A method and system for repairing missing historical power output data of wind farms

Abstract

This invention discloses a method and system for repairing missing historical power output data of wind farms, comprising the following steps: identifying missing historical power output values ​​of wind farms and their positions in the corresponding time series of a pre-constructed historical feature database of wind farms; extracting feature data of the missing historical power output values ​​based on the missing values ​​and their positions in the corresponding time series of the historical feature database; supplementing the remaining feature data based on a feature matching mechanism; treating each feature in the supplemented feature data as a graph node; constructing a feature graph based on graph theory; training a basic model through the feature graph to establish a nonlinear mapping from graph data to wind power output values, and training a model for repairing missing historical power output data of wind farms. This invention's method for repairing missing historical power output data of wind farms will effectively improve the accuracy of missing data repair.

Description

Technical Field

[0001] This invention relates to the field of data preprocessing, specifically to a method and system for repairing missing historical power output data of wind farms. Background Technology

[0002] With the rapid development of wind power, the total installed capacity and growth rate are constantly increasing. However, the volatility and randomness of wind farm output power pose challenges to the operation and planning of power systems. To ensure the safety and stability of the power system, it is necessary to use Supervisory Data Acquisition (SCADA) systems to collect historical data to accurately predict the output power of wind farms. However, the data collected in this way may be incomplete because SCADA systems are often affected by internal and external factors, such as sensor failures, network attacks, and communication congestion. Therefore, patching up missing historical data from wind farms is of great significance for wind power forecasting.

[0003] Therefore, the repair of missing wind farm data is an effective means to improve the accuracy of wind power forecasting, and many scholars have conducted extensive and in-depth research on it. Methods for repairing missing wind farm data can be summarized into four categories: interpolation-based methods, regression-based methods, similarity-based methods, and parameter estimation-based methods. These methods each have the following problems: 1) When the scale of missing data is large, they cannot accurately and effectively repair the missing data; 2) For existing historical databases, they do not fully utilize the temporal characteristics of the data; 3) During the data repair process, they fail to effectively utilize external information such as meteorological, temporal, and geographical data, resulting in low repair accuracy; 4) They require manual parameter tuning or manual construction of probability functions, leading to high randomness. Summary of the Invention

[0004] To address the issue of low data repair accuracy in practical engineering applications amidst the increasing impact of wind power output on the power system, this invention proposes a training method, repair method, repair model training system, repair system, electronic equipment, and computer-readable storage medium for repairing missing historical output data in wind farms. This invention will effectively improve the accuracy of missing data repair.

[0005] This invention proposes a method for repairing missing historical power output data of wind farms. Compared with other existing repair methods, this method has higher repair accuracy and can cope with the situation of large-scale missing wind farm data.

[0006] The objective of this invention is achieved at least through the following technical solutions:

[0007] In a first aspect, embodiments of the present invention also provide a method for training a model to repair missing historical power output data of a wind farm, including:

[0008] Identify the missing historical power output values ​​of wind farms in a pre-built historical feature database and their positions in the corresponding time series of the database;

[0009] Based on the missing values ​​and their corresponding time series positions in the wind farm historical feature database, feature data of missing historical power output values ​​of wind farms are extracted. The remaining feature data are supplemented based on the feature matching mechanism. Each feature in the feature data after the supplementation of the remaining feature data is taken as a graph node. A feature graph is constructed based on graph theory.

[0010] The basic model is trained by the feature graph to establish a nonlinear mapping from graph data to wind power output values, and a model for repairing missing historical power output data of wind farms is trained. The basic model is built on a graph convolutional neural network framework.

[0011] As a further improvement of the present invention, the method for constructing the pre-built historical feature database of wind farms includes:

[0012] Historical power output data from wind farms is obtained from the metering system, and meteorological data of the same time and granularity, including wind speed, wind direction, humidity, temperature, air pressure, and air density, are obtained from the weather station. Missing or abnormal data are marked with preset characters, and all relevant data are merged based on time information. ,in The data characteristics of power, wind speed, wind direction, humidity, temperature, air pressure, and air density at time t are used to form a historical feature database of wind farms.

[0013] As a further improvement of the present invention, the step of identifying missing historical power output values ​​of wind farms and their positions in the corresponding time series of the pre-constructed historical feature database of wind farms includes:

[0014] Obtain missing or abnormal data pre-marked with preset characters from the wind farm historical feature database, and in conjunction with the data usage instructions of the wind farm historical feature database, identify missing values ​​of historical power output of the wind farm and mark their positions in the corresponding time series of the wind farm historical feature database;

[0015] Identifying missing historical output values ​​of wind farms involves obtaining the missing data types and corresponding missing values ​​for wind power output data; missing data types include random missing or consecutive missing.

[0016] As a further improvement of the present invention, the step of extracting feature data of missing historical power output values ​​of wind farms based on the missing values ​​and their positions in the corresponding time series of the wind farm historical feature database includes:

[0017] Meteorological data corresponding to the historical power output data of the wind farm were extracted, and wind power output values ​​at the corresponding missing power time and three time scales before and after were extracted and fused to obtain feature data:

[0018]

[0019] in, For the wind speed, wind direction, humidity, temperature, air pressure, and air density at the moment of power loss, This represents the wind power output values ​​at three time scales prior to the moment of power loss. And so on, among which This represents the wind power output value at the moment of power loss.

[0020] As a further improvement of the present invention, the supplementation of other feature data based on the feature matching mechanism includes:

[0021] Extracted feature data The dataset is divided into feature datasets corresponding to missing historical power output values ​​of wind farms. and other feature datasets Then calculate the similarity value between the two;

[0022] Arrange the calculated similarity values ​​from smallest to largest, and extract the value corresponding to the smallest similarity value. Corresponding wind power output The P is then merged into the remaining feature data. .

[0023] As a further improvement to this invention, the similarity value between the two is calculated using the Euclidean distance method, as follows:

[0024]

[0025] in They are respectively and A certain feature data in, i The corresponding feature data includes power, wind speed, wind direction, humidity, temperature, air pressure, and air density.

[0026] As a further improvement of the present invention, the step of treating each feature in the feature data after supplementing the remaining feature data as a graph node and constructing a feature graph based on graph theory includes:

[0027] Feature data Each feature in the graph is treated as a graph node, thus forming a graph data node. Each graph data node is connected in a fully connected manner to form a feature graph.

[0028] As a further improvement of the present invention, the step of training a basic model through the feature graph to establish a nonlinear mapping from graph data to wind power output values ​​includes:

[0029] By combining the attention mechanism to learn the mapping relationship between different input features and between input and output, an attention mechanism network is constructed, and different weights are assigned to each graph node. Based on the mapping relationship between node features and the power to be repaired, node feature information is mined.

[0030] By combining the GAT algorithm to automatically update the weights of graph connection edges and the feature information of aggregated graphs, a GAT network is constructed to mine feature information at the graph space level.

[0031] By combining the attention mechanism network and the GAT network, the feature graph is learned and features are extracted, and a nonlinear mapping from graph data to wind power output values ​​is established.

