Wind power prediction method and system based on vmd and entropy condition flexible network

Abstract

The application discloses a wind power prediction method and system based on VMD and entropy condition flexible network, which first acquires wind power data and meteorological characteristic data, obtains intrinsic mode function components through variational mode decomposition and calculates sample entropy; then inputs the sample entropy into an entropy encoder to generate an entropy embedding vector, splices the meteorological characteristic data and the mode components to form time sequence input features; further inputs the time sequence input features and the entropy embedding vector into an entropy condition flexible time convolution network, modulates channel attention, inflation rate mixing coefficients and residual jump coefficients through the entropy embedding vector, and outputs flexible time sequence features; then inputs the flexible time sequence features and the entropy embedding vector into an entropy condition flexible bidirectional long short-term memory network, embeds the entropy embedding vector in the gating calculation to modulate memory strength, and forms bidirectional time sequence representation; finally outputs a prediction value through a full connection layer.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, and particularly relates to a short-term wind power prediction method based on variational mode decomposition (VMD) and entropy-conditional flexible neural network. Background Technology

[0002] As an important component of clean energy, wind power has advantages such as low cost and sustainability, and plays an increasingly important role in the global energy structure transformation. The number of installed wind turbines has grown rapidly in recent years. However, due to the complex influence of meteorological factors, wind power systems are subject to uncertainty, which poses a great challenge to the safe and stable operation of the power grid.

[0003] Current wind power forecasting methods can be categorized into three types: physical models, statistical learning, and machine learning. Physical models, based on meteorological and wind turbine operating principle equations, have clear physical meaning; however, their complex modeling process limits their ability to meet ultra-short-term forecasting needs. Statistical learning methods, such as ARIMA and Support Vector Regression (SVR), perform well under stable meteorological conditions, but their adaptability to non-stationary wind speed sequences is insufficient. In recent years, deep learning models such as LSTM and TCN have made significant progress in time series forecasting; however, single models often struggle to simultaneously predict different frequency components, leading to a decline in forecasting performance.

[0004] Signal decomposition techniques provide a theoretical foundation for solving the aforementioned problems. Variational Mode Decomposition (VMD), by constructing a variational optimization framework, effectively suppresses mode aliasing and achieves more stable frequency domain decomposition. However, existing research typically builds prediction models independently for each sub-mode after VMD decomposition. This results in a lack of information sharing between sub-models, making it difficult to capture the correlation characteristics between modes. Furthermore, it ignores the differences in the intrinsic complexity of different modal components and fails to utilize prior information on modal complexity to guide the network's differentiated learning. Sample entropy (SE) can characterize the features of each mode, but how to embed entropy information as prior knowledge into the prediction model has not been fully explored. Summary of the Invention

[0005] This invention provides a wind power prediction method and system based on VMD and entropy-conditional flexible networks, addressing the problems of existing prediction techniques neglecting modal complexity differences and failing to adaptively adjust network behavior. This invention achieves differentiated adaptive modeling of multi-scale modes by fusing sample entropy priors with entropy-conditional gating mechanisms, thereby improving the accuracy and robustness of short-term wind power prediction.

[0006] According to one aspect of the present invention, a wind power prediction method based on VMD and entropy-conditional flexible networks is provided, comprising: Acquire wind power data and meteorological characteristic data; Variational mode decomposition is performed on the wind power data to obtain K intrinsic mode function components, and the sample entropy of each intrinsic mode function component is calculated. The sample entropy of each intrinsic mode function component is input into the entropy encoder to generate an entropy embedding vector; The meteorological feature data is concatenated with the intrinsic mode function components to form a time-series input feature; The temporal input features and entropy embedding vectors are jointly input into an entropy-conditional flexible temporal convolutional network. The entropy-conditional flexible temporal convolutional network generates channel attention weights using the entropy embedding vectors to perform channel weighting on the temporal input features, constructing a dual-branch causal dilated convolutional architecture. The dilation rate mixing coefficient is calculated using the entropy embedding vectors to fuse the convolutional outputs of the two branches, and the jump coefficients are generated using the entropy embedding vectors to modulate the residual jump connections, outputting flexible temporal features. The flexible temporal features and entropy embedding vectors are jointly input into the entropy-conditional flexible bidirectional long short-term memory network. The entropy-conditional flexible bidirectional long short-term memory network uses the flexible temporal features as sequence input. The entropy embedding vectors are embedded in the forget gate, input gate and output gate calculation of the forward long short-term memory network to modulate the gate strength. The entropy embedding vectors are also embedded in each gate of the reverse long short-term memory network. The final hidden states of the forward and reverse long short-term memory networks are spliced ​​together to form a bidirectional temporal representation. The bidirectional time series representation is mapped to wind power prediction values ​​through a fully connected layer.

