Wind power prediction method and device based on deep learning and error correction

Abstract

The invention provides a wind power prediction method and device based on deep learning and error correction, and relates to the technical field of electric power systems and power prediction, and the method integrates a deep learning GRU model and an attention mechanism to establish an EDA-GRU model for power prediction and power error prediction. VMD and SVMD are respectively used for modal division by utilizing the fluctuation characteristic of wind power, and discrete wavelet transform (DWT) is used for capturing feature information and fusing the feature information into TimeGAN for data enhancement, so that the authenticity of the data is enhanced, the scale of the data is expanded, the problems that the fluctuation of the wind power is strong and the prediction precision is difficult to improve in wind power prediction are solved, and the prediction efficiency is improved. Compared with a traditional model CNN, GRU, LSTM and a combined model CNN-GRU, the prediction accuracy is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of power systems and power prediction technology, and in particular to a wind power prediction method and device based on deep learning and error correction. Background Art

[0002] Wind power forecasting is a key link in ensuring the safe and stable operation of the power grid and optimizing energy dispatch. Existing wind power forecasting methods are usually a combination of multiple forecasting models. While combining the advantages of each model, they also often face the characteristics of high complexity. Error correction is mostly based on statistical methods or direct prediction of error sequences. In addition, the data expansion of error sequences often lacks strong interpretability, making it difficult to clarify the causes of model results or identify key influencing factors. This not only limits the practicality and promotion of the model, but also makes it difficult to effectively utilize or correct forecast errors in power grid operation.

[0003] Traditional power forecasting methods mainly focus on using statistical methods, physical methods and their hybrid methods, such as autoregressive moving average models and exponential smoothing methods. These methods are more suitable for stable conditions, but are usually not effective when faced with highly dynamic and nonlinear wind power output. The emerging machine learning methods, such as decision trees, random forests and support vector machines, have better flexibility and accuracy for nonlinear problems, but are still unable to adapt to highly dynamic problems. Deep learning models, especially neural networks and their variants, have improved their adaptability to complex problems in wind power forecasting due to their superior time series data processing capabilities. At present, short-term power forecasting methods do not fully consider the strong volatility of the power curve and the complexity of the error causes, and the model also lacks interpretability, making it difficult to accurately predict the changing trend of wind power.

[0004] The existing solutions still cannot take into account the following issues: (1) The main influencing factor of wind power is wind speed. Due to the nonlinearity and non-stationarity of wind speed, the hidden information in its data is difficult to be utilized.

[0005] (2) Deep learning models usually lack interpretability and focus more on the prediction results of the model, but it is difficult to reasonably explain the reasons behind the results. By introducing the attention mechanism, the key time periods or characteristic variables that play a major role in power prediction can be effectively identified, thus providing important reference value for prediction analysis.

[0006] (3) Commonly used error correction methods are often relatively simple and do not fully consider the changing characteristics of the errors, the causes of the errors, and the interpretability of the error correction, which reduces the accuracy of the correction.

[0007] Accessing a high proportion of renewable energy and electronic equipment to the power grid and optimizing and regulating the operation of various energy sources are of great significance for improving the resilience of the power grid, optimizing energy results, and protecting the environment.

[0008] Therefore, how to solve the problem of strong volatility of wind power in wind power forecasting and the difficulty in improving prediction accuracy has become an urgent problem to be solved. Summary of the invention

[0009] In order to overcome the deficiencies in the background technology, the present invention provides a wind power prediction method and device based on deep learning and error correction.

[0010] In order to achieve the above-mentioned invention object, the present invention adopts the following technical scheme: In a first aspect, the present invention provides a wind power prediction method based on deep learning and error correction, comprising the following steps: S1. Collect wind power data within a preset time period, and divide the wind power data using variational mode decomposition based on a particle swarm optimization algorithm to obtain multiple sub-modal wind power data; S2, dividing the sub-modal wind power data into training set data, validation set data and test set data; S3, input the training set data into the preset EDA-GRU model for training to obtain a power prediction model, and then input the validation set data and the test set data into the power prediction model to obtain the validation set power prediction result and the test set power prediction result; S4, calculating the residual as the prediction error based on the power prediction result of the validation set and the actual wind power value of the validation set data, and obtaining the validation set prediction error sequence; S5. Divide the prediction error sequence of the validation set into multiple error subsequences through continuous variational mode decomposition, extract wavelet error features using discrete wavelet transform, and use TimeGAN to perform data expansion on each error subsequence to obtain an enhanced error sequence. S6, inputting the enhanced error sequence into the preset EDA-GRU model in step S3 for training to obtain an error prediction model, and predicting a test set prediction error sequence with the same time series as the test set power prediction result through the error prediction model; S7. Construct an error correction coefficient through the test set prediction error to correct the test set power prediction result to obtain the final prediction result.

[0011] Specifically, step S1 divides the wind power data using variational modal decomposition based on the particle swarm optimization algorithm, specifically: using a combined objective function to optimize the key parameters of variational modal decomposition; the key parameters of variational modal decomposition include the number of modes, decomposition bandwidth and penalty factor; the combined objective function is shown in formula (1): , Among them, MSE is the function of minimizing the reconstruction error; ST is the function of maximizing the signal smoothness; POK is the mode number penalty function, which is used to limit the mode number K within a reasonable range; To minimize the reconstruction error weight factor, To maximize the signal smoothness weight factor, is the modal number penalty weight factor, which is used to control the importance of the MSE function, ST function and POK function in the combined objective function; α is the decomposition bandwidth, and β is the penalty factor.