[0032] As a further improvement to the present invention, the step of training a basic model through the feature graph to establish a nonlinear mapping from graph data to wind power output values ​​specifically includes:

[0033] For the constructed feature data Three matrices are obtained through network training, namely: and feature data vector Multiplying each of these three matrices yields three new eigenvectors. , , Then the vector with vector Multiplication yields a new eigenvector. , will vector After normalization using the softmax function, it is then compared with the matrix. Multiplying them together yields the final new feature data vector. The normalization operation of the softmax function is as follows:

[0034]

[0035] in, For feature vectors, For the nth element, the number of elements is related to the vector. X The dimensions are consistent. The resulting new feature vector;

[0036] In the constructed feature graph, multiple GAT layers are used to aggregate information between feature graph nodes. Then, a fully connected layer is used to transform the dimensions of the features, realizing a non-linear mapping from graph data to wind power output values. The graph node information aggregation method is as follows:

[0037]

[0038]

[0039] Among them, parameters W Used to perform feature dimension transformation for each node, parameters a Used to calculate the correlation weights between nodes, || represents vector concatenation. a i，j Indicates in a The calculation obtained below i, j Weights between nodes Represents a non-linear activation function. h i For the updated i Node feature information;

[0040] The calculation formula for the fully connected layer is as follows:

[0041]

[0042] in Indicates the first k Layer feature vectors, For the first k Layer weight vector, For the first k Layer bias.

[0043] Secondly, embodiments of the present invention also provide a method for repairing missing historical power output data of wind farms, including:

[0044] Obtain the missing historical power output values ​​of the wind farm at the time point to be repaired;

[0045] Based on the missing historical power output values ​​of wind farms and their corresponding time series positions in the historical feature database of wind farms, feature data of the missing historical power output values ​​of wind farms are extracted. The remaining feature data are supplemented based on the feature matching mechanism. Each feature in the feature data after the supplementation of the remaining feature data is taken as a graph node. A feature graph is constructed based on graph theory.

[0046] The supplemented feature data and feature graph are input into the wind farm historical output missing data repair model trained by the method described above, and the output is the wind farm historical output missing data repair result.

[0047] Thirdly, embodiments of the present invention also provide a training system for a wind farm historical power output missing data repair model, comprising:

[0048] The identification module is used to identify the missing historical output values ​​of wind farms in the pre-built historical feature database of wind farms and their positions in the corresponding time series of the historical feature database of wind farms;

[0049] The construction module is used to extract feature data of missing historical output values ​​of wind farms based on the missing values ​​and their positions in the corresponding time series of the wind farm historical feature database, supplement the remaining feature data based on the feature matching mechanism, and take each feature in the feature data after supplementing the remaining feature data as a graph node, and construct a feature graph based on graph theory.

[0050] The training module is used to train a basic model through the feature graph, establish a nonlinear mapping from graph data to wind power output values, and train a model to repair missing historical power output data of wind farms. The basic model is built on a graph convolutional neural network framework.

[0051] As a further improvement of the present invention, the method for constructing the pre-built historical feature database of wind farms in the identification module includes:

[0052] Historical power output data from wind farms is obtained from the metering system, and meteorological data of the same time and granularity, including wind speed, wind direction, humidity, temperature, air pressure, and air density, are obtained from the weather station. Missing or abnormal data are marked with preset characters, and all relevant data are merged based on time information. ,in The data characteristics of power, wind speed, wind direction, humidity, temperature, air pressure, and air density at time t are used to form a historical feature database of wind farms.

[0053] As a further improvement of the present invention, the identification module, which identifies the missing historical output values ​​of wind farms and their positions in the corresponding time series of the pre-constructed historical feature database of wind farms, includes:

[0054] Obtain missing or abnormal data pre-marked with preset characters from the wind farm historical feature database, and in conjunction with the data usage instructions of the wind farm historical feature database, identify missing values ​​of historical power output of the wind farm and mark their positions in the corresponding time series of the wind farm historical feature database;

[0055] Identifying missing historical output values ​​of wind farms involves obtaining the missing data types and corresponding missing values ​​for wind power output data; missing data types include random missing or consecutive missing.

[0056] As a further improvement of the present invention, in the construction module, the step of extracting feature data of missing historical power output values ​​of wind farms based on the missing values ​​and their corresponding time series in the wind farm historical feature database includes:

[0057] Meteorological data corresponding to the historical power output data of the wind farm were extracted, and wind power output values ​​at the corresponding missing power time and three time scales before and after were extracted and fused to obtain feature data:

[0058]

[0059] in, For the wind speed, wind direction, humidity, temperature, air pressure, and air density at the moment of power loss, This represents the wind power output values ​​at three time scales prior to the moment of power loss. And so on, among which This represents the wind power output value at the moment of power loss.

[0060] As a further improvement of the present invention, the step of supplementing the remaining feature data based on the feature matching mechanism in the construction module includes:

[0061] Extracted feature data The dataset is divided into feature datasets corresponding to missing historical power output values ​​of wind farms. and other feature datasets Then calculate the similarity value between the two;

[0062] Arrange the calculated similarity values ​​from smallest to largest, and extract the value corresponding to the smallest similarity value. Corresponding wind power output The P is then merged into the remaining feature data. .

[0063] As a further improvement of the present invention, in the construction module, the step of treating each feature in the feature data after supplementing the remaining feature data as a graph node and constructing a feature graph based on graph theory includes:

[0064] Feature data Each feature in the graph is treated as a graph node, thus forming a graph data node. Each graph data node is connected in a fully connected manner to form a feature graph.

[0065] As a further improvement of the present invention, in the construction module, the step of employing a graph attention network model, combined with the constructed feature graph, to establish a nonlinear mapping from graph data to wind power output values ​​includes:

[0066] By combining the attention mechanism to learn the mapping relationship between different input features and between input and output, an attention mechanism network is constructed, and different weights are assigned to each graph node. Based on the mapping relationship between node features and the power to be repaired, node feature information is mined.

[0067] By combining the GAT algorithm to automatically update the weights of graph connection edges and the feature information of aggregated graphs, a GAT network is constructed to mine feature information at the graph space level.

[0068] By combining the attention mechanism network and the GAT network, the feature graph is learned and features are extracted, and a nonlinear mapping from graph data to wind power output values ​​is established.

[0069] Fourthly, embodiments of the present invention also provide a system for repairing missing historical power output data of wind farms, comprising:

[0070] The acquisition module is used to obtain the missing historical output values ​​of the wind farm at the time point to be repaired;

[0071] The feature construction module is used to extract feature data of missing historical power output values ​​of wind farms based on the missing historical power output values ​​of wind farms and their corresponding time series positions in the historical feature database of wind farms. It supplements the remaining feature data based on the feature matching mechanism, and uses each feature in the supplemented feature data as a graph node, and constructs a feature graph based on graph theory.

[0072] The output module is used to input the feature data and feature graph after supplementing the remaining feature data into the wind farm historical power output missing data repair model trained by the method, and output the wind farm historical power output missing data repair result.

[0073] Fifthly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the wind farm historical power output missing data repair model training method or the wind farm historical power output missing data repair method.

[0074] In a sixth aspect, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the wind farm historical power output missing data repair model training method or the wind farm historical power output missing data repair method.

[0075] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0076] This invention introduces a graph to depict the relationship between known and missing data. Simultaneously, it mines information about the target missing data hidden within each known feature based on an attention mechanism and a GAT network. The attention mechanism enables the mining and updating of node features, while the GAT network mines graph space features. The proposed model can fully utilize the temporal and external feature information of the data to mine hidden features. Compared to other existing repair methods, this method has higher repair accuracy and can handle situations with large-scale missing wind farm data. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of the present invention or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This is a flowchart of a training method for a wind farm historical power output missing data repair model according to the present invention;

[0079] Figure 2 This is a flowchart of a method for repairing missing historical power output data of a wind farm according to an embodiment of the present invention;

[0080] Figure 3 This invention provides a training system for a wind farm historical power output missing data repair model;

[0081] Figure 4 This is a schematic diagram of an electronic device provided by the present invention. Detailed Implementation

[0082] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0083] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0084] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0085] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0086] With the continuous advancement of artificial intelligence algorithms such as deep learning and machine learning, the powerful feature transformation capabilities of attention mechanisms and the powerful feature aggregation and mining capabilities of GAT networks have been widely applied in many fields and have achieved relatively good results. If attention mechanisms and GAT network technology can be applied to the repair of missing historical output data in wind farms, the repair effect can be effectively improved.