[0007] As a further technical solution, after acquiring wind power data and meteorological characteristic data, the following steps are also included: using a Gaussian mixture model to detect outliers in wind power and meteorological data, identifying outliers through probability density estimation, and using the K-nearest neighbor regression algorithm to repair the outlier data.

[0008] As a further technical solution, the meteorological feature data is obtained by calculating the Pearson correlation coefficient between each meteorological variable and wind power and then filtering the strongly correlated meteorological features according to the correlation threshold.

[0009] As a further technical solution, the variational mode decomposition includes: establishing a processing model for variational mode decomposition, setting the number of modes, penalty factor, and convergence threshold; introducing Lagrange multipliers and a quadratic penalty term to construct an augmented Lagrange function; using the alternating direction multiplier method to iteratively update the frequency domain representation and center frequency of each mode; terminating the iteration when the mode change between adjacent iterations is less than the convergence threshold, and outputting K eigenmode function components.

[0010] As a further technical solution, after calculating the sample entropy of each intrinsic mode function component, the method further includes: constructing a K-dimensional sample entropy feature vector, where each element corresponds to a complexity measure of an intrinsic mode function component, and normalizing the sample entropy feature vector.

[0011] As a further technical solution, when constructing a dual-branch causal dilated convolutional architecture in the entropy-conditional flexible temporal convolutional network, the two branches are set with different dilation rates, and the dilation rate mixing coefficient is calculated through the entropy embedding vector to fuse the convolutional outputs of the two branches.

[0012] As a further technical solution, in the feedforward long short-term memory network of the entropy-conditional flexible bidirectional long short-term memory network, the calculation formulas for the forget gate, input gate, and output gate are as follows: , Among them, f t i t o t These are the forget gate, input gate, and output gate, respectively. t Input for the current time, h t-1 The hidden state of the previous moment, z e W is the entropy embedding vector. f W i W o V is the forward input weight matrix. f V i V o For the forward entropy embedding weight matrix, U f U i U o Let b be the forward hidden state weight matrix. f b i b o Let σ be the forward bias vector, and σ be the Sigmoid function.

[0013] According to one aspect of the present invention, a wind power prediction system based on VMD and entropy-conditional flexible networks is provided, comprising: The data acquisition module is used to acquire wind power data and meteorological characteristic data; The variational mode decomposition and entropy calculation module is used to perform variational mode decomposition on the wind power data to obtain K intrinsic mode function components, and to calculate the sample entropy of each intrinsic mode function component. The entropy coding module is used to input the sample entropy of each intrinsic mode function component into the entropy encoder to generate an entropy embedding vector; The feature splicing module is used to splice the meteorological feature data with the intrinsic mode function components to form time-series input features; The entropy-conditional flexible temporal convolutional network module is used to receive the temporal input features and the entropy embedding vector, generate channel attention weights with the entropy embedding vector to perform channel weighting on the temporal input features, construct a dual-branch causal dilated convolutional architecture, calculate the dilation rate mixing coefficient through the entropy embedding vector to fuse the convolutional outputs of the two branches, and generate jump coefficients through the entropy embedding vector to modulate the residual jump connection, outputting flexible temporal features; The entropy-conditional flexible bidirectional long short-term memory network module is used to receive the flexible temporal features and the entropy embedding vector. The flexible temporal features are used as sequence input. The entropy embedding vector is embedded in the calculation of the forget gate, input gate and output gate of the forward long short-term memory network to modulate the gate strength. The entropy embedding vector is also embedded in each gate of the reverse long short-term memory network. The final hidden states of the forward and reverse long short-term memory networks are spliced ​​together to form a bidirectional temporal representation. The output module is used to map the bidirectional time series representation into wind power prediction values ​​through a fully connected layer.

