A wind power output short-term prediction method and system

Abstract

The application discloses a kind of wind power output short-term prediction method and system, the method includes: from the historical feature database of wind farm of pre-constructed wind farm scene under cold wave weather is screened, and constitutes cold wave weather small sample scene set;The cold wave weather small sample scene set is input into pre-trained cold wave sample scene generator, and cold wave wind power operation scene sample is generated;Based on graph theory knowledge, cold wave characteristic information graph is constructed in combination with the cold wave wind power operation scene sample;Combining the cold wave characteristic information graph trained basic model, the nonlinear mapping of cold wave characteristic information graph to wind power output power value is established, and wind power output short-term prediction model is obtained, and the basic model is based on graph convolutional neural network framework Graph Attention Network model is built.The application is wind power output short-term prediction method, which can effectively improve the precision of wind power output short-term prediction under cold wave extreme weather.

Description

Technical Field

[0001] This invention relates to the field of wind power output prediction, and specifically to a method and system for short-term wind power output prediction. Background Technology

[0002] With the rapid development of wind power, the installed capacity of wind power in the power grid is constantly increasing. However, the volatility and randomness of wind power output pose challenges to the operation and planning of the power system. In addition, cold waves, as a typical meteorological disaster, have consistently had a severe impact on wind turbine operation, posing challenges to accurate wind power output forecasting and seriously endangering power grid security. In recent years, with changes in climate and other factors, cold waves have become more frequent, and their characteristics, such as strong winds and freezing rain, have become more extreme. During prolonged cold waves, the accuracy of wind power forecasting decreases significantly. Therefore, it is necessary to fully explore the climatic characteristics of cold waves and their impact on wind power output, and to conduct targeted research on accurate wind power forecasting methods under cold wave conditions to improve the power grid's supply capacity and ensure its safe and stable operation.

[0003] In recent years, artificial intelligence algorithms such as deep learning and machine learning have made continuous progress. However, the corresponding models require a large amount of sample data for training. Cold wave weather is a small-sample event, and its sample size is insufficient to support the model's full training. To address this issue, some scholars have proposed short-term wind power output prediction methods based on sample expansion during cold waves. However, these methods suffer from problems such as the inability to accurately capture the meteorological characteristics of cold waves, resulting in a significant discrepancy between the generated meteorological samples and the actual cold wave weather. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a method and system for short-term wind power output prediction. This method will effectively improve the similarity rate of wind power operation scenario sample generation during cold waves, thereby improving the accuracy of short-term wind power output prediction under cold wave weather.

[0005] The objective of this invention is achieved at least through the following technical solutions:

[0006] The first aspect of this invention is to provide a method for training a short-term wind power output prediction model, comprising:

[0007] Wind farm operation scenarios under cold wave weather were selected from a pre-built historical feature database of wind farms to form a small sample scenario set for cold wave weather.

[0008] The small sample scene set of cold wave weather is input into a pre-trained cold wave sample scene generator to generate cold wave wind power operation scene samples.

[0009] Based on graph theory, a cold wave feature information graph is constructed by combining the cold wave wind power operation scenario samples.

[0010] By training a basic model using the constructed cold wave feature information map, a nonlinear mapping from the cold wave feature information map to the wind power output value is established, resulting in a short-term wind power output prediction model. The basic model is based on a graph attention network model built on a graph convolutional neural network framework.

[0011] As an optional embodiment of the present invention, the pre-constructed historical feature database of wind farms is obtained by fusing historical wind farm output data and meteorological data, and the specific method includes:

[0012] Historical power output data of wind farms are 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 meteorological station. The data characteristics of all meteorological data are then integrated based on time information to form a historical feature database of wind farms.

[0013] As an optional embodiment of the present invention, the step of selecting wind farm operation scenarios under cold wave weather from a pre-constructed historical feature database of wind farms to form a small sample scenario set for cold wave weather includes:

[0014] Using M as the time window, temperature data is extracted from a pre-built historical feature database of wind farms. A day contains m time segments, and each temperature segment is: T = {T1, T2, ... T} k}, k = M / m;

[0015] Based on the requirements for cold wave weather, select the time period within the time window where the temperature will drop: t i If, during this period, the weather is determined to be a cold wave based on the criteria for cold wave weather, the relevant meteorological data of the wind farm operation on that day will be extracted and used as the operating scenario of the wind farm under cold wave weather, thus forming a small sample scenario set of cold wave weather.