[0012] Specifically, the preset EDA-GRU model is obtained by establishing an encoder-decoder structure and combining a GRU model with an attention mechanism, wherein the encoder is composed of three layers of GRU units, and then the hidden features of the input sequence are extracted, concatenated and linearly reduced to obtain a vector as the input of the attention mechanism, and then the attention mechanism calculates the attention score, and passes it to the decoder after integration; the decoder dynamically obtains context information in combination with the attention mechanism and normalizes and further processes the features, and finally generates the final prediction result through the fully connected layer.

[0013] Specifically, the vector obtained by concatenating and linearly reducing the hidden features of the wind power data is used as the input of the attention mechanism, and then the attention mechanism calculates the attention score as follows: First, concatenate the hidden states of the multi-layer GRU to form a high-dimensional vector , as shown in formula (9): , in, Represents the hidden features of the Lth layer GRU of the encoder at time step t; Concat() is the concatenation function; Secondly, the low-dimensional vector is obtained by dimensionality reduction through the linear layer , as shown in formula (10): , in, is a high-dimensional vector concatenated from the hidden states of all layers at time step t; W is the weight matrix of the linear layer, and b is the bias term; yes Low-dimensional vector after linear dimensionality reduction; Then, calculate the attention score , as shown in formula (11): , in, is the decay weight based on the time step distance; exp() is the exponential function; i is the time step index of the input sequence; t is the current decoder time step; is an adaptive attenuation coefficient to control the long-distance influence, ; is the hidden feature of the decoder at time step t and the encoder hidden state at time step i The dot product of .

[0014] Specifically, the adaptive attenuation coefficient The calculation formula is shown in formula (12): , in, is the weight matrix; is the bias vector to improve dynamics; the softplus function outputs a non-negative value to ensure , meeting the requirement of attenuation coefficient, as shown in formula (13): .

[0015] Specifically, the decoder combines the attention mechanism to dynamically obtain context information and normalizes the further processed features, and finally generates the final prediction result through the fully connected layer as follows: First, normalize the attention scores of all time steps to get the attention weights , as shown in formula (14): , Where T is the length of the input sequence; Based on attention weight , calculate the context vector , as shown in formula (15): , in, is the hidden state of the encoder at time step t; The context vector It will be used as the input of the decoder to generate the final prediction result.

[0016] Specifically, step S5 specifically includes the following steps: S51, dividing the prediction error sequence of the validation set into multiple error subsequences by continuous variational mode decomposition; S52, performing data normalization on each error subsequence to obtain a normalized error subsequence, then using discrete wavelet transform to decompose the normalized error subsequence to obtain low-frequency components and high-frequency components, and obtaining wavelet error features of each error subsequence based on the low-frequency components and the high-frequency components; S53, generating similar error features of the wavelet error features through TimeGAN; S54, restoring similar error features by inverse wavelet transform to obtain a normalized enhanced error sequence; S55. Denormalize the normalized enhanced error sequence to obtain an enhanced error sequence, and finally obtain an enhanced error sequence of each error subsequence.

[0017] Specifically, step S7 is as follows: the test set power prediction result is corrected by formula (23): , in, is the corrected predicted value corresponding to time point i, in kilowatts; is the preliminary predicted wind power value corresponding to time point i, in kilowatts; is the error correction coefficient corresponding to time point i, The value range of is (0, 1); The error correction coefficient is composed of a correction coefficient and a bias coefficient, as shown in formula (24): , in, is the error correction coefficient corresponding to time point i; is the correction coefficient corresponding to time point i; is the bias coefficient corresponding to time point i; Correction factor Specifically, as shown in formula (25): , in, represents the actual wind power value at time point i; represents the preliminary predicted wind power value at time point i; Represents the prediction error of the test set at time point i; Bias coefficient Specifically, as shown in formula (26): , in, The gust wind speed at time point i, in meters per second, is the wind speed at time point i, in meters per second.

[0018] Specifically, the wind power data includes wind power generation power per hour, unit: kilowatt; wind speed, unit: meter / second; gust wind speed, unit: meter / second; the preset time period includes a first preset time period and a second preset time period; the sub-modal wind power data in the first preset time period is sorted in chronological order, with the first 80% as training set data and the last 20% as validation set data, and the sub-modal wind power data in the second preset time period is used as test set data.

[0019] In a second aspect, the present invention provides a wind power prediction device based on deep learning and error correction, comprising the following units: A decomposition optimization unit is used to collect wind power data within a preset time period, and to divide the wind power data using variational mode decomposition based on a particle swarm optimization algorithm to obtain multiple sub-modal wind power data; A data division unit, used for dividing the sub-modal wind power data into training set data, verification set data and test set data; The first prediction unit is used to input the training set data into a preset EDA-GRU model for training to obtain a power prediction model, and then input the validation set data and the test set data into the power prediction model to obtain a validation set power prediction result and a test set power prediction result; An error calculation unit is used to calculate the residual as the prediction error based on the power prediction result of the verification set and the actual wind power value of the verification set data, so as to obtain a verification set prediction error sequence; The error enhancement unit is used to divide the validation set prediction error sequence into multiple error subsequences through continuous variational mode decomposition, extract wavelet error features using discrete wavelet transform, and use TimeGAN to perform data expansion on each error subsequence to obtain an enhanced error sequence; The second prediction unit is used to input the enhanced error sequence into a preset EDA-GRU model for training to obtain an error prediction model, and predict a test set prediction error sequence with the same time series as the test set power prediction result through the error prediction model; The prediction correction unit is used to construct an error correction coefficient through the test set prediction error to correct the test set power prediction result to obtain the final prediction result.

[0020] In a third aspect, the present invention also provides an electronic device comprising a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the program includes instructions for executing the steps in the method described in the first aspect.

