A wind power generation power prediction method and system

Abstract

The present invention discloses a wind power prediction method and system, belonging to the technical field of wind power generation processing. A multivariate time series prediction data set is constructed based on the historical data of the target to be predicted, and a power prediction pre-training model is obtained by training a multivariate time series prediction model based on the multivariate time series prediction data set; the power prediction pre-training model is used to predict the current parameter data of the target to be predicted to obtain the generated power of the target to be predicted within a set future time. Based on the multivariate time series prediction data set, the present invention performs excellently in time series prediction tasks in multiple fields. Experimental results show that the method proposed by the present invention can effectively capture the long-term and short-term dependencies in the multivariate time series and improve the accuracy and stability of wind power prediction by using physical constraints.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation processing, and in particular relates to a wind power generation power prediction method and system. Background Art

[0002] Accurate prediction of wind power generation is conducive to grid dispatchers to formulate reasonable power generation plans and operation modes in advance, and coordinate the peak load regulation and frequency regulation of the power system, which is of great significance for the continuous and stable operation of the power system. There are many factors that affect the size of wind power generation. In addition to stable geographical factors such as topography and landforms, there are also meteorological factors with strong randomness, including wind speed, wind direction, temperature, atmospheric pressure and humidity, which leads to strong volatility and intermittency of wind power in the time dimension. Therefore, modeling of wind power generation requires not only in-depth consideration of the fluctuation regularity brought about by the long-term series of historical power generation, but also full exploration of the correlation between multiple related variables and future power generation.

[0003] Traditional physical and statistical methods are limited in both modeling difficulty and prediction accuracy, and it is difficult to mine the complex nonlinear relationship between wind power and related variables. Traditional methods often rely on simplified assumptions and cannot effectively capture the complex patterns in actual data, resulting in low prediction accuracy. Deep learning methods have significant advantages in fitting the complex nonlinear relationship between multiple influencing factors and power data. It can explore the fluctuation law of wind power in the time dimension, so that the prediction results can achieve higher accuracy. For example, the self-attention mechanism represented by the long short-term memory (LSTM) network and the Transformer (transformer model) is widely used in the field of sequence modeling, but the single LSTM has the disadvantages of insufficient feature extraction ability and easy overfitting of the model. Although the Transformer performs well in capturing dependencies, the quadratic complexity of its attention mechanism limits its application in long-distance time series prediction; therefore, the existing methods have low accuracy in wind power prediction and high computational complexity when dealing with long-distance time dependencies. Summary of the invention

[0004] The object of the present invention is to provide a method and system for predicting wind power generation, so as to overcome the problem of low accuracy of wind power generation prediction in the prior art.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting wind power generation power comprises the following steps:

[0007] Construct a multivariate time series prediction dataset based on the historical data of the target to be predicted, and train a power prediction pre-training model for the multivariate time series prediction model based on the multivariate time series prediction dataset;

[0008] Use the power prediction pre-training model to predict the current parameter data of the target to be predicted to obtain the generated power of the target to be predicted within a set future time.

[0009] Furthermore, the data in the multivariate time series prediction dataset has time characteristics and physical characteristics, and the data in the multivariate time series prediction dataset meets the requirements of uniform time intervals and constant feature quantities.

[0010] Furthermore, the time characteristics of the data in the multivariate time series prediction dataset refer to that the data in the multivariate time series prediction dataset includes the date and time stamp when the data is collected, and the date and time stamp formats of all data collections are the same.

[0011] Furthermore, the physical characteristics of the data in the multivariate time series prediction dataset refer to the high-dimensional time series features obtained by fusing word embedding, time embedding, and space embedding, and the physical characteristics extracted from the high-dimensional time series features.

[0012] Furthermore, construct a multivariate time series prediction dataset based on the historical data of the target to be predicted, and perform cleaning and standardization processing on the data in the obtained multivariate time series prediction dataset:

[0013] Perform cleaning processing on the data in the multivariate time series prediction dataset, which specifically includes the following steps: fill the missing values filled with NULL characters in the multivariate time series prediction dataset with the mean of the two values before and after the missing interval; if both of the two values before and after are missing, search for the closest non-missing value step by step for mean calculation;

[0014] Perform standardization processing on the data in the cleaned multivariate time series prediction dataset, specifically to make the mean of the data in the multivariate time series prediction dataset 0 and the variance 1.