[0016] As an optional embodiment of the present invention, the training method of the pre-trained cold wave sample scene generator includes:

[0017] The definition of cold wave weather is transformed into a mathematical expression and embedded into the loss function of a generative adversarial network model. A cold wave sample scene generator is obtained by training on a small sample set of cold wave weather scenarios.

[0018] As an optional embodiment of the present invention, the generative adversarial network model includes a generator G and a discriminator D, wherein the generator G is used to generate a set of scenarios with a similar distribution to real cold wave wind power operation scenarios, and the discriminator D is used to distinguish the set of scenarios generated by the generator G from the real sample set; both the generator G and the discriminator D networks are built from fully connected layers, and the calculation formula for the fully connected layer is as follows:

[0019]

[0020] Where, x (k+1) Let x represent the feature vector of the (k+1)th layer. (k) W represents the feature vector of the k-th layer, where k is an integer. (k) Let b be the weight vector of the k-th layer. (k) This is the bias of the k-th layer.

[0021] As an optional embodiment of the present invention

[0022] The loss function W(D,G,L) of the generative adversarial network is:

[0023]

[0024] Among them, L cold Define the loss function for the embedding model for cold wave weather; E z p represents the expected distribution of sample z. data (x) represents the probability distribution of the true sample x, p z (z) represents the probability distribution of the generated sample z, G(z) represents the generator neural network, and D(x) represents the discriminator neural network;

[0025] L cold It consists of two parts:

[0026] L cold1 =|T max -T min -T 降温 | 2

[0027] L cold2 =|T min -T 低温 | 2

[0028] Among them, T 降温 To determine the temperature drop conditions for a cold wave, T 低温 To determine the lowest temperature during a cold wave, T max The highest temperature during the cold wave was T. min The lowest temperature during the cold wave.

[0029] As an optional embodiment of the present invention, the construction of a cold wave feature information map based on graph theory knowledge and combined with the cold wave wind power operation scenario samples includes:

[0030] A graph data representation in graph theory is G(A,X), where A represents the connectivity matrix and X represents the node features. The feature data D = {D...} from the cold wave weather sample database... 1，n D 2，n ...D96，n Each time segment in the graph is treated as a graph node, thus forming a graph data node. The graph nodes are connected in a fully connected manner to form a cold wave characteristic information graph.

[0031] As an optional embodiment of the present invention, the step of training the basic model with the constructed cold wave feature information map, establishing a nonlinear mapping from the cold wave feature information map to the wind power output value, and obtaining a short-term wind power output prediction model includes:

[0032] A graph attention network is built based on the graph convolutional neural network framework to learn and extract features from the cold wave feature information graph data, establish a nonlinear mapping from graph data to wind power output values, and use short-term wind power output prediction under cold wave weather as the output result to train and obtain a short-term wind power output prediction model.

[0033] As an optional embodiment of the present invention, the establishment of the nonlinear mapping from the cold wave characteristic information map to the wind power output value includes:

[0034] In the cold wave feature information map, each node represents a different time segment. The information between the nodes of the cold wave feature information map is aggregated by using a multi-layer graph attention layer. Then, a fully connected layer is used to transform the dimension of the features to obtain a non-linear mapping from the cold wave feature information map to wind power output.

[0035] The method for aggregating information between nodes in a cold wave feature graph using a multi-layer graph attention layer is as follows:

[0036]

[0037]

[0038] In the formula, parameter W is used to perform feature dimension transformation for each node, parameter a is used to calculate the correlation weights between nodes, || denotes vector concatenation, and a i，j Let σ represent the weights between nodes i and j calculated under a, and h represent the nonlinear activation function. i The updated feature information for node i is given, where k is an integer.