[0021] The present invention provides a wind power prediction method and device based on deep learning and error correction. The method integrates the deep learning GRU model and the attention mechanism to establish the EDA-GRU model for power prediction and power error prediction. The volatility characteristics of wind power are used to use VMD and SVMD for modal division, and discrete wavelet transform (DWT) is used to capture feature information and integrate it into TimeGAN for data enhancement, which enhances the authenticity of the data and expands the scale of the data. The strong volatility of wind power and the difficulty in improving the prediction accuracy in wind power prediction are solved, and the prediction accuracy is greatly improved compared with the traditional models CNN, GRU, LSTM and the combined model CNN-GRU. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 is a schematic diagram of a wind power prediction method based on deep learning and error correction provided according to an embodiment of the present invention; Figure 2 It is a schematic diagram of the EDA-GRU model structure diagram provided according to an embodiment of the present invention; Figure 3 is a schematic diagram of a GRU model framework diagram used in an embodiment of the present invention; Figure 4 is a schematic diagram of a TimeGAN data enhancement framework combined with wavelet transform provided according to an embodiment of the present invention; Figure 5 is a schematic diagram of a wind power prediction device based on deep learning and error correction provided according to an embodiment of the present invention; Figure 6 It is a schematic diagram of a wind power prediction device based on deep learning and error correction provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The present invention can be explained in detail through the following embodiments. The purpose of providing the present invention is to protect all technical improvements within the scope of the present invention. In the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "front", "back", "left", "right", etc. indicating directions or positional relationships, they only correspond to the drawings of the present application for the convenience of describing the present invention, and do not indicate or imply that the device or element referred to must have a specific direction.

[0025] Embodiment 1

[0026] refer to Figure 1 , this embodiment provides a wind power prediction method based on deep learning and error correction, comprising the following steps: S1. Collect wind power data within a preset time period, and divide the wind power data using variational mode decomposition based on a particle swarm optimization algorithm to obtain multiple sub-modal wind power data; The wind power data includes hourly wind power generation, unit: kilowatt; wind speed, unit: meter / second; gust wind speed, unit: meter / second; the preset time period includes 2013-1-1 to 2016-12-31; The preset time period includes a first preset time period and a second preset time period; the first preset time period is from 2013-1-1 to 2015-12-31; the second preset time period is from 2016-1-1 to 2016-12-31; The collected data include: hourly wind power generation (unit: kilowatt), wind speed (unit: meter / second), gust wind speed (unit: meter / second), which refers to the wind speed that suddenly increases in a short period of time.

[0027] Equipment: Data acquisition system (including wind speed sensor, generator power monitoring equipment); data storage tool (saved as CSV file); data cleaning tool (Pandas library).

[0028] Specifically, a combined objective function is constructed by a particle swarm optimization algorithm to optimize the key parameters of variational modal decomposition, and then the optimized variational modal decomposition is used to divide the wind power data into multiple sub-modes; the key parameters of the variational modal decomposition include the number of modes, decomposition bandwidth and penalty factor; Particle swarm optimization (PSO) is used to find the optimal parameters of variational mode decomposition, and then the wind power data is divided into multiple sub-modal wind power data using variational mode decomposition (VMD) optimized by PSO; Variational mode decomposition (VMD) is an adaptive signal decomposition method that decomposes the original signal into several intrinsic mode functions (IMFs), i.e., sub-modes. Each sub-mode represents a different frequency component of the signal. The advantage of VMD is that it can retain more local features in signal decomposition and can effectively handle non-stationary and nonlinear signals.

[0029] Particle swarm optimization (PSO) is a group intelligence optimization algorithm based on the behavior model of bird flocks foraging in nature. It finds the optimal solution to the problem by simulating the movement of "particles" in the search space.

[0030] In this embodiment, the PSO algorithm is used to optimize the key parameters in VMD: the number of modes, the decomposition bandwidth, and the penalty factor.

[0031] The PSO algorithm used in this embodiment defines a combined objective function to optimize the key parameters in VMD. The combined objective function is shown in formula (1): , Among them, MSE is the function of minimizing the reconstruction error; ST is the function of maximizing the signal smoothness; POK is the mode number penalty function, which is used to limit the mode number K within a reasonable range; To minimize the reconstruction error weight factor, To maximize the signal smoothness weight factor, is the modal number penalty weight factor, which is used to control the importance of the MSE function, ST function and POK function in the combined objective function; α is the decomposition bandwidth, and β is the penalty factor.

[0032] Preferably, , , The value of depends on the importance of the three factors: the number of modes, the decomposition bandwidth and the penalty factor, but all three of them should be between 0 and 1. , , The values ​​are 0.4, 0.4, and 0.2 respectively; Number of modes (K): The number of modes selected during VMD decomposition will affect the granularity of the results. If the number of modes is too large, it may lead to overfitting; if the number of modes is too small, it may not be able to effectively capture the characteristics of the signal. PSO can find the optimal decomposition structure by adjusting the number of modes; its value is an integer selected in the range of [2, 10]. Preferably, K is 6 in this embodiment; Decomposition bandwidth (α): This parameter controls the bandwidth of each mode, thus affecting the frequency distribution of the mode. PSO can optimize the decomposition effect by adjusting the bandwidth to ensure that the mode can reasonably represent the frequency components of the signal; its value is a real number selected in the range of [0,10]. Preferably, α is 5 in this embodiment.

[0033] Penalty factor (β): This parameter affects the decomposition accuracy of the mode. PSO can optimize this parameter to make the decomposed mode smoother and without excessive high-frequency noise. Its value is a real number selected in the range of [0, 10]. Preferably, β is 5 in this embodiment.

[0034] VMD is optimized through PSO, and the initial data is finally divided into multiple sub-modes.

[0035] The combined objective function includes the following parts: A. Minimize reconstruction error (MSE): The mean square error between the reconstructed signal and the original signal, reflecting the quality of decomposition. B. Maximize signal smoothness (ST): By adjusting the α and β values, the frequency difference between the modes is optimized, making the signal smoother and less noisy.