[0014] According to one aspect of the present invention, an electronic device is provided, characterized in that it comprises: At least one processor; And a memory, which stores computer programs that can run on the processor; When the processor executes the computer program, it implements the wind power prediction method of VMD and entropy-conditional flexible network.

[0015] According to one aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the wind power prediction method of VMD and entropy-conditional flexible networks.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By fusing variational mode decomposition with adaptive sample entropy grouping, the problem of neglecting the differences in modal complexity in traditional unified modeling methods is effectively solved, and the accuracy of short-term wind power prediction is significantly improved; 2. An entropy encoder, dual-branch dilated convolution, and entropy modulation dilation rate mixing and residual jump mechanism are introduced to improve the adaptive modeling capability of the temporal convolutional network; an entropy-conditionally gated bidirectional long short-term memory network is designed to enable the model to adaptively adjust the memory strength and receptive field size according to the sample entropy prior, thereby enhancing the feature representation capability of wind power prediction. 3. It avoids the problems of high training costs and insufficient information sharing caused by independent modeling of traditional multi-sub-models, and provides a reliable technical guarantee for wind power prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the method in an embodiment of the present invention; Figure 2 This is a schematic diagram of VMD in an embodiment of the present invention; Figure 3 This is a diagram of the TCN structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the BiLSTM principle in an embodiment of the present invention; Figure 5 This is a complete prediction flowchart of the hybrid model in this embodiment of the invention. Detailed Implementation

[0019] To address the problems existing in current technologies, constructing a prediction framework that can sense differences in modal complexity and adaptively adjust network behavior accordingly is a key approach to improving prediction performance.

[0020] Against this technological backdrop, there is an urgent need to propose a novel intelligent wind power prediction method that can effectively address the challenges of multi-scale fluctuations and modal complexity differences in power sequences. This method should achieve deep integration and adaptive modeling of variational mode decomposition and deep learning networks, improving the accuracy, stability, and robustness of short-term wind power prediction under non-stationary meteorological conditions. To this end, this invention proposes a short-term wind power prediction method based on VMD and entropy-conditional flexible networks. By organically combining VMD-SE (Variational Mode Decomposition-Sample Entropy) with entropy-conditional TCN-BiLSTM (Flexible Temporal Convolutional Network-Flexible Bidirectional Long Short-Term Memory Network) and globally embedding sample entropy priors, differentiated modeling and high-precision prediction are achieved. This effectively solves the problems of traditional unified modeling methods neglecting modal complexity differences and the poor adaptability of single networks.

[0021] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0023] This invention primarily leverages Variational Mode Decomposition (VMD) theory and deep learning techniques to improve the accuracy and stability of short-term wind power prediction. It proposes a short-term wind power prediction method based on VMD and an entropy-conditional flexible network. This invention fully utilizes the adaptive frequency domain decomposition of VMD and the complex quantification advantages of sample entropy. By integrating these two methods with an entropy-conditional flexible network, it achieves differentiated mode prediction. A feature extraction strategy is designed for the power sequence, effectively synergizing entropy encoding, expansion rate mixing, residual skipping, and bidirectional memory. The prediction results obtained through this invention are more accurate and stable, effectively improving the accuracy and robustness of wind power prediction.

[0024] The method provided by this invention can be implemented using computer simulation software technology. The overall system architecture diagram in an embodiment of this invention is shown below. Figure 1 As shown.

[0025] Let x be the input point, P(x) be the probability density function of the data point x, and π j The weights are Gaussian distributions, μ j ,∑ j Let K be the mean and covariance matrix of a Gaussian distribution. p denoted as the number of Gaussian distributions.

[0026] The accuracy of ultra-short-term wind power forecasts is affected by data quality. To improve data usability, a combination of Gaussian Mixture Model (GMM) and K-Nearest Neighbors (KNN) is used for data preprocessing. GMM models the data by estimating the generation probability of each data point, and is mainly responsible for identifying outliers through a probabilistic model. .

[0027] Define x q w is an outlier i Let y be the weight of the i-th nearest neighbor.i K represents the true target value of the i-th nearest neighbor. q This represents the number of outliers.

[0028] KNN is used to repair outliers marked by GMM, thereby improving data quality. , The data processed by GMM-KNN is divided into training and test sets according to a ratio.