[0015] Furthermore, train a power prediction pre-training model for the multivariate time series prediction model based on the multivariate time series prediction dataset. The multivariate time series prediction model includes an embedding module layer, a preprocessing module layer, a feature enhancement module layer, and a prediction module layer;

[0016] The embedding module layer is used to fuse word embedding, time embedding, and space embedding for the input data to obtain high-dimensional time series features;

[0017] The preprocessing module layer uses the multi-head attention mechanism to preprocess the high-dimensional time series features output by the embedding module layer, fusing the high-dimensional features in different subspaces; then further processes the fused high-dimensional features through a non-linear activation function, summation, and normalization.

[0018] The feature enhancement module layer enhances the input sequence that has been further processed through a recursive mechanism.

[0019] The prediction module layer is used to predict the fused features after enhancement processing to obtain a prediction result.

[0020] Furthermore, the feature enhancement module layer includes a state space module and a physical space module. The state space module is used to extract the data state features of the high-dimensional features of the time series data after further processing, and the physical space module is used to extract the physical constraints of the high-dimensional features of the time series data after further processing.

[0021] In the second aspect of the present invention, a wind power prediction system includes a pre-training module and a prediction module;

[0022] The pre-training module constructs a multivariate time series prediction data set based on the historical data of the target to be predicted, and trains a multivariate time series prediction model based on the multivariate time series prediction data set to obtain a power prediction pre-training model.

[0023] The prediction module predicts the current parameter data of the target to be predicted based on the power prediction pre-training model to obtain the generated power of the target to be predicted within a set future time.

[0024] In the third aspect of the present invention, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned wind power prediction method are implemented.

[0025] In the fourth aspect of the present invention, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned wind power prediction method are implemented.

[0026] Compared with the prior art, the present invention has the following beneficial technical effects:

[0027] A wind power prediction method of the present invention constructs a multivariate time series prediction data set based on historical data of a target to be predicted, and trains a power prediction pre-training model for a multivariate time series prediction model based on the multivariate time series prediction data set; the power prediction pre-training model is used to predict current parameter data of the target to be predicted to obtain the generated power of the target to be predicted within a set future time. The present invention is based on the multivariate time series prediction data set and performs excellently in time series prediction tasks in multiple fields. Experimental results show that the method proposed by the present invention can effectively capture long-term and short-term dependencies in the multivariate time series and improve the accuracy and stability of wind power prediction by using physical constraints.

[0028] Preferably, the multivariate time series prediction model of the present invention adopts an architecture of an embedding module layer, a preprocessing module layer, and a feature enhancement module layer, which can effectively reduce the cumulative error brought by the traditional recurrent neural network prediction method, reduce the deployment and use costs. Through the feature enhancement module layer, the prediction value at each moment does not depend on the prediction results of other moments, avoiding the cumulative error caused by recursion, and at the same time reducing the risk of gradient explosion or gradient disappearance caused by the recurrent network. The input data is fused with word embedding, time embedding, and space embedding to obtain high-dimensional time series features, making the feature expression mode of the input vector richer and more stable, thereby improving the feature expression ability and enhancing the prediction accuracy.

[0029] Preferably, the present invention introduces a physical constraint loss into the loss function to further ensure the physical consistency of the prediction results. Description of the Drawings

[0030] Figure 1 It is a schematic flowchart of the wind power prediction method in the embodiment of the present invention. Detailed Embodiments

[0031] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0032] It should be noted that the terms "first", "second", etc. in the specification, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0033] As Figure 1 shown, the present invention provides a wind power generation power prediction method, which specifically includes the following steps:

[0034] S1. Construct a multivariate time series prediction data set based on the historical data of the target to be predicted, and train a multivariate time series prediction model based on the multivariate time series prediction data set to obtain a power prediction pre-training model;

[0035] S2. Use the power prediction pre-training model to predict the current parameter data of the target to be predicted to obtain the generated power of the target to be predicted within a set future time.

[0036] The current parameter data of the target to be predicted includes the longitude, latitude, height, wind speed, wind direction, temperature, atmospheric pressure and humidity of the target to be predicted.