[0087] To address the issue that most current wind farm historical missing data repair algorithms rarely fully consider the temporal and external features of the data, resulting in low accuracy in practical engineering applications, given the increasing impact of wind power output on the power system, this invention introduces a graph to depict the relationship between known and missing data. Simultaneously, it mines the information of the target missing data contained in each known feature based on an attention mechanism and a GAT network, thereby improving the accuracy of wind farm missing data repair.

[0088] The first objective of this invention is to provide a method for training a model to repair missing historical power output data of wind farms, comprising the following steps:

[0089] S1. Identify the missing historical output values ​​of wind farms in the pre-constructed historical feature database of wind farms and their positions in the corresponding time series of the historical feature database of wind farms;

[0090] S2. Based on the missing values ​​and their corresponding time series positions in the wind farm historical feature database, extract the feature data of the missing historical output values ​​of the wind farm, supplement the remaining feature data based on the feature matching mechanism, and take each feature in the feature data after supplementing the remaining feature data as a graph node, and construct a feature graph based on graph theory.

[0091] S3. Train the basic model through the feature graph to establish a nonlinear mapping from graph data to wind power output value, and train the wind farm historical power output missing data repair model. The basic model is built on the graph convolutional neural network framework.

[0092] Prior to S1, it also includes: integrating historical wind farm output data and meteorological data to construct a historical characteristic database of wind farms.

[0093] Among them, an attention mechanism network was built, and an attention + GAT (Graph Attention Network) network model was built based on the graph convolutional neural network framework. The model was trained in combination with the constructed feature graph to establish a nonlinear mapping from graph data to wind power output values, and a model for repairing missing historical power output data of wind farms was trained.

[0094] This invention addresses the problems existing in the repair of missing wind farm data by making full use of the temporal and external features of the data. This network can fully utilize the temporal and external features of the data to mine hidden features of the data. Compared with other existing repair methods, it can mine hidden features of the data and form a high-precision method for repairing missing historical output data of wind farms.

[0095] The method proposed in this invention has the following advantages:

[0096] 1. Constructing a Feature Graph: 1) Innovatively constructing a heterogeneous feature graph, where the features between nodes are heterogeneous, such as one node representing wind speed and another representing wind direction; 2) The feature graph includes wind power output values ​​at missing times, set to 0, to facilitate learning the mapping relationship between known node data and the power to be repaired during graph feature extraction; 3) To improve feature utilization, a matching method based on Euclidean distance is used to mine similar wind power output values ​​from historical databases. Other feature representation methods mostly describe the relationship between feature data in the form of data matrices, while the feature graph restores the correlation between features in the form of a non-Euclidean space graph, providing a more comprehensive description.

[0097] 2. The wind power output missing data repair method proposed in this invention is based on an attention mechanism and the GAT algorithm. The attention mechanism network assigns different weights to each graph node, thereby learning the mapping relationship between node features and the power to be repaired, thus realizing the mining of node feature information. The GAT algorithm, by automatically updating the weights of graph connection edges and aggregating graph feature information, constructs a GAT network to achieve feature information mining at the graph space level. These two algorithms mine features at different levels; therefore, merging the two networks to achieve graph feature mining results in a more comprehensive feature mining compared to other algorithms. This is also the first application of these two algorithms in the field of new energy related to power grids.

[0098] The specific steps of this invention are described below:

[0099] Step 1: Based on database systems such as metering systems and meteorological station systems, collect all relevant data including historical power output, wind speed, wind direction, humidity, temperature, air pressure, air density, and date of a wind farm to construct a historical characteristic database of the wind farm.

[0100] As an example of this embodiment, step 1 includes the following steps:

[0101] Historical power output data from wind farms is obtained from the metering system, and meteorological data of the same time and granularity, including wind speed, wind direction, humidity, temperature, air pressure, and air density, are also obtained from the weather station. All relevant data are then integrated based on date and other time information. ,in For the corresponding data characteristics at time t, such as power, wind speed, wind direction, humidity, temperature, air pressure, and air density, The data length is typically between one and three years, thus forming a historical characteristic database of wind farms.

[0102] Step 2: Based on the historical feature database of wind farms, identify the missing historical output values ​​of wind farms and their positions in the time series of the historical database;

[0103] As an example of this embodiment, step 2 includes the following steps:

[0104] Data collected from database systems such as metering systems and meteorological station systems are generally marked with preset characters such as nan and -999 to indicate missing or abnormal data. Combined with the database data usage instructions, missing historical output values ​​of wind farms are identified and their corresponding positions in the historical database time series are marked.

[0105] Step 3: Combine historical time series features, meteorological features and other feature data, supplement the remaining feature data based on the feature matching mechanism, and take each feature in the feature data after supplementing the remaining feature data as a graph node, and construct a feature graph based on graph theory;

[0106] As an example of this embodiment, step 3 includes feature data extraction, feature matching, and constructing a feature graph.

[0107] The data extraction process includes the following steps:

[0108] First, meteorological data such as wind speed and direction were extracted for the historical power output of the wind farm at the corresponding time. At the same time, power output data at three time scales before and after the corresponding time were extracted and fused to obtain the following feature data:

[0109]

[0110] in, For meteorological data characteristics such as wind speed, wind direction, humidity, temperature, air pressure, and air density at the corresponding time of power loss, This represents the wind power output values ​​at three time scales prior to the moment of power loss. And so on, among which The power at the moment of power loss is set to 0.

[0111] Feature matching includes the following steps:

[0112] First, the extracted feature data The dataset is divided into feature datasets corresponding to missing historical power output values ​​of wind farms. and other feature datasets Then, the similarity between the two is calculated using Euclidean distance, as follows:

[0113]

[0114] in They are respectively and A certain feature data in, i The corresponding feature data includes power, wind speed, wind direction, humidity, temperature, air pressure, air density, and other data characteristics.

[0115] Sort the calculated Euclidean distances from smallest to largest, and extract the ones with the smallest distances. Corresponding wind power output Merged into feature data .

[0116] The construction of the feature graph includes the following steps:

[0117] A graph data in graph theory can be represented as: ,in A Represents the connection matrix. X Representing node features, here, the obtained feature data... Each feature in the graph is treated as a graph node, thus forming a graph data node. Each graph node is connected in a fully connected manner to form a feature graph.

[0118] Step 4: Build an attention mechanism network and build an attention + GAT (Graph Attention Network) network model based on the graph convolutional neural network framework. Train the model with the constructed feature graph to obtain the missing historical output data repair values ​​of the wind farm.