[0029] Define a i and b i For observations with different characteristics, and Here, represents the mean of different features, and n is the sample size. Since wind power generation is affected by various factors, the Pearson correlation coefficient is used to analyze each input feature. The expression for the Pearson coefficient is as follows: .

[0030] The Pearson correlation coefficients between each meteorological variable and wind power are calculated based on the above formula. Strongly correlated meteorological features are then selected based on the correlation threshold to construct a set of input variables for the model.

[0031] The VMD schematic diagram in this embodiment of the invention is as follows: Figure 2 As shown, VMD, as an adaptive non-recursive signal decomposition method, can decompose the original signal into several intrinsic mode functions (IMFs) with specific center frequencies and finite bandwidths by constructing a constrained variational optimization problem, while ensuring that the sum of the bandwidths of the analytic signals after Hilbert transform of each mode is minimized.

[0032] Define K as the preset number of modes, u k (t) represents the k-th IMF component, ω k Let δ(t) be the corresponding center frequency, and δ(t) be the Dirac function. For convolution operations, This involves time differentiation. Given a wind power time series f(t), the core idea of ​​VMD is to solve the following constrained variational problem: , To solve the above constrained optimization problem, a Lagrange multiplier λ(t) and a quadratic penalty factor α are introduced for processing: .

[0033] Define f(ω) as the Fourier transform of the signal f(t). This is the frequency domain representation of the k-th IMF in the (n+1)-th iteration; For the frequency domain representation of the Lagrange multipliers in the nth iteration, For center frequency updates, τ is the step size for Lagrange multiplier updates.

[0034] Secondly, the Alternating Directional Multiplier Method (ADMM) is used for iterative solution in the frequency domain. In the (n+1)th iteration, the frequency domain representation update and center frequency update for each mode are as follows: , , When the condition is met When the iteration stops, This is the convergence threshold.

[0035] Sample entropy (SE) is a nonlinear dynamic metric used to quantify the complexity and irregularity of time series data. Compared to approximate entropy, it offers better consistency and noise resistance. A higher sample entropy value indicates stronger irregularity in the sequence; conversely, a lower value indicates stronger regularity. Let m be the embedding dimension, N be the time series length, r be the similarity tolerance, and A and B be the conditional probabilities of satisfying the tolerance r for dimensions m+1 and m, respectively. The expression for SE is as follows: .

[0036] After obtaining the sample entropy, perform sample entropy-driven modal adaptive grouping: First, construct a K-dimensional sample entropy feature vector, where each element corresponds to the complexity measure of an IMF component; then, normalize the sample entropy feature vector; finally, input the normalized sample entropy vector into the entropy encoder.

[0037] The TCN structure diagram in the embodiment of the present invention is as follows: Figure 3 As shown, Temporal Convolutional Networks (TCNs) model long sequence dependencies through causal convolution and dilated convolution. However, traditional TCNs use uniform convolution kernel parameters for all input features, making them unable to adapt to temporal features of different modalities with varying complexity. Therefore, this invention proposes an entropy-conditional flexible TCN, which achieves differentiated modeling of modalities by modulating the dilation rate mixing and residual skip connections of the network using sample entropy vector modulation.

[0038] W1 and W2 are defined as learnable weight matrices; b1 and b2 are bias vectors. This two-layer fully connected network extracts a high-dimensional representation of the entropy vector through nonlinear transformation, which serves as prior knowledge for subsequent flexible gating.

[0039] First, an entropy encoder is constructed, which combines K-dimensional samples into an entropy embedding vector z. e middle: .

[0040] Define w c Here, W represents the channel weight vector, ⊙ represents element-wise multiplication, and W represents the channel weight vector. ccThis is the channel attention weight matrix. Before inputting into the flexible TCN layer, the channel attention weights need to be calculated using the entropy embedding vector. The input feature X then needs to be weighted by the channels. .

[0041] After completing the above operations, a flexible TCN module can be constructed. This mainly involves the following steps: 1) Two-branch dilated convolution Define d l Where p1 and p2 are the padding sizes, and k is the kernel size, X is the base dilation rate. l-1 This is the output of layer (l-1). and For the gated outputs of branch 1 and branch 2; and This is the convolution output of branch 1 and branch 2.