[0036] C. Balanced number of modes (K): An appropriate K value helps balance the granularity of decomposition. Too many modes may lead to overfitting, while too few modes may fail to capture the details of the signal.

[0037] This embodiment reconstructs the three partial functions of the objective function as follows: a) Minimize the reconstruction error (MSE): Use the mean square error to represent the difference between the reconstructed signal and the original signal, as shown in formula (2): , in, is the original signal at time The value of is the signal obtained by variational mode decomposition (VMD) at time The reconstructed value of , N is the number of samples, α is the decomposition bandwidth, and β is the penalty factor.

[0038] The MSE function itself is a simple error metric that directly calculates the error between the original signal and the reconstructed signal. It does not require knowing the specific VMD parameters. The adjustment of VMD parameters affects the accuracy and smoothness of signal decomposition, and indirectly determines the quality of the reconstructed signal, thereby affecting the value of MSE. , , It means that MSE is related to these three parameters, but it is not the input for calculating MSE.

[0039] b) Maximizing the smoothness (ST) of the signal: This is achieved by calculating the difference between the modal spectra, as shown in formula (3): , in, is the center frequency of the nth mode, is the center frequency of the n+1th mode, K is the number of modes, and a smaller frequency difference means a higher smoothness.

[0040] The SK function is similar to the MSE function. , It means that SK is related to these three parameters, but it is not the input for calculating SK.

[0041] Smoothness calculates the difference between adjacent modal frequencies. If the modal frequencies differ greatly, the smoothness will be low, indicating that the signal decomposition is not smooth enough. By optimizing K, α and β, we hope to maximize the smoothness, that is, to minimize the frequency variation of the modal.

[0042] c) Balanced number of modes (K): This is a penalty term used to control the number of modes K to avoid selecting too many or too few modes. Too many modes may lead to overfitting, while too few modes may lead to insufficient decomposition, so a penalty mechanism is needed to ensure the rationality of the number of modes.

[0043] , in: and is a reasonable range of the number of modes, which can usually be set according to the characteristics of the signal, such as =2 and =10.

[0044] If the modal number K exceeds the predetermined range , the penalty term will become infinite, indicating that this is an unacceptable solution. Therefore, through the optimization process, PSO will avoid choosing an unreasonable number of modes.

[0045] The PSO algorithm is to In the range of , choose the maximum of are the parameters of the model.

[0046] The VMD is optimized through PSO, and the initial data (wind power data) is finally divided into multiple sub-modes.

[0047] S2, dividing the sub-modal wind power data into training set data, validation set data and test set data; The first 80% of the submodal wind power data in the first preset time period is used as the training set data, and the last 20% is used as the validation set data in chronological order, and the submodal wind power data in the second preset time period is used as the test set data; The hourly wind power generation is used for single feature prediction, a two-dimensional input matrix consisting of the time series is constructed, and the wind speed and gust wind speed are used for bias coefficient modeling in step S7.

[0048] The training set is used for model training; the validation set is used to calculate the prediction error sequence and predict the prediction error of the test set; the test set is used to finally evaluate the generalization ability of the model and measure the performance of the model in real scenarios.

[0049] S3, input the training set data into the preset EDA-GRU model for training to obtain a power prediction model, and then input the validation set data and the test set data into the power prediction model to obtain the validation set power prediction result and the test set power prediction result; The preset EDA-GRU model is obtained by establishing an encoder-decoder structure and combining a GRU model with an attention mechanism; Specifically, the preset EDA-GRU model adopts the following Figure 2 The deep learning network structure shown in the figure has an input encoder composed of three layers of GRU units, which enhances the feature extraction capability and reduces the impact of noise. The hidden features of the extracted input sequence are then concatenated and linearly reduced to obtain a vector as the input of the attention mechanism. The attention mechanism then calculates the attention score and passes it to the decoder after integration. The decoder combines the attention mechanism to dynamically obtain context information and normalizes the features for further processing, and finally generates a prediction result through the fully connected layer.

[0050] In this embodiment, the training set data obtained in step S2 is input into a preset EDA-GRU model for training to obtain a power prediction model; Input the validation set data into the power prediction model to obtain the validation set power prediction result, and input the test set data into the power prediction model to obtain the test set power prediction result; It can be understood that each sub-mode obtained by VMD decomposition of the original data represents a different frequency component of the original data, and the number of data in this sub-mode is the same as that of the original data, and adding the sub-modes obtained by VMD decomposition can accurately reconstruct the original signal. This is because one of the goals of VMD is to ensure the orthogonality and completeness of signal decomposition.

[0051] It can be seen from steps S1 and S2 that the validation set data contains multiple sub-modal wind power data, and the time series of each sub-modal wind power data is the same. Therefore, after the validation set data is input into the power prediction model for prediction, the power prediction result data of the multiple sub-modal wind power data are added together to finally obtain the validation set power prediction result; Similarly, the test set data also contains multiple sub-modal wind power data, and the time series of each sub-modal wind power data is the same. Therefore, after the test set data is input into the power prediction model for prediction, the power prediction result data of the multiple sub-modal wind power data are added together to finally obtain the test set power prediction result; The validation set power prediction results and the test set power prediction results both contain a two-dimensional matrix sequence of predicted power results corresponding to time points.