[0039] A second aspect of the present invention is to provide a method for short-term prediction of wind power output, comprising:

[0040] By integrating wind farm output data and meteorological data for the predicted date, a sample of wind power operation scenarios for the predicted cold wave is obtained.

[0041] Based on graph theory, a feature information graph of the cold wave to be predicted is constructed by combining the wind power operation scenario samples of the cold wave to be predicted.

[0042] The cold wave feature information map to be predicted is input into the short-term wind power output prediction model trained by the method described above to obtain the short-term prediction result of wind power output under cold wave weather.

[0043] A third aspect of the present invention is to provide a training system for a short-term wind power output prediction model, comprising:

[0044] The filtering module is used to filter wind farm operation scenarios under cold wave weather from a pre-built historical feature database of wind farms, forming a small sample scenario set for cold wave weather.

[0045] The generation module is used to input the small sample scene set of cold wave weather into a pre-trained cold wave sample scene generator to generate cold wave wind power operation scene samples.

[0046] The module is used to construct a cold wave feature information map based on graph theory knowledge and combined with the cold wave wind power operation scenario samples;

[0047] The training module is used to train the basic model by combining the constructed cold wave feature information map, establish a nonlinear mapping from the cold wave feature information map to the wind power output value, and obtain a short-term wind power output prediction model. The basic model is based on a graph attention network model built on a graph convolutional neural network framework.

[0048] As a further improvement of the present invention, in the filtering module, the pre-constructed historical feature database of wind farms is obtained by fusing historical wind farm output data and meteorological data, and the specific method includes:

[0049] 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 meteorological station. All relevant data are then fused together based on the time information, resulting in D = {D...} 1，n D 2，n ...D t，n}, where D t,n ={D t,P D t,W ... D t,T 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.

[0050] As a further improvement of the present invention, the filtering module, which filters wind farm operation scenarios under cold wave weather from a pre-constructed historical feature database of wind farms to form a small sample scenario set for cold wave weather, includes:

[0051] Using M as the time window, temperature data is extracted from a pre-built historical feature database of wind farms. A day contains m time segments, and each temperature segment is: T = {T1, T2, ... T}k}, k = M / m;

[0052] Based on the requirements for cold wave weather, select the time period within the time window where the temperature will drop: t i When the temperature drops, if, during this period, the weather is determined to be a cold wave based on the cold wave weather criteria, then the relevant wind farm operation data for that day, D = {D}, is used. 1，n D 2，n ...D k，n The extracted data serves as the operating scenario for wind farms during cold waves, thus forming a small sample scenario set for cold wave weather.

[0053] As a further improvement of the present invention, the training method of the pre-trained cold wave sample scene generator in the generation module includes:

[0054] The definition of cold wave weather is transformed into a mathematical expression and embedded into the loss function of a generative adversarial network model. A cold wave sample scene generator is obtained by training on a small sample set of cold wave weather scenarios.

[0055] As a further improvement of the present invention, the generative adversarial network model in the generation module includes a generator G and a discriminator D. The generator G is used to generate a set of scenarios with a similar distribution to real cold wave wind power operation scenarios, and the discriminator D is used to distinguish the set of scenarios generated by the generator G from the real sample set. Both the generator G and the discriminator D networks are built from fully connected layers, and the calculation formula for the fully connected layer is as follows:

[0056]

[0057] Where, x (k) W represents the feature vector of the k-th layer. (k) Let b be the weight vector of the k-th layer. (k) This is the bias of the k-th layer.

[0058] As a further improvement of the present invention, the loss function W(D,G,L) of the generative adversarial network is:

[0059]

[0060] Among them, L cold Define the loss function for the embedding model for cold wave weather; E z p represents the expected distribution of sample z. data (x) represents the probability distribution of the true sample x, p z (z) represents the probability distribution of the generated sample z, G(z) represents the generator neural network, and D(x) represents the discriminator neural network;

[0061] Lcold It consists of two parts:

[0062] L cold1 =|T max -T min -T 降温 | 2

[0063] L cold2 =|T min -T 低温 | 2

[0064] Among them, T 降温 To determine the temperature drop conditions for a cold wave, T 低温 To determine the lowest temperature during a cold wave, T max The highest temperature during the cold wave was T. min The lowest temperature during the cold wave.