[0042] Each flexible TCN layer contains two parallel dilated convolution branches. For the l-th layer, the dilation rates of the two branches are as follows: , Each branch contains convolutional branches and gated branches: .

[0043] 2) Entropy-modulated expansion rate mixing Define σ(-) as the Sigmoid function, w α These are learnable parameters. First, the mixing coefficient α is calculated using the entropy embedding vector. c : , The outputs of the two branches are calculated according to α. c Weighted fusion: .

[0044] Finally, gated linear units (GLUs) are used for activation: .

[0045] 3) Entropy-modulated residual connection Define w β Generate weights for the jump coefficients. Calculate the jump coefficient β: .

[0046] Define X res For residual connections, C out C is the number of output channels. in This represents the number of input channels. When the number of input and output channels is inconsistent, the residuals are adjusted through convolution. .

[0047] Final output X l The weighted sum of residuals and gated activations: , Parameter α c The design allows for control over the expansion rate mixing and β controls the residual ratio, thus enabling the model to adaptively adjust the receptive field size based on the sample entropy.

[0048] Long Short-Term Memory (LSTM) networks effectively mitigate the gradient vanishing problem through a gating mechanism, making this method suitable for capturing long-term dependencies in time-series data.

[0049] The schematic diagram of BiLSTM in this embodiment of the invention is as follows: Figure 4 As shown, BiLSTM simultaneously models both forward and backward temporal information, further enhancing its feature representation capabilities. This invention proposes an entropy-conditional flexible BiLSTM: incorporating the entropy embedding vector into the gating computation of LSTM to achieve adaptive memory control for modalities of different complexity.

[0050] Define f t i t o t These are the forward forget gate, input gate, and output gate, respectively; W f W i W o W c V is the forward input weight matrix; f V i V o Forward entropy embedding weight matrix; U f U i U o U c b is the forward hidden state weight matrix; f b i b o b c It is the forward bias vector; Forward candidate cell state, x t The input at time t; h t In the forward-hidden state, c t It is in the cellular state.

[0051] The principle of LSTM is mainly to control the information flow through three gating units (forget gate, input gate, and output gate) and cell state updates, thereby ensuring that the network can better learn long-range dependencies. In entropy-conditional LSTM, an entropy embedding vector z is introduced into each gating unit. e This ensures that the model in the feedforward LSTM can adaptively adjust the forgetting and remembering strength according to the complexity of the signal: .

[0052] definition , , These are the reverse forget gate, input gate, and output gate, respectively. , , , The input weight matrix is ​​the reverse input; , , The weight matrix is ​​embedded as the inverse entropy; , , , This is the reverse hidden state weight matrix; , , , It is the reverse bias vector; This represents the reverse candidate cell state. This is the reverse hidden state. This is the reverse cell state.

[0053] For the inverse LSTM, the same gating structure is used, but the state propagates backward in time: .

[0054] Define h bi The hidden state after bidirectional splicing, h T This represents the hidden state of the forward LSTM at time step T. The final hidden state is obtained after the entire sequence has been processed in reverse order by the backward LSTM. The forward and backward LSTMs use independent parameter matrices, thus avoiding information confusion caused by parameter sharing. Finally, the BiLSTM obtains a bidirectional representation by concatenating the final hidden states of the forward and backward passes. .

[0055] The training and optimization of the overall prediction model composed of TCN-BiLSTM in this invention also includes: setting network hyperparameters including the number of TCN channels, convolutional kernel size, number of BiLSTM hidden units and learning rate; and mapping BiLSTM features to wind power prediction values ​​through a fully connected output layer.

[0056] For predictive performance evaluation and comparative verification, this invention also includes: calculating three types of evaluation indicators: root mean square error, mean absolute error, and coefficient of determination; comparing performance with the baseline model; and verifying the effectiveness of the sample entropy adaptive grouping and entropy conditional modulation mechanisms.

[0057] The complete prediction flowchart of the hybrid model in this embodiment of the invention is as follows: Figure 5 As shown, in order to clarify the collaborative working mechanism of each module in the VMD-SE-flexible TCN-BiLSTM hybrid model, this invention elaborates on the ultra-short-term wind power prediction process based on sample entropy conditional modulation.