[0037] The data in the multivariate time series prediction data set has time characteristics and physical characteristics; the data in the multivariate time series prediction data set meets the requirements of uniform time interval and constant number of features.

[0038] For the requirement that the data meets the uniform time interval, specifically: the historical data of the target to be predicted collected is time series data, that is, sorted in time order, and the time difference between adjacent data is constant; if data loss occurs in the time series data, a time stamp is added at this time node, and the NULL character (null character) is used to fill in the position where the data is missing.

[0039] For the requirement that the data meets the constant number of features, specifically: the data in the multivariate time series prediction data set has the same feature type and number of features at each moment. If the value of the data is missing at some moments, a time stamp is added at this time node, and the NULL character is used to fill in the position where the data is missing.

[0040] The time characteristics of the data in the multivariate time series prediction dataset refer to that in addition to recording the data itself, the data in the multivariate time series prediction dataset also records the date and time stamp when the data is collected, and the date and time stamp formats of all data collections are consistent.

[0041] The physical characteristics of the data in the multivariate time series prediction dataset refer to the high-dimensional time series features obtained by fusing word embedding, time embedding, and space embedding, and the physical characteristics extracted from the high-dimensional time series features; the specific physical characteristics extracted include the time rate of change, flow rate, and divergence. To improve the accuracy of wind power prediction, the data in the multivariate time series prediction dataset includes physical quantity information related to physical constraints, including longitude, latitude, altitude, wind speed, wind direction, temperature, atmospheric pressure, and humidity.

[0042] Clean and standardize the data in the multivariate time series prediction dataset obtained in the above steps:

[0043] Clean the data in the multivariate time series prediction dataset, which specifically includes the following steps: Fill the missing values filled with NULL characters in the multivariate time series prediction dataset with the mean of the two values before and after the missing interval; if both of the two values before and after are missing, search for the closest non-missing value step by step for mean calculation.

[0044] Standardize the data in the cleaned multivariate time series prediction dataset, which specifically includes the following steps:

[0045] Perform standardization processing using the following formula to make the mean of the data in the multivariate time series prediction dataset 0 and the variance 1:

[0046]

[0047] In the formula, represents the i-th data point in the multivariate time series prediction dataset, represents taking the mean of the sequence, represents calculating the standard deviation of the sequence, represents the data in the multivariate time series prediction dataset.

[0048] Train a power prediction pre-training model based on the multivariate time series prediction dataset. The multivariate time series prediction model includes an embedding module layer, a preprocessing module layer, a feature enhancement module layer, and a prediction module layer;

[0049] The embedding module layer is used to fuse word embedding, time embedding, and space embedding for the input data to obtain high-dimensional time series features; the data input to the embedding module layer is the data after standardization processing.

[0050] The embedding module layer is the fundamental part of the multivariate time series prediction model. Its function is to transform the input time series data and related context information into high-dimensional time series features, so as to effectively process and capture complex patterns and features in the data subsequently. In this application, the embedding module layer is used to perform the fusion of word embedding, time embedding, and spatial embedding, and the calculation formula is as follows.

[0051]

[0052]

[0053]

[0054]

[0055] In the formula, represents the input time series data, represents the time context information, represents the spatial information, and respectively represent one-dimensional convolution and linear transformation, , , and respectively represent the total data embedding, word embedding, time embedding, and spatial embedding obtained by calculation.

[0056] The preprocessing module layer uses the multi-head attention mechanism to preprocess the high-dimensional time series features output by the embedding module layer, fusing the high-dimensional features in different subspaces; then further processes the fused high-dimensional features through a non-linear activation function, summation, and normalization.

[0057] The multi-head self-attention mechanism can capture the correlation between different positions in the input high-dimensional time series features by parallel computing of multiple attention heads, thereby enhancing the feature extraction ability of the model.

[0058] (11)

[0059] (12)

[0060] (13)

[0061] (14)

[0062] (15)

[0063] In the formula, , and are respectively the Query, key, and value vectors of each attention head; , and are learnable weight matrices for queries, keys, and values respectively; is the output of the th attention head, represents the projection matrix of the multi-head attention output; is the scaling factor, is the dimension of the key vector, used to prevent the gradient from vanishing due to too large dot product values; represents the output of the multi-head attention mechanism.