[0065] As a further improvement of the present invention, the construction module, which involves constructing a cold wave feature information graph based on graph theory knowledge and combining it with the cold wave wind power operation scenario samples, includes:

[0066] A graph data representation in graph theory is G(A,X), where A represents the connectivity matrix and X represents the node features. The feature data D = {D...} from the cold wave weather sample database... 1，n D 2，n ...D 96，n Each time segment in the graph is treated as a graph node, thus forming a graph data node. The graph nodes are connected in a fully connected manner to form a cold wave characteristic information graph.

[0067] As a further improvement of the present invention, the step of training the basic model with the constructed cold wave feature information map to establish a nonlinear mapping from the cold wave feature information map to the wind power output value, thereby obtaining a short-term wind power output prediction model, includes:

[0068] A graph attention network is built based on the graph convolutional neural network framework to learn and extract features from the cold wave feature information graph data, establish a nonlinear mapping from graph data to wind power output values, and use short-term wind power output prediction under cold wave weather as the output result to train and obtain a short-term wind power output prediction model.

[0069] As a further improvement of the present invention, the establishment of the nonlinear mapping from the cold wave characteristic information map to the wind power output value includes:

[0070] In the cold wave feature information map, each node represents a different time segment. The information between the nodes of the cold wave feature information map is aggregated by using a multi-layer graph attention layer. Then, a fully connected layer is used to transform the dimension of the features to obtain a non-linear mapping from the cold wave feature information map to wind power output.

[0071] The method for aggregating information between nodes in a cold wave feature graph using a multi-layer graph attention layer is as follows:

[0072]

[0073]

[0074] In the formula, parameter W is used to perform feature dimension transformation for each node, parameter a is used to calculate the correlation weights between nodes, || denotes vector concatenation, and a i，j Let σ represent the weights between nodes i and j calculated under a, and h represent the nonlinear activation function. i The updated feature information for node i is given, where k is an integer.

[0075] A fourth aspect of the present invention is to provide a short-term wind power output prediction system, comprising:

[0076] The fusion module is used to fuse wind farm output data and meteorological data for the predicted day to obtain a sample of wind power operation scenarios during the predicted cold wave.

[0077] The graph theory module is used to construct a feature information graph of the cold wave to be predicted based on graph theory knowledge and combined with the wind power operation scenario samples to be predicted.

[0078] The prediction module is used to input the cold wave feature information map to be predicted into the short-term wind power output prediction model trained by the method to obtain the short-term prediction result of wind power output under cold wave weather.

[0079] A fifth aspect of the present 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, wherein the processor executes the computer program to implement the wind power output short-term prediction model training method or the wind power output short-term prediction method.

[0080] A sixth aspect of the present invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the wind power output short-term prediction model training method or the wind power output short-term prediction method.

[0081] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0082] This invention proposes a method for generating wind power operation scenario samples during cold waves. It integrates the meteorological mechanism of cold wave weather into a generative adversarial network to achieve an effective combination of "knowledge + model". This guides the model to generate a sample scenario set that is more similar to the real scenario set, thereby expanding the number of wind power operation scenario samples. The constructed graph convolutional neural network framework builds a graph attention network model. The graph attention network model has powerful feature aggregation and mining capabilities. This network can effectively represent the temporal and feature correlation between wind power sample data in the form of a graph, and deeply mine the mapping relationship between known data and the power to be predicted, thereby achieving high-precision short-term prediction of wind power output. Attached Figure Description

[0083] Figure 1 This is a flowchart of a method for constructing a short-term wind power output prediction model according to the present invention;

[0084] Figure 2 This is a flowchart of a short-term wind power output prediction method according to the present invention;

[0085] Figure 3 This is a flowchart of the steps of a short-term wind power output prediction method in an embodiment of the present invention;

[0086] Figure 4 This is a schematic diagram of a short-term wind power output prediction system in an embodiment of the present invention;

[0087] Figure 5 This is a schematic diagram of an electronic device provided by the present invention. Detailed Implementation

[0088] 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.

