Short-term wind power cluster power prediction method supporting conventional meteorological typing and extreme weather small sample expansion

Abstract

The invention discloses a short-term wind power cluster power prediction method supporting conventional meteorological typing and extreme weather small sample expansion, and the method comprises the steps: comprehensively utilizing NWP data, wind power plant operation data and other related information; the technologies such as a Granger causal relationship analysis algorithm, a hierarchical clustering method and a neural network are utilized to realize conventional meteorological scene typing and extreme weather small sample amplification, so that the defects caused by fuzzy prediction scene, low sensitivity to meteorological changes and insufficient consideration to extreme conditions in traditional wind power prediction can be overcome; and moreover, the precision of the wind power short-term prediction model can be improved, and the problems that the prediction model cannot cope with the influence of extreme weather and the power prediction result still has large deviation are solved.

Description

Technical Field

[0001] The application of the present invention belongs to the field of wind power technology, and the technical means of the present invention belongs to the field of time series data analysis and mining technology, and also belongs to the field of big data analysis and application technology, and particularly relates to a wind power prediction method based on NWP data. Background Art

[0002] Driven by the goal of building a new power system, it has become an inevitable trend to build a new power system with wind power, photovoltaic and other new energy sources as the main body. Among them, the volatility and uncertainty of wind power have brought great challenges to the operation, dispatching and control of the power system. Accurate prediction of new energy power generation can provide reliable decision-making information for the dispatching and control of power systems at different levels, thereby ensuring the safe, stable, economical and green operation of the power system.

[0003] How to establish an accurate wind power prediction model has always been a difficult problem faced by researchers. Traditional wind power prediction modeling methods generally use historical meteorological data and historical power data of local wind farms to establish single prediction or combined prediction models. This prediction mechanism has the following limitations: First, the prediction model input is a single typical meteorological factor, and the prediction information that can be mined is insufficient, and the application scenarios and limitations of traditional feature selection methods are not considered; second, due to the inertia of the meteorological system, the power of a wind farm can be correlated with wind farms in other locations in time and space. The traditional method has a single information dimension and does not effectively mine spatiotemporal correlation information; third, the traditional method does not adequately analyze the systematic errors of overestimation or underestimation caused by inaccurate numerical weather forecast data, and fails to make full use of the error law, so that there is still room for further improvement in prediction accuracy; fourth, the number of samples is short in extreme weather, and conventional wind power prediction technology cannot be applied to extreme weather scenarios.

[0004] Given that wind power is directly affected by characteristic meteorological factors such as wind speed, wind direction, air pressure, temperature, and humidity, the increasing maturity of numerical weather prediction (NWP) technology provides the necessary conditions for the development of high-precision wind power prediction methods that meet the needs of future power systems. However, this technology has not been widely used in intraday wind power prediction. Therefore, it is necessary to make full use of the meteorological data and power data contained in the wind power cluster area based on big data statistical analysis theory, deeply explore the impact of these high-dimensional meteorological data on the power of wind power clusters, effectively reduce the feature dimension, and apply it to the wind power prediction framework. At the same time, due to the overestimation or underestimation of prediction errors caused by inaccurate numerical weather forecast data, it is also necessary to establish an error classification correction model to further improve the prediction accuracy. Furthermore, with the increase in the proportion of wind power grid connection, extreme weather has a significant impact on the normal operation of wind power. However, the probability of extreme weather is low and the number of samples is small, which cannot meet the data requirements for wind power prediction. Summary of the invention

[0005] In order to overcome the above-mentioned deficiencies of the prior art, in a first aspect, the present invention provides a short-term wind power cluster power prediction method that supports conventional meteorological classification and small sample expansion of extreme weather, comprising the following steps: Construct NWP information feature set; Use Granger causality analysis algorithm to extract key meteorological factors; Hierarchical clustering method is used to identify and classify the meteorological types of local weather in the wind farms within the cluster; Establish a short-term wind power prediction model that integrates meteorological and power spatiotemporal information; Use time series generative adversarial networks to generate data that conforms to the real sequence trends under extreme weather conditions; The wind power data is modeled using variational mode decomposition, and convolutional neural network and long short-term memory neural network are used as combined prediction models; A wind power error matching model is established for the amplitude error and phase error of wind power.

[0006] Based on the above scheme, the method of constructing the NWP information feature set is as follows: in the feature set composed of the NWP features of m wind farms Select A subset of features ; Select the optimal subset with greater redundancy according to the maximum correlation principle ; Add constraints based on minimum redundancy so that The average mutual information between the included features is the smallest.