[0064] After the output of the multi-head attention mechanism, in order to further enhance the non-linear representation and stability of features, this application introduces non-linear activation functions, summation, and normalization to enhance the output of the multi-head attention mechanism. The purpose is to enhance the effect of feature extraction, improve the expression ability and training stability of the model, and the output after enhancement processing is obtained .

[0065]

[0066] In the formula represents layer normalization processing; represents activating the output of the multi-head attention mechanism.

[0067] The feature enhancement module layer is the core component in the multivariate time series prediction model. The feature enhancement module layer enhances the input sequence that has been further processed through a recursive mechanism, capturing the temporal features in the input sequence while maintaining the consistency of the physical properties of high-dimensional features.

[0068] Specifically, the feature enhancement module layer processes the high-dimensional features preprocessed by the multi-head attention mechanism through a recursive mechanism to enhance the expression ability of feature representation; it can effectively capture local features, physical constraints, and temporal dependencies in the input data, thereby improving the prediction performance of the model.

[0069] The feature enhancement module layer includes a state space module and a physical space module. The state space module is used to extract the data state features of the high-dimensional features of the further processed time series data, and the physical space module is used to extract the physical constraints of the high-dimensional features of the further processed time series data. The physical constraints include the channel velocity field and the time divergence field. The specific calculation steps are as follows:

[0070] The specific process of the state space module for extracting the data state features of the high-dimensional features of time series data is as follows:

[0071]

[0072]

[0073]

[0074] In the formula is the predicted output of the state space module, which is the output after convolution and non-linear activation is input into the state space module to capture temporal dependencies. Among them, the input data , is the batch size, is the time dimension, is the number of high-dimensional feature channels. , and are the parameter matrices of the state space module respectively, , are respectively the hidden states at time and is the time series.

[0075] Next, the linear transformation result output by the state space module is multiplied by the original input through a residual connection to form a new feature representation, and then a linear transformation is performed again to match the dimension of the original input, obtaining the data state feature extracted by the state space module for the time series data :

[0076]

[0077] The physical space module uses the following formula to extract the physical constraints of the high-dimensional features of the time series data:

[0078]

[0079] In the physical space module, first, the enhanced feature is transposed so that the subsequent one-dimensional convolution operation can calculate the channel velocity on the feature channel axis; in the formula, represents the transpose of represents the channel velocity, represents the time difference between two time points in the time series.

[0080] After obtaining the channel velocity, the channel acceleration is calculated using the same method:

[0081]

[0082] In the formula: represents Channel speed at a moment denotes the channel speed at a moment

[0083] For the untransposed input features, the partial derivatives are calculated and summed over multiple feature channels at the same time, which is the divergence of the features on the multi-channel axis at this moment, reflecting the dilation and compression of the temporal features in the time series denoted as the divergence of the features at a moment, reflecting its dilation and compression on d channels at a moment:

[0084]

[0085] Input the channel speed and acceleration representations after convolution into the channel speed field of the physical space module to extract the speed field information in the high-dimensional feature channels :

[0086]

[0087]

[0088] In the formula, respectively denote the velocity matrices in the physical space model; denotes the hidden state of the velocity field at a moment, denotes the hidden state of the velocity field at a moment, denotes the velocity field adjustment coefficient in the physical space module;

[0089] Similarly, input the temporal divergence after convolution into the temporal divergence field of the physical space module to extract the divergence field information of the high-dimensional features at multiple moments :

[0090]

[0091]

[0092] In the formula, respectively denote the divergence matrices in the physical space model, denotes the hidden state of the divergence field at a moment; denotes the hidden state of the divergence field at a moment, denotes the divergence field adjustment coefficient in the physical space module.

[0093] Transpose the output of the temporal divergence field, perform a dot product with the channel speed field, and then obtain the physical constraint of the physical space module through a linear transformation :

[0094]

[0095] Finally, the state space and physical space obtained from the time series data through the state space module and the physical space module are fused, so that while the time series data and its physical characteristics maintain consistency constraints, the data state characteristics and physical constraints can also be fused to support subsequent loss calculation. Finally, the fused features after fusion are:

[0096]

[0097] Prediction module layer: The fused features output by the feature enhancement module layer are linearly transformed, and under the constraint of the physical loss function, they are input into the prediction layer for prediction to generate the final prediction result:

[0098]

[0099] where, is the final predicted power generation; is the high-dimensional feature matrix of the prediction module layer; is the weight matrix; is the bias term.