[0119] As an example of this embodiment, step 4 includes the following steps:

[0120] Repairing missing historical power output data in wind farms is essentially a process of mining the high-dimensional relationships implicit in the feature graph. Feature mining of the feature graph includes mining node feature information and graph information. Combining the characteristics of attention mechanisms in learning the mapping relationships between different input features and between input and output, an attention mechanism network is constructed, assigning different weights to each graph node to learn the mapping relationship between node features and the power to be repaired, thus achieving node feature information mining. Simultaneously, combining the characteristics of the GAT algorithm in automatically updating graph connection edge weights and aggregating graph feature information, a GAT network is constructed to achieve feature information mining at the graph space level. Finally, the attention mechanism network and the GAT network are merged to learn and extract features from the feature graph, establishing a non-linear mapping from graph data to wind power output values, outputting the results of repairing missing historical power output data of the wind farm, and training to obtain a model for repairing missing historical power output data of the wind farm.

[0121] Preferably, the attention mechanism network is constructed by the following steps:

[0122] For the constructed feature data When analyzing the impact of all eigenvalues ​​on the output, three matrices are first obtained through network training, namely... and feature data vector Multiplying each of these three matrices yields three new eigenvectors. , , Then the vector with vector Multiplication yields a new eigenvector. , will vector After normalization using the softmax function, it is then compared with the matrix. Multiplying them together yields the final new feature data vector. The normalization operation of the softmax function is as follows:

[0123]

[0124] in, For feature vectors, For the nth element, the number of elements is related to the vector. X The dimension is consistent, and the value is 14. The resulting new feature vector;

[0125] Setting up a GAT network includes the following steps:

[0126] In the constructed feature graph, each node represents different feature information, while the feature information of each edge is not yet defined and needs to be mined using a GAT network. Multiple GAT layers are used to aggregate the information between nodes in the feature graph, and then a fully connected layer is used to transform the dimensions of the features, achieving a non-linear mapping from graph data to wind power output values. The graph node information aggregation method is as follows:

[0127]

[0128]

[0129] Among them, parameters W Used to perform feature dimension transformation for each node, parameters a Used to calculate the correlation weights between nodes, || represents vector concatenation. a i，j Indicates in a The calculation obtained below i, j Weights between nodes Represents a non-linear activation function. h i For the updated i Node characteristic information.

[0130] The calculation formula for a fully connected layer is as follows:

[0131]

[0132] in Indicates the first k Layer feature vectors, For the first k Layer weight vector, For the first k Layer bias.

[0133] This invention also provides a method for repairing missing historical power output data of wind farms, including:

[0134] Obtain feature data of the time point to be repaired;

[0135] Feature data is supplemented using feature matching mechanisms, and feature graphs are constructed using graph theory.

[0136] The feature graph is input into the wind farm historical power output missing data repair model trained by the method described above, and the wind farm historical power output missing data repair result is output.

[0137] The specific steps are as follows: Obtain feature data of the time point to be repaired. The feature matching mechanism described above is used to supplement feature data. Then, a feature graph is constructed using graph theory and input into the constructed attention + GAT network to output the results of repairing missing historical power output data of the wind farm.

[0138] The second objective of this invention is to provide a method for repairing missing historical power output data of wind farms, comprising:

[0139] Obtain the missing historical power output values ​​of the wind farm at the time point to be repaired;

[0140] Based on the missing historical power output values ​​of wind farms and their corresponding time series positions in the historical feature database of wind farms, feature data of the missing historical power output values ​​of wind farms are extracted. The remaining feature data are supplemented based on the feature matching mechanism. Each feature in the feature data after the supplementation of the remaining feature data is taken as a graph node. A feature graph is constructed based on graph theory.

[0141] The supplemented feature data and feature graph are input into the wind farm historical output missing data repair model trained by the method described above, and the output is the wind farm historical output missing data repair result.

[0142] The method of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0143] Example 1:

[0144] like Figure 2 As shown in the figure, this embodiment provides a method for repairing missing historical power output data of a wind farm, including the following steps:

[0145] Step 1: To construct a historical characteristic database of wind farms, data collection is required. This involves collecting data uploaded from monitoring equipment to the metering system, obtaining historical power output data for a specific wind farm from the metering system, and simultaneously acquiring meteorological data of the same time and granularity from the weather station, including wind speed, wind direction, humidity, temperature, air pressure, and air density. All relevant data are then integrated based on date and other time information. ,in For the corresponding data characteristics at time t, such as power, wind speed, wind direction, humidity, temperature, air pressure, and air density, The data length is typically one to three years, thus forming a historical characteristic database of wind farms.

[0146] Step 2: Organize the data in the historical feature database of wind farms, set the missing data type of wind power output to random missing (missing location is random), and the missing value is 10% of the total data.

[0147] Step 3: Extract features, match features, and construct a feature graph. Figure 3 Partial composition.

[0148] Data extraction begins by extracting meteorological data such as wind speed and direction corresponding to the historical power output of the wind farm at specific times. Simultaneously, power output data at three time scales before and after the corresponding time are extracted and fused to obtain the following feature data:

[0149]

[0150] in, For meteorological data characteristics such as wind speed, wind direction, humidity, temperature, air pressure, and air density at the corresponding time of power loss, This represents the wind power output values ​​at three time scales prior to the moment of power loss. And so on, among which The power at the moment of power loss is set to 0.

[0151] Feature matching first involves matching the extracted feature data... The dataset is divided into feature datasets corresponding to missing historical power output values ​​of wind farms. and other feature datasets Then, the similarity between the two is calculated using Euclidean distance, as follows:

[0152]

[0153] in They are respectively and A certain feature data in, i The corresponding feature data includes power, wind speed, wind direction, humidity, temperature, air pressure, air density, and other data characteristics.

[0154] Sort the calculated Euclidean distances from smallest to largest, and extract the ones with the smallest distances. Corresponding wind power output Merged into feature data .

[0155] Constructing a feature graph: In graph theory, graph data can be represented as... ,in A Represents the connection matrix. X Representing node features, here, the obtained feature data... Each feature in the graph is treated as a graph node, forming a graph data node. These graph nodes are connected using a fully connected approach, creating a feature graph. Finally, all feature graphs are divided into training and testing sets in an 8:2 ratio.

[0156] Step 4: First, build an attention mechanism network, then build a GAT network based on the graph convolutional neural network framework, merge the two networks to build an attention + GAT model, learn and extract features from the feature graph, establish a nonlinear mapping from graph data to wind power output values, divide the training samples in the training set into several batches for training, obtain a well-parameter attention + GAT model, and finally achieve the repair of missing historical power output data of the wind farm.

[0157] Construct an attention mechanism network for the constructed feature data. When analyzing the impact of all eigenvalues ​​on the output, three matrices are first obtained through network training, namely... and feature data vector Multiplying each of these three matrices yields three new eigenvectors. , , Then the vector with vector Multiplication yields a new eigenvector. , will vector After normalization using the softmax function, it is then compared with the matrix. Multiplying them together yields the final new feature data vector. The normalization operation of the softmax function is as follows:

[0158]

[0159] in, For feature vectors, For the nth element, the number of elements is related to the vector. X The dimension is consistent, and the value is 14. This is the new feature vector obtained.

[0160] In the constructed feature graph, each node represents different feature information, while the feature information of each edge is not yet defined. The GAT network is used to mine the graph's feature information. Multiple GAT layers are used to aggregate the information between nodes in the feature graph. Then, a fully connected layer is used to transform the dimensionality of the features, achieving a non-linear mapping from graph data to wind power output values. The graph node information aggregation method is as follows:

[0161]

[0162]

[0163] Among them, parameters W Used to perform feature dimension transformation for each node, parameters a Used to calculate the correlation weights between nodes, || represents vector concatenation. a i，j Indicates in a The calculation obtained below i, j Weights between nodes Represents a non-linear activation function. h i For the updated i Node characteristic information.