[0007] Based on the above scheme, the specific steps of using the hierarchical clustering method to identify and classify the meteorological types of wind farms in the cluster are as follows: Select the weather forecast data for a certain period of time in the future as input; Construct a feature matrix based on the reduced-dimensional conventional weather NWP data and historical power data; Use the Euclidean distance metric to characterize the matrix data; Classification modeling is performed for different meteorological modes.

[0008] Based on the above scheme, the time series generative adversarial network includes an autoencoder network composed of an embedding function and a recovery function, and an adversarial network composed of a generator and a discriminator; The embedding part of the autoencoder network composed of the embedding function and the recovery function encodes the input high-dimensional data as a latent space to obtain the most valuable feature information in the high-dimensional vector; the recovery part restores the latent space to the initial dimension, and the data reconstruction loss is Get the optimal solution of the latent space under the minimum objective; Introduce random Gaussian noise into the time series generative adversarial network and build a generator network model to generate real samples; based on real samples and generated samples, build a discriminator network structure to distinguish generated samples from real samples; The time series generative adversarial network completes the training of the generator and the discriminator in the embedding space.

[0009] Based on the above scheme, the model training is as follows: data is reconstructed in the autoencoder based on real time series data, and the embedding and reproduction functions are defined as: ; ; Defines a data type as static data and time series data Two categories; denote the embedding function and recurring function of the corresponding variables respectively; and Latent spaces corresponding to static data and time series data; and is the input data after decoding of the reproduction function; The generating function and the adversarial function are defined as: ; ; and Represent the generator function and the discriminator function respectively; Represents two types of initial noise of the generator; and These are the two data formats after passing through the generator; and The discrimination result of the corresponding data.

[0010] Based on the above scheme, joint training is specifically as follows: using data reconstruction loss Optimize the encoding and decoding of the autoencoder to generate more efficient low-dimensional potential representation of data; Introduce real multivariate data as the supervision item of the generator, by defining the supervised loss between the generator and the real data , evaluate the generator’s ability to learn latent representations that reflect temporal correlations and real data features; Defining the adversarial loss for unsupervised GANs , to implement feedback to the generator.

[0011] On the basis of the above scheme, the specific steps of the frequency division modeling are: decomposing the wind power into a low-frequency trend component and a high-frequency fluctuation component in the frequency domain by using the variational mode decomposition method; Convolutional neural network and long short-term memory neural network are used as combined prediction models. The convolutional neural network model training set takes NWP related factors as input and the high-frequency power fluctuation process as output; the long short-term memory neural network model training set takes low-frequency NWP related factors as input and the low-frequency power component as output. A short-term prediction model for wind power group based on the fusion of meteorological and power spatiotemporal information is established.

[0012] On the basis of the above scheme, the power prediction error is obtained by comparing the power prediction results of the wind power cluster with the actual wind power measured results. At the same time, the displacement characteristics of the site meteorological information and the measured power error in the time series are analyzed. Based on the nearest measured wind power, a hierarchical displacement model for the amplitude error of wind power and a sequence displacement model for the phase error are established.

[0013] In a second aspect, an electronic device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0014] A third aspect: A computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0015] Beneficial effects of the present invention: The present invention comprehensively utilizes NWP data and related information such as wind farm operation data to realize conventional meteorological scene classification and extreme weather small sample amplification. It can not only make up for the shortcomings of traditional wind power forecasting, such as fuzzy prediction scenes, low sensitivity to meteorological changes, and insufficient consideration of extreme situations, but also improve the accuracy of short-term wind power forecasting, and solve the problems that the prediction model cannot cope with the impact of extreme weather and that the power prediction results still have large deviations. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The technical roadmap used by the short-term wind power cluster power prediction method of the present invention to support conventional meteorological classification and small sample expansion of extreme weather; Figure 2 A schematic diagram of meteorological classification used in the short-term wind power cluster power prediction method supporting conventional meteorological classification and small sample expansion of extreme weather in the present invention. DETAILED DESCRIPTION

[0017] In order to make the objects, advantages and features of the present invention more obvious, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, the specific steps in a specific embodiment of the present invention are as follows: Step A: Construct the NWP information feature set based on the maximum correlation and minimum redundancy feature analysis.