[0100] A loss function is set in the prediction module layer, and a physical constraint loss term is introduced into the loss function to measure the deviation between the model prediction result and the physical law. Its formula is:

[0101]

[0102] where, represents the loss function, represents the hyperparameter that adjusts the weight between the prediction error loss and the physical constraint loss, is the total number of all samples in the dataset.

[0103] In the specific training process, the multivariate time series prediction dataset is divided into a training set, a test set, and a validation set according to the ratio of 7:2:1; after each round of training, the loss is tested on the test set. When the loss increases for three consecutive rounds, the training is stopped, and the weight file with the smallest loss on the test set is stored.

[0104] Based on the power prediction pre-trained model obtained after training is completed, the test set is predicted. The test set used for prediction is independently and identically distributed with the training set used for training. The present invention shows the characteristics of accuracy and speed in practical tasks.

[0105] In the experimental verification, this application evaluates the prediction results of multiple models at the future 15-minute moment on the prediction data. The experimental results show that the method of this application performs excellently in the wind power prediction task. The specific results are shown in Table 1. In the table, respectively represent the mean absolute error, root mean square error, and coefficient of determination between the predicted value and the actual value.

[0106] Table 1 Comparison of the prediction results of the method of the present invention and the prior art at the future 15 min moment

[0107]

[0108] Table 2 Prediction results of the method of the present invention and the prior art at the future 30 min moment

[0109]

[0110] Table 3 Comparison of the prediction results of the method of the present invention and the prior art at the future 1 h moment

[0111]

[0112] Table 4 Comparison of the prediction results of the method of the present invention and the prior art at the future 2 h moment

[0113]

[0114] Table 5 Comparison of the prediction results of the method of the present invention and the prior art at the future 4 h moment

[0115]

[0116] In the research on wind power prediction, we used a variety of deep learning models for comparison, including LSTM (Long Short-Term Memory Network), GRU (Gated Recurrent Unit), BiGRU (Bidirectional Gated Recurrent Unit), Transformer (Transformer model), and Mamba (Selective Structured State Space Model). The LSTM and GRU models can effectively capture the long-term and short-term dependencies in time series data through their unique gate mechanisms. The BiGRU further enhances the ability to extract bidirectional features of the data. The Transformer model adopts the self-attention mechanism and performs excellently in processing long time series, especially in capturing global dependencies. The Mamba model is an existing improved model method that further optimizes the structure of the deep learning model and improves its performance in wind power prediction.

[0117] By comparing the experimental data in Tables 1 to 5, it can be clearly seen that the method of the present invention exhibits excellent performance in wind power prediction. Compared with other models, the method of the present invention shows higher accuracy in predictions for multiple time windows (such as the next 15 minutes, 30 minutes, 1 hour, 2 hours, and 4 hours). In terms of the mean absolute error ( ), and the root mean square error ( ), the error values of the method of the present invention are lower than those of other models in all time periods, especially with the smallest prediction errors in long time windows (such as 2 hours and 4 hours). At the same time, the coefficient of determination ( ) of the method of the present invention is close to or exceeds 0.9 in each time period, significantly higher than other models. This indicates that the method of the present invention can not only capture the long-term and short-term dependence characteristics of wind power generation, but also better adapt to the complex changes in the data, so it has higher prediction accuracy and stability in practical applications.

[0118] In one embodiment of the present invention, a wind power prediction system includes a pre-training module and a prediction module;

[0119] The pre-training module constructs a multivariate time series prediction data set based on the historical data of the target to be predicted, and trains a multivariate time series prediction model based on the multivariate time series prediction data set to obtain a power prediction pre-training model;

[0120] The prediction module predicts the current parameter data of the target to be predicted based on the power prediction pre-training model to obtain the generated power of the target to be predicted within a set future time.

[0121] In another embodiment of the present invention, the present invention also provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a terminal device for storing programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and, of course, the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space that stores the operating system of the terminal. And, in this storage space, there are also stored one or more instructions suitable for being loaded and executed by a processor, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed memory or a non-volatile memory, such as at least one disk memory.