[0164] The calculation formula for a fully connected layer is as follows:

[0165]

[0166] in Indicates the first k Layer feature vectors, For the first k Layer weight vector, For the first k Layer bias.

[0167] The parameters selected for the GAT network are: two GAT layers with 15 hidden nodes, and three fully connected layers with 100, 50, and 1 node respectively.

[0168] Model training and testing: Based on the constructed attention mechanism network and GAT network model, the two models are merged to obtain the attention + GAT model. Based on this model, the feature graph is learned and features are extracted to establish a nonlinear mapping from graph data to wind power output values. The training samples in the training set are divided into several batches (50 in this embodiment) for training to obtain a well-parameterized attention + GAT model and finally achieve the repair of missing historical data of wind farms.

[0169] Example 2:

[0170] like Figure 1 As shown in the figure, this embodiment provides a method for repairing missing historical power output data of a wind farm, including the following steps:

[0171] Step 1: To construct a historical characteristic database of wind farms, data collection is required. This involves collecting data uploaded from monitoring equipment to the metering system, obtaining historical power output data for a specific wind farm from the metering system, and simultaneously acquiring meteorological data of the same time and granularity from the weather station, including wind speed, wind direction, humidity, temperature, air pressure, and air density. All relevant data are then integrated based on date and other time information. ,in For the corresponding data characteristics at time t, such as power, wind speed, wind direction, humidity, temperature, air pressure, and air density, The data length is typically one to three years, thus forming a historical characteristic database of wind farms.

[0172] Step 2: Organize the data in the historical feature database of wind farms, set the missing data type of wind power output to continuous missing (the missing positions are continuous in time), and the missing value is 10% of the total data.

[0173] Step 3: Extract features, match features, and construct a feature graph. Figure 3 Partial composition.

[0174] Data extraction begins by extracting meteorological data such as wind speed and direction corresponding to the historical power output of the wind farm at specific times. Simultaneously, power output data at three time scales before and after the corresponding time are extracted and fused to obtain the following feature data:

[0175]

[0176] in, For meteorological data characteristics such as wind speed, wind direction, humidity, temperature, air pressure, and air density at the corresponding time of power loss, This represents the wind power output values ​​at three time scales prior to the moment of power loss. And so on, among which The power at the moment of power loss is set to 0.

[0177] Feature matching first involves matching the extracted feature data... The dataset is divided into feature datasets corresponding to missing historical power output values ​​of wind farms. and other feature datasets Then, the similarity between the two is calculated using Euclidean distance, as follows:

[0178]

[0179] in They are respectively and A certain feature data in, i The corresponding feature data includes power, wind speed, wind direction, humidity, temperature, air pressure, air density, and other data characteristics.

[0180] Sort the calculated Euclidean distances from smallest to largest, and extract the ones with the smallest distances. Corresponding wind power output Merged into feature data .

[0181] Constructing a feature graph: In graph theory, graph data can be represented as... ,in A Represents the connection matrix. X Representing node features, here, the obtained feature data... Each feature in the graph is treated as a graph node, forming a graph data node. These graph nodes are connected using a fully connected approach, creating a feature graph. Finally, all feature graphs are divided into training and testing sets in an 8:2 ratio.

[0182] Step 4: First, build an attention mechanism network, then build a GAT network based on the graph convolutional neural network framework, merge the two networks to build an attention + GAT model, learn and extract features from the feature graph, establish a nonlinear mapping from graph data to wind power output values, divide the training samples in the training set into several batches for training, obtain a well-parameter attention + GAT model, and finally achieve the repair of missing historical power output data of the wind farm.

[0183] Construct an attention mechanism network for the constructed feature data. When analyzing the impact of all eigenvalues ​​on the output, three matrices are first obtained through network training, namely... and feature data vector Multiplying each of these three matrices yields three new eigenvectors. , , Then the vector with vector Multiplication yields a new eigenvector. , will vector After normalization using the softmax function, it is then compared with the matrix. Multiplying them together yields the final new feature data vector. The normalization operation of the softmax function is as follows:

[0184]

[0185] in, For feature vectors, For the nth element, the number of elements is related to the vector. X The dimension is consistent, and the value is 14. This is the new feature vector obtained.

[0186] In the constructed feature graph, each node represents different feature information, while the feature information of each edge is not yet defined. The GAT network is used to mine the graph's feature information. Multiple GAT layers are used to aggregate the information between nodes in the feature graph. Then, a fully connected layer is used to transform the dimensionality of the features, achieving a non-linear mapping from graph data to wind power output values. The graph node information aggregation method is as follows:

[0187]

[0188]

[0189] Among them, parameters W Used to perform feature dimension transformation for each node, parameters a Used to calculate the correlation weights between nodes, || represents vector concatenation. a i，j Indicates in a The calculation obtained below i, j Weights between nodes Represents a non-linear activation function. h i For the updated i Node characteristic information.

[0190] The calculation formula for a fully connected layer is as follows:

[0191]

[0192] in Indicates the first k Layer feature vectors, For the first k Layer weight vector, For the first k Layer bias.

[0193] The parameters selected for the GAT network are: two GAT layers with 15 hidden nodes, and three fully connected layers with 100, 50, and 1 node respectively.

[0194] Model training and testing: Based on the constructed attention mechanism network and GAT network model, the two models are merged to obtain the attention + GAT model. Based on this model, the feature graph is learned and features are extracted to establish a nonlinear mapping from graph data to wind power output values. The training samples in the training set are divided into several batches (50 in this embodiment) for training to obtain a well-parameterized attention + GAT model and finally achieve the repair of missing historical data of wind farms.

[0195] Example 3:

[0196] like Figure 1 As shown in the figure, this embodiment provides a method for repairing missing historical power output data of a wind farm, including the following steps:

[0197] Step 1: To construct a historical characteristic database of wind farms, data collection is required. This involves collecting data uploaded from monitoring equipment to the metering system, obtaining historical power output data for a specific wind farm from the metering system, and simultaneously acquiring meteorological data of the same time and granularity from the weather station, including wind speed, wind direction, humidity, temperature, air pressure, and air density. All relevant data are then integrated based on date and other time information. ,in For the corresponding data characteristics at time t, such as power, wind speed, wind direction, humidity, temperature, air pressure, and air density, The data length is typically one to three years, thus forming a historical characteristic database of wind farms.

[0198] Step 2: Organize the data in the historical feature database of wind farms, set the missing data type of wind power output to random missing (missing location is random), and the missing value is 50% of the total data.

[0199] Step 3: Extract features, match features, and construct a feature graph. Figure 3 Partial composition.

[0200] Data extraction begins by extracting meteorological data such as wind speed and direction corresponding to the historical power output of the wind farm at specific times. Simultaneously, power output data at three time scales before and after the corresponding time are extracted and fused to obtain the following feature data:

[0201]

[0202] in, For meteorological data characteristics such as wind speed, wind direction, humidity, temperature, air pressure, and air density at the corresponding time of power loss, This represents the wind power output values ​​at three time scales prior to the moment of power loss. And so on, among which The power at the moment of power loss is set to 0.

[0203] Feature matching first involves matching the extracted feature data... The dataset is divided into feature datasets corresponding to missing historical power output values ​​of wind farms. and other feature datasets Then, the similarity between the two is calculated using Euclidean distance, as follows:

[0204]

[0205] in They are respectively and A certain feature data in, i The corresponding feature data includes power, wind speed, wind direction, humidity, temperature, air pressure, air density, and other data characteristics.