[0019] Taking wind speed as an example, the NWP information of wind farms in the region is selected. The specific method is to select the feature set consisting of the NWP wind speed of m wind farms. Select A subset of features ,Right now ,and Such candidate subsets have Among these candidate subsets, the optimal subset selected according to the maximum correlation principle The characteristic variables contained in it should be consistent with the target variable regional wind power The average value of the mutual information reaches the maximum value, that is: ; In the formula yes Middle NWP wind speed and wind power The mutual information between them. The optimal subset selected by the maximum correlation principle Has greater redundancy, so There is also a large correlation between the features. On this basis, we add constraints based on minimum redundancy so that The average value of the mutual information between the included features is the smallest, that is: ; In the formula, yes Features in and The mutual information between them.

[0020] Combining the above two constraints, we can get the maximum relevance and minimum redundancy principle for feature selection: ; Assume that according to the mRMR principle, features, and the subset they constitute is , the remaining The subset composed of features is denoted as In order to get , to Find the feature, making the Features and Combination Still meets mRMR. The conditions that a feature should satisfy are: ; Step B. Use the Granger causality analysis algorithm to extract key meteorological factors and achieve data dimensionality reduction of high-dimensional meteorological information.

[0021] Granger causality test is used to test whether one set of time series is the cause of another set of time series. In this embodiment, SPSSAU is used to test the feature set Granger causality test was performed to obtain the optimal feature set of NWP. At the same time, the static information of the wind farm under investigation was combined to obtain the reduced-dimensional NWP information, which served as the data basis for subsequent meteorological classification and small sample expansion.

[0022] Step C. Use the hierarchical clustering method to identify and classify the meteorological types of the local weather of the wind farms within the cluster.

[0023] This embodiment uses the weather forecast data from 00:00 the next day to the next 24 to 72 hours as input to predict the confidence interval or probability distribution of wind power at the corresponding time. The method of the present invention takes into account the four meteorological elements (wind speed, wind direction, temperature, and air pressure), constructs a feature matrix based on the reduced dimension conventional weather NWP data and historical power data, characterizes the matrix data using the Euclidean distance metric, uses the hierarchical clustering method to identify and classify the meteorological types of wind farms in the cluster, and performs classification modeling for different meteorological modes.

[0024] Hierarchical clustering uses Euclidean distance to calculate the distance (i.e., similarity) between data points of different categories. The smaller the distance, the higher the similarity. The two data points or categories with the closest distance are combined to generate a clustering tree. The following is the calculation formula of Euclidean distance and the steps of hierarchical clustering: ; (1) Initialization: all samples are grouped into one cluster; (2) In the same cluster, calculate the distance between any two samples, find the two sample points a and b with the farthest distance, and take a and b as the centers of the two clusters; (3) Calculate the distances of the remaining sample points in the original cluster from a and b. The one closest to the center is assigned to the cluster. (4) Repeat steps (2) and (3) until the distance between the two farthest clusters is less than the threshold, or the number of clusters reaches the specified value, and then terminate the algorithm. Figure 2 shown.

[0025] Step D. For different meteorological types, short-term wind power prediction models integrating meteorological and power spatiotemporal information are established in combination with graph neural networks.

[0026] According to the meteorological classification results obtained in step C, the reduced-dimensional NWP data of different meteorological modes and the corresponding historical power data are used as input, and the wind power prediction model corresponding to the meteorological mode is established based on the characteristic matrix of the meteorological mode and combined with the graph neural network.

[0027] The prediction model is MTGNN (a graph neural network), and its framework and process are as follows: 1. Graph representation construction: First, the graph data (analyzing the correlation of high-dimensional meteorological data features to construct edges and then construct a graph) is represented as a collection of nodes and edges. Each node and edge has a feature representation vector that describes its attributes and relationships. These features can be node attribute information, such as node labels, degrees, etc.; they can also be edge attribute information, such as edge type, weight, etc.

[0028] 2. Information passing: MTGNN uses a message passing mechanism to update the representation vector of a node. By iteratively passing and aggregating the information of nodes and edges, each node can gradually obtain the feature information of the nodes and edges around it and update its own representation.

[0029] 3. Graph Convolutional Neural Network (GCN): In the process of information transmission, MTGNN uses the idea of ​​graph convolutional neural network. GCN is a neural network based on convolution operation, which shares parameters between node neighbors and updates the representation of nodes by aggregating the features of neighboring nodes.

[0030] 4. Graph self-attention mechanism: In addition to using GCN, MTGNN also introduces a graph self-attention mechanism. The self-attention mechanism can dynamically adjust the relationship between nodes, so that each node can selectively pay attention to different neighboring nodes, and better understand and simulate the local structure of the graph during information transmission.