[0122] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the power grid broadband oscillation monitoring method in the above embodiments; the one or more instructions in the computer-readable storage medium are loaded and executed by the processor to perform the following steps: obtaining a power grid signal and performing spectral analysis to obtain an initial spectral analysis result; according to the initial spectral analysis result, obtaining all oscillation monitoring signals in non-power-frequency bands of the power grid signal; selecting each oscillation detection signal to be corrected among all oscillation monitoring signals in non-power-frequency bands, where the oscillation detection signal to be corrected is an oscillation monitoring signal whose amplitude is greater than the average amplitude of all oscillation monitoring signals in non-power-frequency bands; according to each oscillation monitoring signal to be corrected and the oscillation monitoring signals within its preset frequency range, correcting the initial spectral analysis result of each oscillation monitoring signal to be corrected to obtain a corrected spectral analysis result of each oscillation monitoring signal to be corrected, and monitoring the power grid broadband oscillation according to the corrected spectral analysis result.

[0123] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-usable program code.

Claims

1. A method for predicting wind power generation, characterized in that: The following steps are involved: A multivariate time series prediction data set is constructed based on the historical data of the target to be predicted, and a multivariate time series prediction model is trained based on the multivariate time series prediction data set to obtain a power prediction pre-training model; Construct a multivariate time series prediction dataset based on the historical data of the target to be predicted, and clean and standardize the data in the obtained multivariate time series prediction dataset: The data in the multivariate time series prediction data set is cleaned, specifically including the following steps: the missing values ​​filled with NULL characters in the multivariate time series prediction data set are filled with the mean of the two values ​​before and after the missing interval; if both the two values ​​before and after are missing, the closest non-missing values ​​are searched level by level for mean calculation; Standardize the data in the cleaned multivariate time series prediction data set, specifically making the mean of the data in the multivariate time series prediction data set 0 and the variance 1; The power prediction pre-training model is used to predict the current parameter data of the target to be predicted to obtain the power generation of the target to be predicted within a set time in the future; the current parameter data of the target to be predicted includes the latitude and longitude, altitude, wind speed, wind direction, temperature, atmospheric pressure and humidity of the target to be predicted; Based on the multivariate time series prediction data set, a multivariate time series prediction model is trained to obtain a power prediction pre-trained model, wherein the multivariate time series prediction model includes an embedding module layer, a preprocessing module layer, a feature enhancement module layer, and a prediction module layer; The embedding module layer is used to fuse word embedding, time embedding and space embedding of the input data to obtain high-dimensional time series features; The preprocessing module layer uses the multi-head attention mechanism to preprocess the high-dimensional time series features output by the embedding module layer and fuse the high-dimensional features of different subspaces; then the fused high-dimensional features are further processed through nonlinear activation functions, summation, and normalization; The feature enhancement module layer enhances the input sequence for further processing through a recursive mechanism; The prediction module layer is used to predict the fusion features after enhancement processing to obtain the prediction results; The data in the multivariate time series prediction data set has time characteristics and physical characteristics, and the data in the multivariate time series prediction data set meets the requirements of uniform time intervals and constant number of features; the physical characteristics of the data in the multivariate time series prediction data set refer to high-dimensional time series features obtained by integrating word embedding, time embedding and space embedding, and physical characteristics extracted from high-dimensional time series features; the feature enhancement module layer includes a state space module and a physical space module, the state space module is used to extract data state characteristics of high-dimensional features of time series data after further processing, and the physical space module is used to extract physical constraints of high-dimensional features of time series data after further processing, and the physical constraints include channel velocity field and time divergence field; The state space module is used to extract the data state features of the high-dimensional features of time series data. The specific process is as follows: In the formula is the predicted output of the state-space module, which is the output of the convolution and nonlinear activation. Input to the state-space module to capture timing dependencies; where the input data , is the batch size, is the time dimension, is the number of high-dimensional feature channels; , and are the parameter matrices of the state-space module, , They are Moment and The hidden state of the moment, is a time series; The linear transformation result of the state space module output With the original input Through residual connection multiplication, a new feature representation is formed, and then linear transformation is performed again to obtain the data state features extracted by the state space module for the time series data. : The physical space module is used to extract the physical constraints of high-dimensional features of time series data using the following formula: In the physical space module, the enhanced features are first transposed; where, express The transpose of represents the channel speed, Represents the time difference between two time points in a time series; After obtaining the channel velocity, calculate the channel acceleration : Where: express The channel speed at the moment, express The channel speed at the moment; For the untransposed input features, the partial derivatives are calculated on multiple feature channels at the same time and added together, which is the divergence of the features on the multi-channel axis at that moment. Represented as a feature in The dispersion of time reflects its Expansion and compression at d channels at a time: The convolved channel velocity and acceleration representations are input into the channel velocity field of the physical space module to extract the velocity field information in the high-dimensional feature channel. : In the formula, They represent the velocity matrix in the physical space model respectively; express The hidden state of the velocity field at time , express The hidden state of the velocity field at time , Represents the velocity field adjustment coefficient in the physical space module; The convolved time divergence is input into the time divergence field of the physical space module to extract the divergence field information of high-dimensional features at multiple times. : In the formula, They represent the scatter matrix in the physical space model, express Hidden state of the momentary divergence field; express The hidden state of the momentary divergence field, Represents the divergence field adjustment coefficient in the physical space module; transpose the time divergence field output, multiply it by the channel velocity field, and then obtain the physical constraints of the physical space module through linear transformation : Finally, the state space and physical space obtained by the time series data through the state space module and the physical space module are fused, and the fused features for: 。 2. A method for predicting wind power generation according to claim 1, characterized in that: The time characteristics of the data in the multivariate time series prediction data set refer to that the data in the multivariate time series prediction data set includes the date and time stamp when the data is collected, and the date and time stamp format of all data collection is consistent.