[0206] Sort the calculated Euclidean distances from smallest to largest, and extract the ones with the smallest distances. Corresponding wind power output Merged into feature data .

[0207] Constructing a feature graph: In graph theory, graph data can be represented as... ,in A Represents the connection matrix. X Representing node features, here, the obtained feature data... Each feature in the graph is treated as a graph node, forming a graph data node. These graph nodes are connected using a fully connected approach, creating a feature graph. Finally, all feature graphs are divided into training and testing sets in an 8:2 ratio.

[0208] Step 4: First, build an attention mechanism network, then build a GAT network based on the graph convolutional neural network framework, merge the two networks to build an attention + GAT model, learn and extract features from the feature graph, establish a nonlinear mapping from graph data to wind power output values, divide the training samples in the training set into several batches for training, obtain a well-parameter attention + GAT model, and finally achieve the repair of missing historical power output data of the wind farm.

[0209] Construct an attention mechanism network for the constructed feature data. When analyzing the impact of all eigenvalues ​​on the output, three matrices are first obtained through network training, namely... and feature data vector Multiplying each of these three matrices yields three new eigenvectors. , , Then the vector with vector Multiplication yields a new eigenvector. , will vector After normalization using the softmax function, it is then compared with the matrix. Multiplying them together yields the final new feature data vector. The normalization operation of the softmax function is as follows:

[0210]

[0211] in, For feature vectors, For the nth element, the number of elements is related to the vector. X The dimension is consistent, and the value is 14. This is the new feature vector obtained.

[0212] In the constructed feature graph, each node represents different feature information, while the feature information of each edge is not yet defined. The GAT network is used to mine the graph's feature information. Multiple GAT layers are used to aggregate the information between nodes in the feature graph. Then, a fully connected layer is used to transform the dimensionality of the features, achieving a non-linear mapping from graph data to wind power output values. The graph node information aggregation method is as follows:

[0213]

[0214]

[0215] Among them, parameters W Used to perform feature dimension transformation for each node, parameters a Used to calculate the correlation weights between nodes, || represents vector concatenation. a i，j Indicates in a The calculation obtained below i, j Weights between nodes Represents a non-linear activation function. h i For the updated i Node characteristic information.

[0216] The calculation formula for a fully connected layer is as follows:

[0217]

[0218] in Indicates the first k Layer feature vectors, For the first k Layer weight vector, For the first k Layer bias.

[0219] The parameters selected for the GAT network are: two GAT layers with 15 hidden nodes, and three fully connected layers with 100, 50, and 1 node respectively.

[0220] Model training and testing: Based on the constructed attention mechanism network and GAT network model, the two models are merged to obtain the attention + GAT model. Based on this model, the feature graph is learned and features are extracted to establish a nonlinear mapping from graph data to wind power output values. The training samples in the training set are divided into several batches (50 in this embodiment) for training to obtain a well-parameterized attention + GAT model and finally achieve the repair of missing historical data of wind farms.

[0221] like Figure 3 As shown, a second objective of this invention is to provide a training system for a wind farm historical power output missing data repair model, comprising:

[0222] The identification module is used to identify the missing historical output values ​​of wind farms in the pre-built historical feature database of wind farms and their positions in the corresponding time series of the historical feature database of wind farms;

[0223] The construction module is used to extract feature data of missing historical output values ​​of wind farms based on the missing values ​​and their positions in the corresponding time series of the wind farm historical feature database, supplement the remaining feature data based on the feature matching mechanism, and take each feature in the feature data after supplementing the remaining feature data as a graph node, and construct a feature graph based on graph theory.

[0224] The training module is used to train a basic model through the feature graph, establish a nonlinear mapping from graph data to wind power output values, and train a model to repair missing historical power output data of wind farms. The basic model is built on a graph convolutional neural network framework.

[0225] It also includes a fusion module, which is used to merge historical wind farm output data and meteorological data to build a historical characteristic database of wind farms;

[0226] As a further improvement to the invention, in the fusion module,

[0227] The method for constructing the pre-built historical feature database of wind farms includes:

[0228] Historical power output data from wind farms is obtained from the metering system, and meteorological data of the same time and granularity, including wind speed, wind direction, humidity, temperature, air pressure, and air density, are obtained from the weather station. Missing or abnormal data are marked with preset characters, and all relevant data are merged based on time information. ,in The data characteristics of power, wind speed, wind direction, humidity, temperature, air pressure, and air density at time t are used to form a historical feature database of wind farms.

[0229] As a further improvement to the invention, the identification module, which identifies missing historical output values ​​of wind farms and their positions in the corresponding time series of the pre-constructed historical feature database of wind farms, includes:

[0230] Data collected from the database systems of metering systems and meteorological stations is marked with preset characters to identify missing or abnormal data. Combined with the database data usage instructions, missing historical output values ​​of wind farms are identified and their corresponding positions in the historical database time series are marked.

[0231] As a further improvement to the invention, in the construction module, the step of extracting feature data of missing historical power output values ​​of wind farms based on the missing values ​​and their corresponding time series in the wind farm historical feature database includes:

[0232] Meteorological data corresponding to the historical power output data of the wind farm were extracted, and wind power output values ​​at the corresponding missing power time and three time scales before and after were extracted and fused to obtain feature data:

[0233]

[0234] in, For the wind speed, wind direction, humidity, temperature, air pressure, and air density at the moment of power loss, This represents the wind power output values ​​at three time scales prior to the moment of power loss. And so on, among which This represents the wind power output value at the moment of power loss.

[0235] As a further improvement to the invention, in the construction module, the supplementation of remaining feature data based on the feature matching mechanism includes:

[0236] Extracted feature data The dataset is divided into feature datasets corresponding to missing historical power output values ​​of wind farms. and other feature datasets Then calculate the similarity value between the two;

[0237] Arrange the calculated similarity values ​​from smallest to largest, and extract the value corresponding to the smallest similarity value. Corresponding wind power output The P is then merged into the remaining feature data. .

[0238] As a further improvement to the invention, in the construction module, the step of treating each feature in the feature data after supplementing the remaining feature data as a graph node and constructing a feature graph based on graph theory includes:

[0239] Feature data Each feature in the graph is treated as a graph node, thus forming a graph data node. Each graph data node is connected in a fully connected manner to form a feature graph.

[0240] As a further improvement to the invention, in the building module,

[0241] The process of training a basic model using the feature graph to establish a nonlinear mapping from graph data to wind power output values, and training a model to repair missing historical power output data of wind farms, includes:

[0242] By combining the attention mechanism to learn the mapping relationship between different input features and between input and output, an attention mechanism network is constructed, and different weights are assigned to each graph node. Based on the mapping relationship between node features and the power to be repaired, node feature information is mined.

[0243] By combining the GAT algorithm to automatically update the weights of graph connection edges and the feature information of aggregated graphs, a GAT network is constructed to mine feature information at the graph space level.

[0244] By combining the attention mechanism network and the GAT network, the feature graph is learned and features are extracted, and a nonlinear mapping from graph data to wind power output values ​​is established.

[0245] More specifically, the fusion module, together with the fused historical wind farm output data and meteorological data, constructs a corresponding historical feature database of wind farms.

[0246] The identification module corresponds to the missing historical output values ​​of wind farms and their positions in the corresponding time series in the wind farm historical feature database that is pre-built.

[0247] The construction module extracts feature data of missing historical output values ​​of wind farms based on the missing values ​​and their corresponding time series positions in the historical feature database of wind farms. It supplements the remaining feature data based on the feature matching mechanism, and treats each feature in the supplemented feature data as a graph node. Based on graph theory, a corresponding feature graph is constructed.