[0031] 5. Convergence and prediction: During the information transmission process, each node and edge undergoes multiple rounds of information transmission, constantly updating and aggregating its own feature representation. Through such iterations, nodes and edges can gradually acquire and integrate the feature information of the nodes and edges around them. Generally, the graph self-attention mechanism is used to dynamically weight and aggregate the representations of nodes and edges according to their importance or mutual relationship. After obtaining the representation of the entire graph, some mapping functions or dimensionality reduction techniques can be used to map the high-dimensional graph representation to a low-dimensional feature space. This can help reduce computational overhead and extract more useful features. Finally, the obtained representation of the entire graph is used to complete the specific prediction task. In the graph regression prediction task, random forest regression (random forest regression is an integrated regression model based on multiple random decision trees. It obtains the final prediction result by averaging or weighted averaging the prediction results of each training tree.) is selected to predict the power.

[0032] Step E. Use a time series generative adversarial network to generate data that conforms to the real sequence trend under extreme weather conditions.

[0033] To address the problem of sample data shortage in extreme weather conditions, a time series generative adversarial network is used, random Gaussian noise is introduced, and a generator network model is built to generate samples that are as realistic as possible.

[0034] Based on real samples and generated samples, a discriminator network structure is built to distinguish generated samples from real samples, generate data that conforms to the real sequence trends under extreme weather conditions, and provide a data basis for neural network modeling.

[0035] The time series generative adversarial network used in the present invention includes an autoencoder network composed of an embedding function and a recovery function, and an adversarial network composed of a generator and a discriminator. The autoencoder provides a latent space for the training of the adversarial network. Through supervised learning, the dynamic time series features of real data and synthetic data are synchronized to obtain the most valuable feature information in the high-dimensional vector; the recovery function restores the latent space to the initial dimension. The two have a significant impact on data reconstruction loss. The optimal solution of the latent space is obtained under the minimum objective. Based on the fact that "even the temporal dynamics of complex systems are often driven by fewer and lower-dimensional factors of change", the generative adversarial network completes the training of the generator and discriminator in the embedding space, which can not only fit the temporal correlation of high-dimensional data, but also effectively reduce the dimensionality difficulty of adversarial model training, thus showing a better learning and training effect.

[0036] From the perspective of the model training process, firstly, data reconstruction is performed in the autoencoder based on the real time series data. The embedding and recovery functions can be defined as: ; ; In order to meet the needs of processing multiple data formats, the data type is defined as static data in the formula. S and time series data X Two categories; Then they represent the embedding function and recovery function of the corresponding variables respectively; and Latent spaces corresponding to static data and time series data; and is the input data after decoding by the recovery function.

[0037] When designing a generative adversarial network, the generating function and the adversarial function can be defined as: ; ; and Represent the generator function and the discriminator function respectively, Represents two types of initial noise of the generator, and These are the two data formats after the generator. and The discrimination result of the corresponding data.

[0038] From the perspective of joint training, we first use the data reconstruction loss The optimization of the encoding and decoding of the autoencoder is realized to generate a more efficient low-dimensional potential representation of the data; secondly, the real multivariate data (that is, the target domain data set composed of the NWP data obtained after dimensionality reduction and the historical power data) is introduced as the supervision item of the generator, and the supervised loss between the generator and the real data is defined. , evaluate the generator's ability to learn the potential representations of temporal correlations and real data features; finally define the adversarial loss of unsupervised GAN , to achieve feedback to the generator. Based on the joint training of each network to minimize the three errors, the model completes the learning of the sequence correlation in the embedding space, thereby generating generated data that conforms to the real time series distribution.

[0039] Step F. Use variational mode decomposition to perform frequency division modeling on wind power data, and use convolutional neural network and long short-term memory neural network as a combined prediction model.

[0040] The wind power data is modeled by frequency division using time-frequency decomposition, and the wind power is decomposed into low-frequency trend components and high-frequency fluctuation components in the frequency domain using variational mode decomposition. Convolutional neural network (CNN) and long short-term memory neural network (LSTM) are used as combined prediction models. The CNN model training set takes NWP-related factors as input and power high-frequency fluctuation process as output. The LSTM model training set takes low-frequency NWP-related factors as input and power low-frequency components as output, and establishes a short-term prediction model for wind power groups that integrates meteorological-power spatiotemporal information.

[0041] Step G: Analyze the displacement characteristics of the site meteorological information and the measured power error in the time series, and establish a wind power error matching model for the amplitude error and phase error of the wind power.

[0042] The power prediction error is obtained by comparing the power prediction results of the wind power cluster with the measured wind power results. At the same time, the displacement characteristics of the site meteorological information and the measured power error in the time series are analyzed. Based on the nearest measured wind power, a hierarchical displacement model for the amplitude error of wind power and a sequence displacement model for the phase error are established.