[0052] GRU (Gated Recurrent Unit) is an improved recurrent neural network model designed for processing time series data. It effectively solves the gradient vanishing problem of traditional RNN by introducing update gates and reset gates, and can capture long-term and short-term dependencies in time series. Its specific structure is as follows Figure 3 As shown, the specific internal calculation process is shown in formulas (5)-(8): , in, , , is the weight matrix corresponding to the network activation function, is the weight matrix of the update gate; To reset the gate weight matrix, is the weight matrix of the candidate hidden state; is the candidate matrix, containing all vectors that may be output from the GRU layer. and are the internal reset gate and update gate outputs of GRU, σ and ɸ are sigmoid and tanh activation functions, respectively. is the hidden state at the previous moment t, is the hidden state at the current time t, is the input at the current time t, is the candidate hidden state at the current time t.

[0053] The attention layer determines the importance of the information contained in the prediction input from the GRU and assigns different weights. The weight calculation process is as follows: First, concatenate the hidden states of the multi-layer GRU to form a high-dimensional vector , as shown in formula (9): , in, Represents the hidden features of the Lth layer GRU of the encoder at time step t; Concat() is the concatenation function; Secondly, the low-dimensional vector is obtained by dimensionality reduction through the linear layer , as shown in formula (10): , in, is a high-dimensional vector concatenated from the hidden states of all layers at time step t; W is the weight matrix of the linear layer, and b is the bias term; yes Low-dimensional vector after linear dimensionality reduction; Then, calculate the attention score , as shown in formula (11): , Among them, calculating the attention score consists of two parts: Part I: is the decay weight based on the time step distance: Where exp() is an exponential function; i is the time step index of the input sequence; t is the current decoder time step; is the adaptive attenuation coefficient, controlling the long-distance influence: Larger λ: The weights of distant time steps decrease faster, and the model pays more attention to nearby time steps; Smaller λ: distant time steps still have a greater impact, and the model focuses on a wider range; This section introduces the proximity assumption of time steps, that is, short-term dependencies in time series are more important; Adaptive attenuation coefficient The calculation method is shown in formula (12): , in, is the hidden state of the decoder at the current time step t, containing the context information of time step t; is the weight matrix; is the bias vector to improve dynamics; the softplus function outputs a non-negative value to ensure , meeting the requirement of attenuation coefficient, as shown in formula (13): ; Part II: is the hidden feature of the decoder at time step t and the encoder hidden state at time step i The dot product of (measure of similarity); Use Softmax to normalize the attention scores of all time steps to get the attention weights , as shown in formula (14): , Where T is the length of the input sequence; Based on attention weight , calculate the context vector , as shown in formula (15): , The context vector It will be used as the input of the decoder to generate the final prediction result; The input of each time step is only the output of the previous time step and the decoder's hidden state ; After introducing the context vector, the input of GRU will be expanded as shown in formula (16): , in, The output generated by the decoder for the previous time step; is the context vector of the current time step (from the attention mechanism); This embodiment uses variational mode decomposition (VMD) based on particle swarm optimization (PSO) to decompose data into multiple modes, and builds an EDA-GRU model based on Tensorflow.

[0054] S4. Calculate the residual as the prediction error based on the power prediction result of the validation set and the actual wind power value of the validation set data, and obtain the validation set prediction error sequence; the calculation formula of the residual is shown in formula (17): , in: is the prediction error at time i; is the actual wind power value at time i; is the initial predicted wind power value at time i; It can be understood that the actual wind power value at different times is subtracted from the predicted wind power value at the corresponding time, thereby obtaining the prediction error sequence of the validation set; S5. Divide the prediction error sequence of the validation set into multiple error subsequences through continuous variational mode decomposition (SVMD), extract wavelet error features using discrete wavelet transform, and use TimeGAN to perform data expansion on each error subsequence to obtain an enhanced error sequence. In order to enhance the authenticity of the data, this method uses discrete wavelet transform (DWT) to extract the characteristics of the real error data, that is, the characteristics of the prediction error sequence of the validation set, so that the error data generated by TimeGAN is more consistent with the real distribution. The data enhancement process is as follows Figure 4 shown.

[0055] S51, dividing the validation set prediction error sequence into multiple error subsequences through continuous variational mode decomposition (SVMD); Continuous Variational Mode Decomposition (SVMD) is a signal processing and data analysis method. It is mainly used to decompose complex signals into a series of mode functions, each of which represents a specific frequency component in the signal.

[0056] The basic principle of SVMD is to decompose the signal into multiple modal functions by variational mode decomposition. In each iteration step, SVMD updates the modal function by minimizing the difference between the signal and the modal function. This process is repeated until convergence. The key feature of SVMD is successive decomposition, that is, in each iteration step, a main frequency component is extracted from the signal and removed from the signal until all frequency components are extracted.

[0057] In this embodiment, the prediction error sequence will obtain multiple error subsequences (modal functions) after continuous variational mode decomposition (SVMD); S52, performing data normalization on each error subsequence to obtain a normalized error subsequence, then using discrete wavelet transform to decompose the normalized error subsequence to obtain low-frequency components and high-frequency components, and obtaining wavelet error features of each error subsequence based on the low-frequency components and the high-frequency components; The data of each error subsequence is normalized, as shown in formula (18): , Where X is the error data in the error subsequence (which may contain values ​​of different ranges), is the minimum value in the error subsequence, is the maximum value in the error subsequence, is the normalized error data; Each normalized error subsequence is decomposed using discrete wavelet transform (DWT), as shown in formula (19): , in, is the validation set error sequence The value of the mth sample at time step t; It is a low-frequency component, reflecting the overall trend of the error; is the high-frequency component of the jth layer, reflecting the detailed changes of the error; j is the number of layers; Calculate wavelet error characteristics As shown in formula (20): , in, is the low frequency component, is the high-frequency component of the jth layer, reflecting the detailed changes of the error; j is the number of layers; S53, generating similar error features of the wavelet error features through TimeGAN; TimeGAN is a deep learning model for generating time series data. It combines the ideas of generative adversarial networks (GAN) and self-supervised learning to generate high-quality data with real time series characteristics. TimeGAN retains the complex relationship between time dependencies and features in the generated time series data.