[0089] To address the impact of extreme weather characteristics such as strong winds and freezing rain during cold waves on wind power output prediction, and considering the scarcity of wind power operation samples during cold waves, which makes it difficult to train wind power output prediction models and thus hinders short-term wind power output prediction, most current wind power output prediction methods for cold waves have failed to adequately solve the problem of sparse samples.

[0090] like Figure 1 As shown, the first objective of this invention is to provide a short-term wind power output prediction method, comprising the following steps:

[0091] S1. Select wind farm operation scenarios under cold wave weather from the pre-built historical feature database of wind farms to form a small sample scenario set for cold wave weather.

[0092] S2, input the small sample scene set of cold wave weather into the pre-trained cold wave sample scene generator to generate cold wave wind power operation scene samples;

[0093] S3. Based on graph theory, construct a cold wave feature information graph by combining the cold wave wind power operation scenario samples;

[0094] S4. Combine the constructed cold wave feature information map to train the basic model, establish a nonlinear mapping from the cold wave feature information map to the wind power output value, and obtain a short-term wind power output prediction model. The basic model is based on a graph attention network model built on a graph convolutional neural network framework.

[0095] This invention addresses the problems existing in cold wave weather sample generation and short-term wind power output prediction. It studies a method of embedding the meteorological mechanism of cold wave weather into a generative adversarial network to achieve a soft combination of knowledge and model, effectively improving the similarity rate of sample generation. At the same time, for the wind power prediction task, a graph attention network is constructed. This network model has powerful feature aggregation and mining capabilities, deeply mining the mapping relationship between known data and the power to be predicted, thereby achieving high-precision short-term prediction of wind power output under cold wave weather.

[0096] As an optional solution, the pre-built historical feature database of wind farms is obtained by integrating historical wind farm output data and meteorological data.

[0097] Optionally, the training method for the pre-trained cold wave sample scene generator includes: converting the definition of cold wave weather into a mathematical expression and embedding it into the loss function of the generative adversarial network model, and obtaining the cold wave sample scene generator by training on the small sample scene set of cold wave weather.

[0098] As an optional approach, the basic model is trained by combining the constructed cold wave feature information map to establish a nonlinear mapping from the cold wave feature information map to the wind power output value, thereby obtaining a short-term wind power output prediction model, including:

[0099] A graph attention network is built based on the graph convolutional neural network framework to learn and extract features from the cold wave feature information graph data, establish a nonlinear mapping from graph data to wind power output values, and use short-term wind power output prediction under cold wave weather as the output result to train and obtain a short-term wind power output prediction model.

[0100] The generation of scenario samples needs to be combined with actual meteorological mechanisms. Generative adversarial networks are currently widely used sample generation models. This invention successfully embeds the meteorological mechanism of cold wave weather using a loss function, realizing a soft combination of knowledge and model, and effectively improving the similarity rate of sample generation. Graph attention network models have powerful feature aggregation and mining capabilities and have been widely used in many fields, achieving good results. This invention applies graph attention networks to wind power output prediction tasks, effectively improving the accuracy of wind power output prediction.

[0101] Furthermore, once the short-term wind power output prediction model is obtained, the necessary data can be input into the model for prediction, eliminating the need to retrain the model for later predictions and improving processing efficiency. Subsequent updates, based on parameter changes or prediction result feedback, can further optimize the model, resulting in a continuously improved model through local optimization.

[0102] As a further improvement, the establishment of the nonlinear mapping from cold wave characteristic information map to wind power output value includes:

[0103] In the cold wave feature information map, each node represents a different time segment. A multi-layer graph attention layer aggregates the information between nodes in the cold wave feature information map. Then, a fully connected layer performs dimensionality transformation on the features, resulting in a nonlinear mapping from the cold wave feature information map to wind power output. This improves the model's accuracy and enables high-precision short-term prediction of wind power output.

[0104] This invention studies a method for mining the meteorological mechanism of cold wave weather, and studies an effective combination method of meteorological mechanism and sample generation model to realize the generation of cold wave weather samples based on "knowledge + model", which greatly improves the accuracy of sample generation. Based on this, a short-term high-precision prediction method for final wind power output is studied.