[0058] The specific steps are as follows:

[0059] 1) Signal decomposition and adaptive grouping: VMD decomposes the original power sequence into K IMFs, calculates the sample entropy of each IMF, and then groups them. The IMFs in the same group are added together to form combined features at different time scales.

[0060] 2) Entropy conditional encoding: The normalized sample entropy vector is mapped to a high-dimensional conditional embedding vector through a two-layer fully connected network. This vector is passed to all subsequent modules as global prior knowledge, representing the complexity differences of each modality.

[0061] 3) Flexible TCN temporal extraction: The grouped IMF and meteorological features are concatenated and input into the flexible TCN. The entropy embedding vector dynamically generates channel attention weights, expansion rate mixing coefficient α, and residual jump coefficient β, realizing adaptive temporal modeling for modes of different complexity.

[0062] 4) Flexible BiLSTM modeling: The output of TCN is input into a flexible BiLSTM for bidirectional time-series modeling, and the entropy embedding vector participates in all gating calculations to adaptively adjust the memory strength.

[0063] 5) Predictive output and collaborative optimization: Finally, the wind power prediction value is generated through a fully connected network.

[0064] All of the above steps can be completed in Python 3.13.

[0065] This invention achieves efficient prediction and adaptive modeling of short-term wind power through a hybrid prediction framework of VMD-SE-flexible TCN-BiLSTM, taking into account both feature extraction and modal differential learning, and provides an effective technical solution for wind farm power prediction.

[0066] This invention also provides a short-term wind power prediction system based on VMD and entropy-conditional flexible networks, used to implement the prediction steps in the above-described method embodiments. (Refer to...) Figure 1 The system architecture diagram shown includes a data acquisition module, a variational mode decomposition and entropy calculation module, an entropy encoding module, a feature splicing module, an entropy-conditional flexible temporal convolutional network module, an entropy-conditional flexible bidirectional long short-term memory network module, and an output module.

[0067] The data acquisition module is used to acquire wind power data and meteorological feature data. The variational mode decomposition and entropy calculation module is used to perform variational mode decomposition on the wind power data to obtain K intrinsic mode function components, and calculate the sample entropy of each intrinsic mode function component. The entropy encoding module is used to input the sample entropy of each intrinsic mode function component into an entropy encoder to generate an entropy embedding vector. The feature concatenation module is used to concatenate the meteorological feature data with the intrinsic mode function components to form a time-series input feature.

[0068] The entropy-conditional flexible temporal convolutional network module is used to receive the temporal input features and the entropy embedding vector, generate channel attention weights using the entropy embedding vector to perform channel weighting on the temporal input features, construct a dual-branch causal dilated convolutional architecture, calculate the dilation rate mixing coefficient through the entropy embedding vector to fuse the convolutional outputs of the two branches, and generate jump coefficients through the entropy embedding vector to modulate the residual jump connections, outputting flexible temporal features.

[0069] The entropy-conditional flexible bidirectional long short-term memory network module is used to receive the flexible temporal features and the entropy embedding vector. Using the flexible temporal features as sequence input, the entropy embedding vector is embedded in the forget gate, input gate, and output gate calculations of the forward long short-term memory network to modulate the gating strength. The entropy embedding vector is also embedded in each gate of the reverse long short-term memory network. The final hidden states of the forward and reverse long short-term memory networks are spliced ​​together to form a bidirectional temporal representation.

[0070] The output module is used to map the bidirectional time series representation into wind power prediction values ​​through a fully connected layer.

[0071] The above modules are in accordance with Figure 5 The complete prediction flowchart shown works in tandem to sequentially complete signal decomposition and adaptive grouping, entropy conditional coding, flexible TCN time series extraction, flexible BiLSTM modeling and prediction output, and finally generate wind power prediction values.

[0072] This invention also provides an electronic device, including at least one processor and a memory. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the short-term wind power prediction method for VMD and entropy-conditional flexible networks as described in any of the above method embodiments.

[0073] Specifically, when the processor executes the computer program, it performs the following operations: acquires wind power data and meteorological feature data; performs variational mode decomposition on the wind power data to obtain K intrinsic mode function components, and calculates the sample entropy of each intrinsic mode function component; inputs the sample entropy of each intrinsic mode function component into an entropy encoder to generate an entropy embedding vector; concatenates the meteorological feature data with the intrinsic mode function components to form a time-series input feature; inputs the time-series input feature and the entropy embedding vector into an entropy conditional flexible temporal convolutional network to output a flexible time-series feature; inputs the flexible time-series feature and the entropy embedding vector into an entropy conditional flexible bidirectional long short-term memory network to form a bidirectional time-series representation; and maps the bidirectional time-series representation to a wind power prediction value through a fully connected layer.