3. A wind power generation power prediction system used in the wind power generation power prediction method according to claim 1, characterized in that: Includes pre-training module and prediction module; A pre-training module constructs a multivariate time series prediction data set based on the historical data of the target to be predicted, and trains the multivariate time series prediction model based on the multivariate time series prediction data set to obtain a power prediction pre-training model; Construct a multivariate time series prediction dataset based on the historical data of the target to be predicted, and clean and standardize the data in the obtained multivariate time series prediction dataset: The data in the multivariate time series prediction data set is cleaned, specifically including the following steps: the missing values ​​filled with NULL characters in the multivariate time series prediction data set are filled with the mean of the two values ​​before and after the missing interval; if both the two values ​​before and after are missing, the closest non-missing values ​​are searched level by level for mean calculation; Standardize the data in the cleaned multivariate time series prediction data set, specifically making the mean of the data in the multivariate time series prediction data set 0 and the variance 1; The prediction module predicts the current parameter data of the target to be predicted based on the power prediction pre-training model to obtain the power generation power of the target to be predicted within a set time in the future; Based on the multivariate time series prediction data set, a multivariate time series prediction model is trained to obtain a power prediction pre-trained model, wherein the multivariate time series prediction model includes an embedding module layer, a preprocessing module layer, a feature enhancement module layer, and a prediction module layer; The embedding module layer is used to fuse word embedding, time embedding and space embedding of the input data to obtain high-dimensional time series features; The preprocessing module layer uses the multi-head attention mechanism to preprocess the high-dimensional time series features output by the embedding module layer and fuse the high-dimensional features of different subspaces; then the fused high-dimensional features are further processed through nonlinear activation functions, summation, and normalization; The feature enhancement module layer enhances the input sequence for further processing through a recursive mechanism; The prediction module layer is used to predict the fusion features after enhancement processing to obtain the prediction results; The data in the multivariate time series prediction dataset has time characteristics and physical characteristics, and the data in the multivariate time series prediction dataset meets the requirements of uniform time intervals and constant number of features; the physical characteristics of the data in the multivariate time series prediction dataset refer to high-dimensional time series features obtained by integrating word embedding, time embedding and space embedding, and physical characteristics extracted from high-dimensional time series features; the feature enhancement module layer includes a state space module and a physical space module, the state space module is used to extract data state characteristics of high-dimensional features of time series data after further processing, and the physical space module is used to extract physical constraints of high-dimensional features of time series data after further processing.

4. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the wind power generation prediction method as claimed in any one of claims 1 to 2 when executing the computer program.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the wind power generation prediction method according to any one of claims 1 to 2 are implemented.

Images