[0058] Since it is a pattern generation, the training goal of TimeGAN is to generate and input features. Similar error characteristics with similar distributions .

[0059] S54, restoring similar error features by inverse wavelet transform to obtain a normalized enhanced error sequence; Therefore, the inverse wavelet transform (IDWT) is used to restore the generated similar error features to a normalized enhanced error sequence. The inverse wavelet transform is shown in formula (21): , in, is the original error sequence after inverse wavelet transform The value of the mth sample at time step t; It is the low-frequency component after inverse wavelet transform, reflecting the overall trend of the error; is the high-frequency component of the jth layer after inverse wavelet transform, reflecting the detailed changes of the error; S55, denormalizing the normalized enhanced error sequence to obtain an enhanced error sequence, and finally obtaining an enhanced error sequence of each error subsequence; The normalized enhanced error sequence is denormalized to obtain generated error data, namely, an enhanced error sequence, and then an enhanced error sequence of each error subsequence is obtained.

[0060] The anti-normalization formula is shown in formula (22): , in, is the error data obtained after normalization (i.e., the data generated by the model), is the minimum value in the error subsequence, is the maximum value in the error subsequence, It is the error data after denormalization, that is, the data restored to the original data range.

[0061] In this embodiment, the residual sequence is decomposed using continuous variational mode decomposition (SVMD), wavelet features are extracted using the PyWavelets library, and the data is expanded based on TimeGAN to twice the original data. The tool used is PyCharm.

[0062] S6, inputting the enhanced error sequence into the preset EDA-GRU model in step S3 for training to obtain an error prediction model, and predicting a test set prediction error sequence with the same time series as the test set power prediction result through the error prediction model; Specifically, the error values ​​at the same time point in multiple enhanced error sequences are averaged to reconstruct a new enhanced error sequence, and the matrix constructed by the new enhanced error sequence and its corresponding time point is used as the input of the EDA-GRU model to train the error prediction model, which can predict and output power error values ​​at different time points.

[0063] The enhanced error sequence is composed of the original error sequence and the error sequence expanded by TimeGAN. The time series of the enhanced error sequence of each error subsequence is the same. After TimeGAN expansion, one time point corresponds to multiple error values. The average of multiple error values ​​is taken as the error at that time point. Then, a matrix is ​​constructed using the error and its corresponding time point as the input of the preset EDA-GRU model, and the error prediction model is obtained by training. By inputting the time point into the error prediction model, the prediction error of the corresponding time point can be obtained.

[0064] In step S3, the test set power prediction result is obtained as a two-dimensional matrix sequence containing time points and predicted power results corresponding to time points. The time series of the test set power prediction results is extracted and input into the power prediction model to obtain the test set prediction error sequence.

[0065] The structure of the preset EDA-GRU model in step S6 is consistent with the EDA-GRU model structure used in step S3; S7. Construct an error correction coefficient through the test set prediction error to correct the test set power prediction result to obtain the final prediction result.

[0066] Specifically, the power prediction result of the test set is corrected by formula (23): , in, is the corrected predicted value corresponding to time point i (kW); is the preliminary predicted wind power value corresponding to time point i (kW); is the error correction coefficient corresponding to time point i, The value range of is (0, 1); This embodiment uses an error correction coefficient to correct the prediction result. The error correction coefficient is composed of a correction coefficient and a bias coefficient, as shown in formula (24): , in, is the error correction coefficient corresponding to time point i; is the correction coefficient corresponding to time point i; is the bias coefficient corresponding to time point i; Correction factor Specifically, as shown in formula (25): , in, represents the actual wind power value at time point i; represents the preliminary predicted wind power value at time point i; Represents the prediction error of the test set at time point i; Bias coefficient Specifically, as shown in formula (26): , in, The gust wind speed at time point i, in meters per second, is the wind speed at time point i, in meters per second; The traditional power prediction method is shown in Table 1:

[0067] This embodiment provides a wind power prediction method based on deep learning and error correction. Compared with traditional CNN, GRU, LSTM and their combined model CNN-GRU, this method can significantly improve the accuracy of wind power prediction. The key to this improvement lies in several aspects: First, the model combines deep learning and error correction mechanism. By performing error correction on the prediction results, it can effectively correct the prediction deviation generated in the traditional method, thereby improving the accuracy. Secondly, the combination of the GRU model and the attention mechanism enables the model to adaptively focus on important time series features, thereby improving the sensitivity to key fluctuations in wind power data. In addition, VMD and SVMD are used for modal separation to effectively extract useful information from each frequency band from complex wind power signals, reduce noise interference, and further improve the accuracy of prediction. Furthermore, discrete wavelet transform (DWT) is used for feature extraction to enhance the model's ability to capture different frequency components in the data. Finally, by using TimeGAN for data enhancement, the model can use more high-quality data for training, which not only improves the generalization ability of the model, but also makes it more stable in the face of uncertainty and volatility of wind power data.

[0068] This embodiment provides a wind power prediction method and device based on deep learning and error correction. The method integrates the deep learning GRU model and the attention mechanism to establish the EDA-GRU model for power prediction and power error prediction. The volatility characteristics of wind power are used to use VMD and SVMD for modal division, and discrete wavelet transform (DWT) is used to capture feature information and integrate it into TimeGAN for data enhancement, which enhances the authenticity of the data and expands the scale of the data. The strong volatility of wind power and the difficulty in improving the prediction accuracy in wind power prediction are solved. Compared with the traditional models CNN, GRU, LSTM and the combined model CNN-GRU, the prediction accuracy is greatly improved.