[0074] The processor may be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), or other programmable logic devices. The memory may be a read-only memory (ROM), a random access memory (RAM), flash memory, or a hard disk, etc.

[0075] This invention also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the short-term wind power prediction method using VMD and entropy-conditional flexible networks as described in any of the above method embodiments.

[0076] The computer-readable storage medium includes, but is not limited to, non-volatile storage media such as read-only memory (ROM), random access memory (RAM), solid-state drive (SSD), optical disk, magnetic tape, and flash memory. A computer program stored on the medium contains several instructions that, when executed by a processor, cause the device to perform all the steps described in the above method embodiments.

[0077] In summary, the short-term wind power prediction method provided by this invention is suitable for solving problems such as the strong randomness and large fluctuations of wind power sequences, as well as the insufficient prediction accuracy of traditional single models. First, a sample entropy-driven variational mode adaptive grouping strategy is constructed. Variational mode decomposition is used to separate the power sequence into multi-scale intrinsic mode functions. The complexity differences of each mode are quantified by sample entropy and then grouped. Second, an entropy-conditional flexible temporal convolutional network architecture is proposed. An entropy encoder is designed to map sample entropy into a high-dimensional conditional embedding vector. A hybrid mechanism of dilation rate mixing of dual-branch dilated convolution and entropy modulation is introduced, and an entropy-modulated residual skip connection is constructed to achieve adaptive adjustment of the receptive field. Then, an entropy-conditional flexible bidirectional long short-term memory network is designed, integrating the entropy embedding vector into the calculation process of the forget gate, input gate, and output gate. Adaptive control of memory strength is achieved through forward and backward bidirectional modeling. Finally, a unified decomposition-prediction modeling framework is established, using sample entropy as global prior knowledge throughout the feature extraction and temporal modeling process. Differentiated modeling for modes of different complexity is achieved through an adaptive gating mechanism of entropy embedding. This invention can effectively improve the characterization ability of multi-scale features of wind power sequences and the adaptive learning level of the model, thereby enhancing the prediction ability of ultra-short-term power forecasting under strong random fluctuation conditions.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A wind power prediction method based on VMD and entropy-conditional flexible networks, characterized in that, include: Acquire wind power data and meteorological characteristic data; Variational mode decomposition is performed on the wind power data to obtain K intrinsic mode function components, and the sample entropy of each intrinsic mode function component is calculated. The sample entropy of each intrinsic mode function component is input into the entropy encoder to generate an entropy embedding vector; The meteorological feature data is concatenated with the intrinsic mode function components to form a time-series input feature; The temporal input features and entropy embedding vectors are jointly input into an entropy-conditional flexible temporal convolutional network. The entropy-conditional flexible temporal convolutional network generates channel attention weights using the entropy embedding vectors to perform channel weighting on the temporal input features, constructing a dual-branch causal dilated convolutional architecture. The dilation rate mixing coefficient is calculated using the entropy embedding vectors to fuse the convolutional outputs of the two branches, and the jump coefficients are generated using the entropy embedding vectors to modulate the residual jump connections, outputting flexible temporal features. The flexible temporal features and entropy embedding vectors are jointly input into the entropy-conditional flexible bidirectional long short-term memory network. The entropy-conditional flexible bidirectional long short-term memory network uses the flexible temporal features as sequence input. The entropy embedding vectors are embedded in the forget gate, input gate and output gate calculation of the forward long short-term memory network to modulate the gate strength. The entropy embedding vectors are also embedded in each gate of the reverse long short-term memory network. The final hidden states of the forward and reverse long short-term memory networks are spliced ​​together to form a bidirectional temporal representation. The bidirectional time series representation is mapped to wind power prediction values ​​through a fully connected layer.

2. The wind power prediction method based on VMD and entropy-conditional flexible networks according to claim 1, characterized in that, After acquiring wind power data and meteorological characteristic data, the process also includes: using a Gaussian mixture model to detect outliers in the wind power and meteorological data, identifying outliers through probability density estimation, and using the K-nearest neighbor regression algorithm to repair the outlier data.