[0043] The error matching model in the present invention takes the wind power prediction value of the prediction model and the wind power actual value as input, outputs a vector error correction result and feeds it back to the power prediction model to correct the prediction error and improve the accuracy of wind power prediction.

[0044] The above implementation modes are only used to illustrate the present invention, but not to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the essence and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims. The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A short-term wind power cluster power prediction method that supports conventional meteorological classification and small sample expansion of extreme weather, characterized in that: The steps include: Construct NWP information feature set; Use Granger causality analysis algorithm to extract key meteorological factors; Hierarchical clustering method is used to identify and classify the meteorological types of local weather in the wind farms within the cluster; Establish a short-term wind power prediction model that integrates meteorological and power spatiotemporal information; Use time series generative adversarial networks to generate data that conforms to the real sequence trends under extreme weather conditions; The wind power data is modeled using variational mode decomposition, and convolutional neural network and long short-term memory neural network are used as combined prediction models; A wind power error matching model is established for the amplitude error and phase error of wind power.

2. The prediction method according to claim 1, characterized in that: The method of constructing the NWP information feature set is as follows: in the feature set consisting of the NWP features of m wind farms Select A subset of features ; Select the optimal subset with greater redundancy according to the maximum correlation principle ; Add constraints based on minimum redundancy so that The average mutual information between the included features is the smallest.

3. The prediction method according to claim 1, characterized in that: The specific steps of using the hierarchical clustering method to identify and classify the meteorological types of wind farms in the cluster are as follows: Select the weather forecast data for a certain period of time in the future as input; Construct a feature matrix based on the reduced-dimensional conventional weather NWP data and historical power data; Use the Euclidean distance metric to characterize the matrix data; Classification modeling is performed for different meteorological modes.

4. The prediction method according to claim 1, characterized in that: The time series generative adversarial network comprises an autoencoder network composed of an embedding function and a recovery function, and an adversarial network composed of a generator and a discriminator; The embedding part of the autoencoder network composed of the embedding function and the recovery function encodes the input high-dimensional data as a latent space to obtain the most valuable feature information in the high-dimensional vector; the recovery part restores the latent space to the initial dimension, and the data reconstruction loss is Get the optimal solution of the latent space under the minimum objective; Introduce random Gaussian noise into the time series generative adversarial network and build a generator network model to generate real samples; based on real samples and generated samples, build a discriminator network structure to distinguish generated samples from real samples; The time series generative adversarial network completes the training of the generator and the discriminator in the embedding space.

5. The prediction method according to claim 4, characterized in that: The specific model training is as follows: based on real time series data, data is reconstructed in the autoencoder, and the embedding and reproduction functions are defined as: ； ； Defines a data type as static data and time series data Two categories; denote the embedding function and recurring function of the corresponding variables respectively; and Latent spaces corresponding to static data and time series data; and is the input data after decoding of the reproduction function; The generating function and the adversarial function are defined as: ； ； and Represent the generator function and the discriminator function respectively; Represents two types of initial noise of the generator; and These are the two data formats after passing through the generator; and The judgment result of the corresponding data.

6. The prediction method according to claim 4, characterized in that: Joint training is specifically: using data reconstruction loss Optimize the encoding and decoding of the autoencoder to generate more efficient low-dimensional potential representation of data; Introduce real multivariate data as the supervision item of the generator, by defining the supervised loss between the generator and the real data , evaluate the generator’s ability to learn latent representations that reflect temporal correlations and real data features; Defining the adversarial loss for unsupervised GANs , to implement feedback to the generator.

7. The prediction method according to claim 1, characterized in that: The specific steps of the frequency division modeling are: decomposing the wind power into a low-frequency trend component and a high-frequency fluctuation component in the frequency domain by using a variational mode decomposition method; Convolutional neural network and long short-term memory neural network are used as combined prediction models. The convolutional neural network model training set takes NWP related factors as input and the high-frequency power fluctuation process as output; the long short-term memory neural network model training set takes low-frequency NWP related factors as input and the low-frequency power component as output. A short-term prediction model for wind power group based on the fusion of meteorological and power spatiotemporal information is established.

8. The prediction method according to claim 1, characterized in that: The power prediction error is obtained by comparing the power prediction results of the wind power cluster with the measured wind power results. At the same time, the displacement characteristics of the site meteorological information and the measured power error in the time series are analyzed. Based on the nearest measured wind power, a hierarchical displacement model for the amplitude error of wind power and a sequence displacement model for the phase error are established.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images