[0069] Embodiment 2

[0070] refer to Figure 5 , this embodiment provides a wind power prediction device based on deep learning and error correction, comprising the following units: A decomposition optimization unit is used to collect wind power data within a preset time period, and to divide the wind power data using variational mode decomposition based on a particle swarm optimization algorithm to obtain multiple sub-modal wind power data; A data division unit, used for dividing the sub-modal wind power data into training set data, verification set data and test set data; The first prediction unit is used to input the training set data into a preset EDA-GRU model for training to obtain a power prediction model, and then input the validation set data and the test set data into the power prediction model to obtain a validation set power prediction result and a test set power prediction result; An error calculation unit is used to calculate the residual as the prediction error based on the power prediction result of the verification set and the actual wind power value of the verification set data, so as to obtain a verification set prediction error sequence; The error enhancement unit is used to divide the validation set prediction error sequence into multiple error subsequences through continuous variational mode decomposition, extract wavelet error features using discrete wavelet transform, and use TimeGAN to perform data expansion on each error subsequence to obtain an enhanced error sequence; The second prediction unit is used to input the enhanced error sequence into a preset EDA-GRU model for training to obtain an error prediction model, and predict a test set prediction error sequence with the same time series as the test set power prediction result through the error prediction model; The prediction correction unit is used to construct an error correction coefficient based on the test set prediction error to correct the test set power prediction result to obtain the final prediction result. Embodiment 3

[0071] refer to Figure 6 , Figure 6: is a schematic diagram of the structure of a wind power prediction device based on deep learning and error correction of this embodiment. The wind power prediction device 20 based on deep learning and error correction of this embodiment includes a processor 21, a memory 22, and a computer program stored in the memory 22 and executable on the processor 21. When the processor 21 executes the computer program, the steps in the above method embodiment are implemented. Alternatively, when the processor 21 executes the computer program, the functions of each module / unit in the above device embodiments are implemented.

[0072] Exemplarily, the computer program can be divided into one or more modules / units, and the one or more modules / units are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments that can complete specific functions, and the instruction segments are used to describe the execution process of the computer program in the wind power prediction device 20 based on deep learning and error correction. For example, the computer program can be divided into the modules in Example 2. For the specific functions of each module, please refer to the working process of the device described in the above embodiment, which will not be repeated here.

[0073] The wind power prediction device 20 based on deep learning and error correction may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art may understand that the schematic diagram is merely an example of the wind power prediction device 20 based on deep learning and error correction, and does not constitute a limitation on the wind power prediction device 20 based on deep learning and error correction, and may include more or fewer components than shown in the figure, or combine certain components, or different components, for example, the wind power prediction device 20 based on deep learning and error correction may also include input and output devices, network access devices, buses, etc.

[0074] The processor 21 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor 21 is the control center of the wind power prediction device 20 based on deep learning and error correction, and uses various interfaces and lines to connect various parts of the wind power prediction device 20 based on deep learning and error correction.

[0075] The memory 22 can be used to store the computer program and / or module. The processor 21 realizes various functions of the wind power prediction device 20 based on deep learning and error correction by running or executing the computer program and / or module stored in the memory 22 and calling the data stored in the memory 22. The memory 22 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.). In addition, the memory 22 can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart MediaCard, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0076] Wherein, if the module / unit integrated in the wind power prediction device 20 based on deep learning and error correction is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor 21, the steps of the above-mentioned method embodiments can be implemented. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0077] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.

[0078] The parts of the present invention that are not described in detail are prior art. It is obvious to those skilled in the art that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention; therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and it is intended that all changes that fall within the meaning and scope of equivalent elements are included in the present invention.

Claims

1. A wind power prediction method based on deep learning and error correction, characterized in that: The specific steps include: S1. Collect wind power data within a preset time period, and divide the wind power data using variational mode decomposition based on a particle swarm optimization algorithm to obtain multiple sub-modal wind power data; S2, dividing the sub-modal wind power data into training set data, validation set data and test set data; S3, input the training set data into the preset EDA-GRU model for training to obtain a power prediction model, and then input the validation set data and the test set data into the power prediction model to obtain the validation set power prediction result and the test set power prediction result; S4, calculating the residual as the prediction error based on the power prediction result of the validation set and the actual wind power value of the validation set data, and obtaining the prediction error sequence of the validation set; S5. Divide the prediction error sequence of the validation set into multiple error subsequences through continuous variational mode decomposition, extract wavelet error features using discrete wavelet transform, and use TimeGAN to perform data expansion on each error subsequence to obtain an enhanced error sequence. S6, inputting the enhanced error sequence into the preset EDA-GRU model in step S3 for training to obtain an error prediction model, and predicting a test set prediction error sequence with the same time series as the test set power prediction result through the error prediction model; S7. Construct an error correction coefficient through the test set prediction error to correct the test set power prediction result to obtain the final prediction result.

2. The wind power prediction method according to claim 1, characterized in that: Step S1 divides the wind power data using variational modal decomposition based on the particle swarm optimization algorithm, specifically: using a combined objective function to optimize the key parameters of variational modal decomposition; the key parameters of variational modal decomposition include the number of modes, decomposition bandwidth and penalty factor; the combined objective function is shown in formula (1): ， Among them, MSE is the function of minimizing the reconstruction error; ST is the function of maximizing the signal smoothness; POK is the mode number penalty function, which is used to limit the mode number K within a reasonable range; To minimize the reconstruction error weight factor, To maximize the signal smoothness weight factor, is the modal number penalty weight factor, which is used to control the importance of the MSE function, ST function and POK function in the combined objective function; α is the decomposition bandwidth, and β is the penalty factor.