[0248] The training module constructs an attention mechanism network and builds a graph attention network model based on the graph convolutional neural network framework. Combined with the constructed feature graph training model, a nonlinear mapping from graph data to wind power output values ​​is established, and a corresponding model for repairing missing historical power output data of wind farms is trained.

[0249] This invention also provides a system for repairing missing historical power output data in wind farms, comprising:

[0250] The acquisition module is used to obtain the missing historical output values ​​of the wind farm at the time point to be repaired;

[0251] The feature construction module is used to extract feature data of missing historical power output values ​​of wind farms based on the missing historical power output values ​​of wind farms and their corresponding time series positions in the historical feature database of wind farms. It supplements the remaining feature data based on the feature matching mechanism, and uses each feature in the supplemented feature data as a graph node, and constructs a feature graph based on graph theory.

[0252] The output module is used to input the feature data and feature graph after supplementing the remaining feature data into the wind farm historical power output missing data repair model trained by the method, and output the wind farm historical power output missing data repair result.

[0253] like Figure 4 As shown, a fourth objective of this invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the wind farm historical power output missing data repair model training method or the wind farm historical power output missing data repair method.

[0254] The training method for the wind farm historical power output missing data repair model includes the following steps:

[0255] Identify the missing historical power output values ​​of wind farms in a pre-built historical feature database and their positions in the corresponding time series of the database;

[0256] Based on the missing values ​​and their corresponding time series positions in the wind farm historical feature database, feature data of missing historical power output values ​​of wind farms are extracted. The remaining feature data are supplemented based on the feature matching mechanism. Each feature in the feature data after the supplementation of the remaining feature data is taken as a graph node. A feature graph is constructed based on graph theory.

[0257] The basic model is trained by the feature graph to establish a nonlinear mapping from graph data to wind power output values, and a model for repairing missing historical power output data of wind farms is trained. The basic model is built on a graph convolutional neural network framework.

[0258] The method for repairing missing historical power output data of wind farms includes:

[0259] Obtain the missing historical power output values ​​of the wind farm at the time point to be repaired;

[0260] Based on the missing historical power output values ​​of wind farms and their corresponding time series positions in the historical feature database of wind farms, feature data of the missing historical power output values ​​of wind farms are extracted. The remaining feature data are supplemented based on the feature matching mechanism. Each feature in the feature data after the supplementation of the remaining feature data is taken as a graph node. A feature graph is constructed based on graph theory.

[0261] The feature data and feature graph after supplementing the remaining feature data are input into the wind farm historical output missing data repair model trained by the method described above, and the output is the wind farm historical output missing data repair result.

[0262] A fifth objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the wind farm historical power output missing data repair model training method or the wind farm historical power output missing data repair method.

[0263] The training method for the wind farm historical power output missing data repair model includes the following steps:

[0264] Identify the missing historical power output values ​​of wind farms in a pre-built historical feature database and their positions in the corresponding time series of the database;

[0265] Based on the missing values ​​and their corresponding time series positions in the wind farm historical feature database, feature data of missing historical power output values ​​of wind farms are extracted. The remaining feature data are supplemented based on the feature matching mechanism. Each feature in the feature data after the supplementation of the remaining feature data is taken as a graph node. A feature graph is constructed based on graph theory.

[0266] The basic model is trained by the feature graph to establish a nonlinear mapping from graph data to wind power output values, and a model for repairing missing historical power output data of wind farms is trained. The basic model is built on a graph convolutional neural network framework.

[0267] The method for repairing missing historical power output data of wind farms includes:

[0268] Obtain the missing historical power output values ​​of the wind farm at the time point to be repaired;

[0269] Based on the missing historical power output values ​​of wind farms and their corresponding time series positions in the historical feature database of wind farms, feature data of the missing historical power output values ​​of wind farms are extracted. The remaining feature data are supplemented based on the feature matching mechanism. Each feature in the feature data after the supplementation of the remaining feature data is taken as a graph node. A feature graph is constructed based on graph theory.

[0270] The feature data and feature graph after supplementing the remaining feature data are input into the wind farm historical output missing data repair model trained by the method described above, and the output is the wind farm historical output missing data repair result.

[0271] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0272] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0273] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0274] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0275] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention without departing from the spirit and scope of the present invention. Any modifications or equivalent substitutions should be covered within the scope of protection of the claims of the present invention. Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention 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, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0276] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0277] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0278] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0279] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for training a wind farm historical power output missing data repairing model, characterized in that, The method comprises the following steps: Identify the missing values of the historical output of the wind farm in the pre-constructed historical feature database of the wind farm and their positions in the corresponding time series in the historical feature database of the wind farm; Based on the missing values and their positions in the corresponding time series in the historical feature database of the wind farm, extract the feature data of the missing values of the historical output of the wind farm, supplement the remaining feature data based on the feature matching mechanism, take each feature in the feature data after supplementing the remaining feature data as a graph node, and construct a feature Graph graph based on graph theory; Train a basic model based on the feature Graph graph, establish a nonlinear mapping from graph data to wind power output value, and train a missing data repair model for the historical output of the wind farm; the basic model is built based on a graph convolutional neural network framework; wherein the training of the basic model based on the feature Graph graph, the establishment of the nonlinear mapping from the graph data to the wind power output value, and the training of the missing data repair model for the historical output of the wind farm comprise: Learn the mapping relationship between different input features and between input and output by combining the attention mechanism, construct an attention mechanism network, assign different weights to each graph node, and mine node feature information based on the mapping relationship between node features and the power to be repaired; Combine the GAT algorithm to automatically update the graph connection edge weight and aggregate the graph feature information, construct a GAT network, and mine the feature information at the graph space level; Combine the attention mechanism network and the GAT network to learn and extract features of the feature Graph graph, and establish a nonlinear mapping from graph data to wind power output value; Output the missing data repair result of the historical output of the wind farm, and train a missing data repair model for the historical output of the wind farm.

2. The wind farm historical output missing data repair model training method according to claim 1, characterized in that, The identification of the missing values of the historical output of the wind farm in the pre-constructed historical feature database of the wind farm and their positions in the corresponding time series in the historical feature database of the wind farm comprises: Obtain the missing or abnormal data marked in advance with a preset character in the historical feature database of the wind farm, identify the missing values of the historical output of the wind farm in combination with the data usage instructions of the historical feature database of the wind farm, and mark their positions in the corresponding time series in the historical feature database of the wind farm; Identify the missing values of the historical output of the wind farm to obtain the missing type of the wind power output data and the corresponding missing values; the missing type includes random missing or continuous missing.

3. The wind farm historical output missing data repair model training method according to claim 1, characterized in that, The extraction of the feature data of the missing values of the historical output of the wind farm based on the missing values and their positions in the corresponding time series in the historical feature database of the wind farm comprises: Extract the meteorological data corresponding to the time of the historical output data of the wind farm, extract the wind power output value at the corresponding power missing time and the three time scales before and after it, and fuse to obtain the feature data: wherein, wind speed, wind direction, humidity, temperature, air pressure, air density corresponding to the power missing moment, wind power output values corresponding to the three time scales before the power missing moment, , , , and so on.

4. The wind farm historical output missing data repair model training method according to claim 3, characterized in that, The supplement of the remaining feature data based on the feature matching mechanism comprises: The extracted feature data is divided into a feature data set corresponding to missing values of historical output of the wind farm and a remaining feature data set , and a similarity value between the two is calculated. The calculated similarity values are arranged from small to large, and the corresponding The corresponding wind power output The P is merged into the feature data as the rest of the feature data .