3. The wind power prediction method based on VMD and entropy-conditional flexible networks according to claim 2, characterized in that, The meteorological feature data is obtained by calculating the Pearson correlation coefficient between each meteorological variable and wind power and then filtering the strongly correlated meteorological features according to the correlation threshold.

4. The wind power prediction method based on VMD and entropy-conditional flexible networks according to claim 1, characterized in that, The variational mode decomposition includes: establishing a processing model for variational mode decomposition, setting the number of modes, penalty factor, and convergence tolerance; introducing Lagrange multipliers and a quadratic penalty term to construct an augmented Lagrange function; using the alternating direction multiplier method to iteratively update the frequency domain representation and center frequency of each mode; terminating the iteration when the mode change between adjacent iterations is less than the convergence threshold, and outputting K intrinsic mode function components.

5. The wind power prediction method based on VMD and entropy-conditional flexible networks according to claim 1, characterized in that, After calculating the sample entropy of each intrinsic mode function component, the method further includes: constructing a K-dimensional sample entropy feature vector, where each element corresponds to a complexity measure of an intrinsic mode function component, and normalizing the sample entropy feature vector.

6. The wind power prediction method based on VMD and entropy-conditional flexible networks according to claim 1, characterized in that, When constructing a dual-branch causal dilated convolutional architecture in the entropy-conditional flexible temporal convolutional network, the two branches are set with different dilation rates, and the dilation rate mixing coefficient is calculated through the entropy embedding vector to fuse the convolutional outputs of the two branches.

7. The wind power prediction method based on VMD and entropy-conditional flexible networks according to claim 1, characterized in that, In the feedforward long short-term memory network of the entropy-conditional flexible bidirectional long short-term memory network, the calculation formulas for the forget gate, input gate, and output gate are as follows: ， Among them, f t i t o t These are the forget gate, input gate, and output gate, respectively. t Input for the current time, h t-1 The hidden state of the previous moment, z e W is the entropy embedding vector. f W i W o V is the forward input weight matrix. f V i V o For the forward entropy embedding weight matrix, U f U i U o Let b be the forward hidden state weight matrix. f b i b o Let σ be the forward bias vector, and σ be the Sigmoid function.

8. A wind power prediction system based on VMD and entropy-conditional flexible networks, characterized in that, include: The data acquisition module is used to acquire wind power data and meteorological characteristic data; The variational mode decomposition and entropy calculation module is used to perform variational mode decomposition on the wind power data to obtain K intrinsic mode function components, and to calculate the sample entropy of each intrinsic mode function component. The entropy coding module is used to input the sample entropy of each intrinsic mode function component into the entropy encoder to generate an entropy embedding vector; The feature splicing module is used to splice the meteorological feature data with the intrinsic mode function components to form time-series input features; The entropy-conditional flexible temporal convolutional network module is used to receive the temporal input features and the entropy embedding vector, generate channel attention weights with the entropy embedding vector to perform channel weighting on the temporal input features, construct a dual-branch causal dilated convolutional architecture, calculate the dilation rate mixing coefficient through the entropy embedding vector to fuse the convolutional outputs of the two branches, and generate jump coefficients through the entropy embedding vector to modulate the residual jump connection, outputting flexible temporal features; The entropy-conditional flexible bidirectional long short-term memory network module is used to receive the flexible temporal features and the entropy embedding vector. The flexible temporal features are used as sequence input. The entropy embedding vector is embedded in the calculation of the forget gate, input gate and output gate of the forward long short-term memory network to modulate the gate strength. The entropy embedding vector is also embedded in each gate of the reverse long short-term memory network. The final hidden states of the forward and reverse long short-term memory networks are spliced ​​together to form a bidirectional temporal representation. The output module is used to map the bidirectional time series representation into wind power prediction values ​​through a fully connected layer.

9. An electronic device, characterized in that, include: At least one processor; And a memory, which stores computer programs that can run on the processor; When the processor executes the computer program, it implements the wind power prediction method of VMD and entropy-conditional flexible network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the wind power prediction method of VMD and entropy-conditional flexible network as described in any one of claims 1 to 7.

Images