3. The wind power prediction method according to claim 1, characterized in that: The preset EDA-GRU model is obtained by establishing an encoder-decoder structure and combining a GRU model with an attention mechanism, wherein the encoder is composed of three layers of GRU units, and then the vector obtained by concatenating and linearly reducing the hidden features of the input sequence is used as the input of the attention mechanism, and then the attention mechanism calculates the attention score, and passes it to the decoder after integration; The decoder combines the attention mechanism to dynamically obtain contextual information and normalizes the features for further processing, and finally generates the final prediction result through the fully connected layer.

4. The wind power prediction method according to claim 3, characterized in that: The vector obtained by concatenating and linearly reducing the hidden features of the wind power data is used as the input of the attention mechanism, and then the attention mechanism calculates the attention score as follows: First, concatenate the hidden states of the multi-layer GRU to form a high-dimensional vector , as shown in formula (9): ， in, Represents the hidden features of the encoder's Lth layer GRU at time step t; Concat() is the concatenation function; Secondly, the low-dimensional vector is obtained by dimensionality reduction through the linear layer , as shown in formula (10): ， in, is a high-dimensional vector concatenated from the hidden states of all layers at time step t; W is the weight matrix of the linear layer, and b is the bias term; yes Low-dimensional vector after linear dimensionality reduction; Then, calculate the attention score , as shown in formula (11): ， in, is the decay weight based on the time step distance; exp() is the exponential function; i is the time step index of the input sequence; t is the current decoder time step; is an adaptive attenuation coefficient to control the long-distance influence, ; is the hidden feature of the decoder at time step t and the encoder hidden state at time step i The dot product of .

5. The wind power prediction method according to claim 4, characterized in that: The adaptive attenuation coefficient The calculation formula is shown in formula (12): ， in, is the weight matrix; is the bias vector to improve dynamics; the softplus function outputs a non-negative value to ensure , meeting the requirements of the attenuation coefficient, as shown in formula (13): 。 6. The wind power prediction method according to claim 4, characterized in that: The decoder combines the attention mechanism to dynamically obtain context information and normalizes the features for further processing, and finally generates the final prediction result through the fully connected layer: First, normalize the attention scores of all time steps to get the attention weights , as shown in formula (14): ， Where T is the length of the input sequence; Based on attention weight , calculate the context vector , as shown in formula (15): ， in, is the hidden state of the encoder at time step t; The context vector It will be used as the input of the decoder to generate the final prediction result.

7. The wind power prediction method according to claim 1, characterized in that: Step S5 specifically includes the following steps: S51, dividing the prediction error sequence of the validation set into multiple error subsequences by continuous variational mode decomposition; S52, performing data normalization on each error subsequence to obtain a normalized error subsequence, then using discrete wavelet transform to decompose the normalized error subsequence to obtain low-frequency components and high-frequency components, and obtaining wavelet error features of each error subsequence based on the low-frequency components and the high-frequency components; S53, generating similar error features of the wavelet error features through TimeGAN; S54, restoring similar error features by inverse wavelet transform to obtain a normalized enhanced error sequence; S55. Denormalize the normalized enhanced error sequence to obtain an enhanced error sequence, and finally obtain an enhanced error sequence for each error subsequence.

8. The wind power prediction method according to claim 1, characterized in that: Step S7 specifically includes: correcting the test set power prediction result by formula (23): ， in, is the corrected predicted value corresponding to time point i, in kilowatts; is the preliminary predicted wind power value corresponding to time point i, in kilowatts; is the error correction coefficient corresponding to time point i, The value range of is (0, 1); The error correction coefficient is composed of a correction coefficient and a bias coefficient, as shown in formula (24): ， in, is the error correction coefficient corresponding to time point i; is the correction coefficient corresponding to time point i; is the bias coefficient corresponding to time point i; Correction factor Specifically, as shown in formula (25): ， in, represents the actual wind power value at time point i; represents the preliminary predicted wind power value at time point i; Represents the prediction error of the test set at time point i; Bias coefficient Specifically, as shown in formula (26): ， in, The gust wind speed at time point i, in meters per second, is the wind speed at time point i, in meters per second.

9. The wind power prediction method according to claim 1, characterized in that: The wind power data includes wind power generation per hour, in kilowatts; Wind speed, unit: m / s; Gust wind speed, unit: m / s; the preset time period includes a first preset time period and a second preset time period; the preset time period includes a first preset time period and a second preset time period; the submodal wind power data in the first preset time period are sorted in chronological order, with the first 80% as training set data and the last 20% as validation set data, wherein the submodal wind power data in the second preset time period is used as test set data.

10. A wind power prediction device based on deep learning and error correction, characterized in that: The following units are included: A decomposition optimization unit is used to collect wind power data within a preset time period, and to divide the wind power data using variational mode decomposition based on a particle swarm optimization algorithm to obtain multiple sub-modal wind power data; A data division unit, used for dividing the sub-modal wind power data into training set data, verification set data and test set data; The first prediction unit is used to input the training set data into a preset EDA-GRU model for training to obtain a power prediction model, and then input the validation set data and the test set data into the power prediction model to obtain a validation set power prediction result and a test set power prediction result; An error calculation unit is used to calculate the residual as the prediction error based on the power prediction result of the verification set and the actual wind power value of the verification set data, so as to obtain a verification set prediction error sequence; The error enhancement unit is used to divide the validation set prediction error sequence into multiple error subsequences through continuous variational mode decomposition, extract wavelet error features using discrete wavelet transform, and use TimeGAN to perform data expansion on each error subsequence to obtain an enhanced error sequence; The second prediction unit is used to input the enhanced error sequence into a preset EDA-GRU model for training to obtain an error prediction model, and predict a test set prediction error sequence with the same time series as the test set power prediction result through the error prediction model; The prediction correction unit is used to construct an error correction coefficient through the test set prediction error to correct the test set power prediction result to obtain the final prediction result.

Images