5. The wind farm historical output missing data repair model training method according to claim 4, characterized in that, The similarity value between the two is calculated by using the Euclidean distance method, which is as follows: wherein respectively and a certain feature data from i power, wind speed, wind direction, humidity, temperature, air pressure, air density in the corresponding feature data.

6. The wind farm historical output missing data repair model training method according to claim 4, characterized in that, The taking of each feature in the feature data after the supplement of the remaining feature data as a graph node and the construction of a feature Graph graph based on graph theory comprise: Each of the feature data in the feature data is taken as a graph node, thereby forming a graph data node, and each graph data node is connected in a fully connected manner, thereby constituting a feature Graph graph.

7. The wind farm historical output missing data repair model training method according to claim 1, characterized in that, The training of the base model through the feature Graph graph establishes a nonlinear mapping of the graph data to the wind power output power value, and specifically comprises: For the constructed feature data , three matrices are obtained by network training, respectively , and the feature data vector is multiplied by the three matrices respectively to obtain three new feature vectors , , , then the vector is multiplied by the vector to obtain a new feature vector , and the vector is normalized by a softmax function and then multiplied by the matrix to obtain the final new feature data vector ; wherein the softmax function normalization operation is: wherein, is a feature vector, is the nth element in which the number of elements is consistent with the dimension of the vector X , is a new feature vector obtained; In the constructed feature Graph graph, the information between the nodes of the feature Graph graph is aggregated by using a multi-layer GAT layer, and then a full connection layer is used for dimension transformation of the features, so as to realize the nonlinear mapping of the graph data to the wind power output power value; wherein the Graph graph node information aggregation method is: Among them, parameters W Used to perform feature dimension transformation for each node, parameters a Used to calculate the correlation weights between nodes; || denotes vector concatenation. a i，j Indicates in a The calculation obtained below i, j Weights between nodes Represents a non-linear activation function. h i For the updated i Node feature information; The full connection layer calculation formula is as follows: wherein represents the k layer feature vector, is the k layer weight vector, is the k bias of the layer.

8. A wind farm historical power output missing data repairing method, characterized in that, It comprises: Obtaining the historical output missing value of the wind farm at the to-be-repaired time point; Based on the historical output missing value of the wind farm and its position in the corresponding time sequence in the historical feature database of the wind farm, the feature data of the historical output missing value of the wind farm is extracted, the remaining feature data is supplemented based on the feature matching mechanism, each feature in the feature data after supplementing the remaining feature data is taken as a graph node, each feature in the feature data after supplementing the remaining feature data is taken as a graph node, and a feature Graph graph is constructed based on graph theory; The feature data after supplementing the remaining feature data and the feature Graph graph are input into the wind farm historical output missing data repair model trained by the method of any one of claims 1-7, and a wind farm historical output missing data repair result is output. 9.A system for training a wind farm historical power output missing data repairing model, characterized in that, It comprises: A recognition module for recognizing the historical output missing value of the wind farm and its position in the corresponding time sequence in the historical feature database of the wind farm; A construction module for extracting the feature data of the historical output missing value of the wind farm based on the missing value and its position in the corresponding time sequence in the historical feature database of the wind farm, supplementing the remaining feature data based on the feature matching mechanism, taking each feature in the feature data after supplementing the remaining feature data as a graph node, and constructing a feature Graph graph based on graph theory; A training module for training a base model through the feature Graph graph to establish a nonlinear mapping of the graph data to the wind power output power value, and training a wind farm historical output missing data repair model, wherein the base model is built based on a graph convolutional neural network framework; wherein the training of the base model through the feature Graph graph to establish a nonlinear mapping of the graph data to the wind power output power value, and the training of the wind farm historical output missing data repair model, comprise: Combining the attention mechanism to learn the mapping relationship between different input features and between the input and the output, constructing an attention mechanism network, assigning different weights to each graph node, and mining node feature information based on the mapping relationship between the node features and the to-be-repaired power; Combining the GAT algorithm to automatically update the graph connection edge weight and aggregate the graph feature information, constructing a GAT network, and mining feature information at the graph space level; Combining the attention mechanism network and the GAT network, learning and feature extracting the feature Graph graph, and establishing a nonlinear mapping of the graph data to the wind power output power value.

10. The wind farm historical output missing data repair model training system of claim 9, wherein, The identification module, the identification pre-constructed wind farm historical feature database, the missing value of wind farm historical output and its position in the corresponding time sequence in the wind farm historical feature database, including: Obtain the missing or abnormal data marked in advance with a preset character in the wind farm historical feature database, combine the data usage instructions of the wind farm historical feature database, identify the missing value of the wind farm historical output, and mark its position in the corresponding time sequence in the wind farm historical feature database; Identify the missing value of the wind farm historical output to obtain the missing type of wind power output data and the corresponding missing value; the missing type includes random missing or continuous missing.

11. The wind farm historical output missing data inpainting model training system of claim 9, wherein, In the construction module, the feature data of the missing value of the wind farm historical output is extracted based on the missing value and its position in the corresponding time sequence in the wind farm historical feature database, including: Extract the meteorological data corresponding to the wind power output data at the moment, extract the wind power output value at the corresponding power missing moment and the three time scales before and after it, and fuse to obtain the feature data: wherein, wind speed, wind direction, humidity, temperature, air pressure, air density at the time of power shortage, wind power output values at the three time scales before the time of power shortage, , , , and so on.

12. The wind farm historical output missing data repairing model training system according to claim 11, wherein, In the construction module, the remaining feature data is supplemented based on the feature matching mechanism, including: The extracted feature data is divided into a feature data set corresponding to missing values of historical output of the wind farm and a remaining feature data set , and a similarity value between the two is calculated. The calculated similarity values are arranged from small to large, and the corresponding one with the smallest similarity value is extracted corresponding wind power output merge the P into the feature data as the rest of the feature data .

13. The wind farm historical output missing data inpainting model training system of claim 12, wherein, In the construction module, each feature in the feature data after supplementing the remaining feature data is taken as a graph node, and a feature Graph graph is constructed based on graph theory, including: Each of the features in the feature data is taken as a graph node, thereby forming a graph data node, and each of the graph data nodes is connected in a fully connected manner, thereby constituting a feature Graph graph.

14. A wind farm historical power output missing data patching system, characterized by, Including: An acquisition module for acquiring the missing value of the wind farm historical output at the time point to be repaired; A feature construction module for extracting feature data of the missing value of the wind farm historical output based on the missing value of the wind farm historical output and its position in the corresponding time sequence in the wind farm historical feature database, supplementing the remaining feature data based on the feature matching mechanism, taking each feature in the feature data after supplementing the remaining feature data as a graph node, taking each feature in the feature data after supplementing the remaining feature data as a graph node, and constructing a feature Graph graph based on graph theory; An output module for inputting the feature data after supplementing the remaining feature data and the feature Graph graph into the wind farm historical output missing data repair model trained by the method of any one of claims 1-7, and outputting to obtain the wind farm historical output missing data repair result.

15. An electronic device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the wind farm historical output missing data repair model training method of any one of claims 1-7 or the wind farm historical output missing data repair method of claim 8.

16. A computer readable storage medium characterized by: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the wind farm historical output missing data repair model training method of any one of claims 1-7 or the wind farm historical output missing data repair method of claim 8.

Images