations, for example, as a server array, a group of blade servers, or a multiprocessor system. Figure 8 Take a processor 10 as an example.

[0060] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0061] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0062] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0063] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0064] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0065] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

Claims

1. A wind power prediction method based on parallel quadratic adaptive decomposition and bidirectional long short-term memory network, characterized in that, Includes the following steps: S1. Construct a meteorological feature screening model based on the maximum information coefficient, quantify the linear and nonlinear correlation between meteorological variables and wind power, and screen out the key meteorological feature set; S2, the original wind power sequence is initially decomposed using the improved adaptive noise complete set empirical mode decomposition to obtain several intrinsic mode components and residuals; S3, calculate the sample entropy of each intrinsic mode component, and divide each intrinsic mode component into high-frequency, mid-frequency and low-frequency components based on the sample entropy value using the fuzzy C-means clustering algorithm. S4. The golden sine and cosine optimization particle swarm algorithm is introduced. The minimum envelope entropy is used as the fitness function to adaptively optimize the number of decomposition modes and the penalty factor of the variational mode decomposition. The optimized variational mode decomposition is then used to perform secondary decomposition on the high-frequency class components to obtain multiple sub-mode components. S5, input the mid-frequency class component, low-frequency class component, residual and sub-mode component obtained in step S3 into the bidirectional long short-term memory network for prediction, and superimpose the prediction results of each bidirectional long short-term memory network to obtain the wind power prediction value. S6, based on a preset multi-scenario comparison framework, uses a solver to train and optimize the model and outputs the final wind power prediction results.

2. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism as described in claim 1, characterized in that, The specific method for step S1 is as follows: Micrometric quantification (MIC) is used to quantify the correlation strength between candidate meteorological features and wind power in a wind farm in Xinjiang, and a key input feature set is constructed based on this. ； ； In the formula, Given a two-dimensional sequence data set, it is divided into several parts in a plane. List, The grid area of ​​the row; Defined as , The maximum information coefficient between them; The mutual information of the dataset after normalization; For search scope The grid within, where The total number of samples, hyperparameters Take 0.6; for , Based on dataset The maximum mutual information.

3. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism as described in claim 1, characterized in that, Step S2 includes performing the first ICEEMDAN decomposition of wind power, specifically as follows: Define the signal to be decomposed as ; This is a local mean operator for the signal; This represents the first [unit / item] obtained through EMD decomposition. First-order modal components; introduced into the original sequence Gaussian white noise with zero group mean and adjustable variance Construct a noisy sequence: ； ； In the formula, The initial noise amplitude coefficient; Calculate the local mean of all noisy signals and take the average to obtain the first-order residual. : ； Obtaining the first-order residual Then, the highest frequency component and the first-order intrinsic mode component in the original signal can be separated. Defined as the original wind power sequence and Difference: ； Continue adding Gaussian white noise and recursively calculate higher-order residuals and modal components: ； Achieve complete decomposition of the target signal, obtaining all modal components and their corresponding residuals.

4. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism as described in claim 1, characterized in that, Step S3 includes establishing a mathematical model for calculating sample entropy: For length of Modal component time series ,structure The sample entropy of the dimensional vector is calculated as follows: ； ； ； ； Repeat the above formula to increase the dimension to Calculate The final sample entropy value is: ； In the formula, Representing vectors and The distance between them; Indicates the embedding dimension as Time vectors Compared with other vectors in the sequence The distance between them is less than the tolerance threshold. The proportion; express The average value.

5. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism as described in claim 4, characterized in that, Step S3 also includes establishing a fuzzy C-means clustering mathematical model: Let the sample set to be clustered be... Each sample The sample entropy value corresponding to an IMF component constitutes 3D feature vectors; the algorithm's objective is to set this... Each sample is optimized and divided into preset categories. In each cluster; (1) The fuzzy C-means algorithm is optimized by minimizing the following objective function: ； In the formula, It is The fuzzy membership matrix; Indicates the first The sample belongs to the first The membership degree of each cluster satisfies And for any sample All ; yes A set of cluster centers; This is a fuzzy weighting index used to control the degree of fuzziness in the clustering results. Indicates sample To the cluster center The Euclidean distance; (2) The fuzzy C-means algorithm optimizes the membership degree alternately. and cluster center Until the objective function converges, given the current cluster centers, recalculate the membership degree of each sample to each cluster based on the inverse relationship between the distances between the samples and each center: ； (3) With the current membership matrix fixed, recalculate the center position of each cluster based on the weighted average of membership degrees: ； The above update steps are iterated alternately until the objective function... The iteration process terminates when the decrease in membership matrix is ​​less than the preset convergence tolerance. With cluster center Once stability is achieved, the membership matrix of each IMF component to high, medium, and low frequency clusters and the corresponding cluster centers are finally obtained.

6. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism according to claim 1, characterized in that, In step S4, the particle swarm optimization algorithm specifically includes: Defined in In the 3D search space, there are A population of particles , of which The position of each particle is represented as Its speed is expressed as The best position in each particle's individual history is denoted as ; The optimal position experienced by all particles in the entire population is recorded as the global optimal position. During each iteration, the particle's velocity and position are updated, with the velocity update as follows: ； In the formula, , They represent particles respectively In the In the nth iteration Dimensional velocity and position; Inertia weighting factor; , As a learning factor, A random number between [0, 1]; , They represent particles respectively In the In the nth iteration The individual's historical optimal position and global optimal position; For inertia weighting factor A linear decreasing weight strategy is used for calculation: ； In the formula, These are the initial inertia weights; The inertial weights are used when the maximum number of generations has been reached. and These are the maximum number of generations to be evolved in the population and the current number of generations to be evolved in the population, respectively. The particle positions are updated as follows: 。 7. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism as described in claim 6, characterized in that, In step S4, when using the golden sine and cosine optimized particle swarm optimization algorithm, the traditional linear superposition cognitive term is replaced with a nonlinear guidance mechanism based on sine and cosine functions, and a dynamic coefficient related to the golden ratio is used for adjustment; specifically, its velocity update formula is modified as follows: ； In the formula, and These are the values ​​of the sine and cosine functions, respectively. The internal random phase angle; This is a dynamic adjustment coefficient. The value is adaptively adjusted according to the golden ratio as the iteration process progresses: ； In the formula, and It is a constant related to the golden ratio; The particle position update in the Golden Sine / Cosine Optimized Particle Swarm Optimization (PSO) algorithm is the same as in the PSO algorithm: ； A golden sine and cosine optimization particle swarm optimization algorithm is introduced, using the minimum envelope entropy as the fitness function to optimize the parameter combination (K, Adaptive optimization is performed.

8. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism according to claim 1, characterized in that, The specific method for step S5 is as follows: Bidirectional Long Short-Term Memory (BSSM) networks introduce cellular states as the main thread of continuous-time memory and dynamically regulate them using three gating mechanisms: forget gate, input gate, and output gate. The mathematical model of the BSSM network is expressed as follows: ； ； ； ； ； ； In the formula, , , These represent the forget gate, input gate, and output gate respectively. Output vector at time step; and These represent the current cell state and the hidden state, respectively. , , and Here is the weight matrix corresponding to each gate unit. , , and Then it is the corresponding bias vector; and These represent the Sigmoid activation function and the hyperbolic tangent activation function, respectively.

9. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism according to claim 1, characterized in that, The multi-scenario comparison framework in step S6 is as follows: Scenario 1: The original wind power sequence is subjected to variational mode decomposition and then input into the BiLSTM model as the predicted value for wind power prediction, i.e., VMD-BiLSTM; Scenario II: The original wind power sequence is optimized by particle swarm optimization and then input into the BiLSTM model as the prediction value for wind power prediction, i.e., PSO-VMD-BiLSTM; Scenario III: The original wind power sequence is decomposed in two stages. The first stage uses CEEMDAN decomposition, and the second stage uses particle swarm optimization to optimize variational mode decomposition. The decomposed sequence is then input into the BiLSTM model as the predicted value for wind power prediction, i.e., CEEMDAN-PSO-VMD-BiLSTM. Scenario IV: The original wind power sequence is decomposed in two stages. The first stage uses ICEEMDAN decomposition, and the second stage uses the Golden Sine Cosine Optimization Particle Swarm Optimization Algorithm to optimize the variational mode decomposition. The decomposed sequence is then input into the LSTM model as the prediction value for wind power prediction, i.e., ICEEMDAN-GDPSO-VMD-LSTM. Scenario V: The original wind power sequence is decomposed in two stages. The first stage uses ICEEMDAN decomposition, and the second stage uses the Golden Sine Cosine Optimized Particle Swarm Optimization (GSO) algorithm to optimize the variational mode decomposition. The decomposed sequence is then input into the BiLSTM model as the predicted value for wind power prediction, i.e., ICEEMDAN-GDPSO-VMD-BiLSTM.

10. The wind power prediction method based on quadratic adaptive decomposition and bidirectional long short-term memory network parallelism as described in claim 9, characterized in that, The specific method for training the model and optimizing the parameters using the solver in step S6, and outputting the final wind power prediction result, is as follows: S601, Initialize prediction model parameters; S602, construct five prediction model architectures corresponding to scenarios I to V respectively, and randomly initialize the weights and bias parameters to be optimized for each model; S603, t=t+1; S604: Input the training dataset into the five scenario models respectively, and calculate the wind power prediction value for each scenario through forward propagation; S605, calculate and retain the prediction error of each scenario model in the current iteration based on the predicted value and the actual value; S606 compares the prediction error of each scenario with the previous generation's best error. If the new error is better, the optimal model parameters and optimal error for that scenario are updated; otherwise, the original optimal parameters and error are retained. S607. If all scene models have reached the maximum number of iterations T or the optimal error change is less than the convergence threshold, the program ends; otherwise, jump to S608. S608, update the model parameters according to the optimization algorithm corresponding to each scenario, and repeat S603-S607.

Images