[0105] Combination Figure 2 and Figure 3 The specific steps of the embodiments of the present invention are described below:

[0106] Step 1: Integrate historical wind farm output data and meteorological data to construct a historical characteristic database of wind farms;

[0107] As an example of this embodiment, step 1 includes the following steps:

[0108] 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 fused together based on the time information, resulting in a data set D = {D...} 1，n D 2，n ...D t，n}, where D t,n ={Dt,P D t,W ... D t,T The data at time t represents the power, wind speed, wind direction, humidity, temperature, air pressure, and air density, while t is the data length, thus forming a historical feature database for the wind farm. As a specific example, the value of t typically ranges from one to three years.

[0109] Step 2: Based on the definition of cold wave weather, select wind farm operation scenarios under cold wave weather from the historical feature database of wind farms to form a small sample scenario set for cold wave weather.

[0110] As an example of this embodiment, step 2 includes the following steps:

[0111] For example, it can be determined according to the national cold wave weather standard: a temperature drop of more than 10°C within 24 hours, or a temperature drop of more than 12°C within 48 hours, while the minimum temperature is below 4°C.

[0112] Therefore, using a 24-hour time window, temperature data is extracted from the historical characteristic database of wind farms. Assuming a data granularity of 15 minutes, a day contains 96 time segments, and the temperature data for each segment is: T = {T1, T2, ... T} 96}

[0113] Based on cold wave weather standards, filter the time periods within the time window where temperatures drop: If during that time period, T max -T min ≥10℃ and T min If the temperature is ≤4℃, it is determined that a cold wave occurred on that day, and the relevant wind farm operation data for that day, D={D 1，n D 2，n ...D 96，n The extracted data serves as the operating scenario for wind farms during cold waves, thus forming a small sample scenario set for cold wave weather.

[0114] Step 3: Transform the definition of cold wave weather into a mathematical expression and embed it into the loss function of the generative adversarial network model. A high-performance cold wave sample scene generator can be obtained by training with a small number of cold wave sample scene sets, so as to generate cold wave wind power operation scene samples.

[0115] As an example of this embodiment, step 3 includes the following steps:

[0116] A generative adversarial network (GAN) was built to generate a large number of cold wave wind power operation scenario samples. The key to its functionality lies in the construction of the network loss function. In addition to the model's own loss function, in order to guide the model to generate operation samples that are more in line with the real cold wave scenario, the temperature change conditions described in the definition of cold wave weather were embedded into the loss function in the form of mathematical expression. The GAN was then trained with the extracted small sample scenario set of cold wave weather to achieve the generation of cold wave wind power operation scenario samples.

[0117] The construction of generative adversarial networks includes:

[0118] Generative Adversarial Networks (GANs) consist of a generator G and a discriminator D. Generator G generates a set of scenarios with a similar distribution to real-world cold-wave wind power operation scenarios, while discriminator D distinguishes the scenario set generated by generator G from the real sample set. Both generator G and discriminator D are constructed using fully connected layers, and the calculation formula for a fully connected layer is as follows:

[0119]

[0120] Among them, x (k+1) Let x represent the feature vector of the (k+1)th layer. (k) W represents the feature vector of the k-th layer, where k is an integer. (k) Let b be the weight vector of the k-th layer. (k) This is the bias of the k-th layer.

[0121] The construction of the loss function includes:

[0122] The loss function of generative adversarial networks is:

[0123] L total =L model +L cold

[0124] Among them, L model L is the model loss function for generative adversarial networks. cold Define the loss function for the embedded model for cold wave weather.

[0125] For example, in a generative adversarial network (GAN), the loss function L of the GAN model is determined based on the training objectives of the generator G and the discriminator D. model The loss function L of the generator and discriminator G and L D The components are as follows:

[0126]

[0127]

[0128] Among them, Lcold Define the loss function for the embedding model for cold wave weather; E x p represents the expected distribution of sample x. data (x) represents the probability distribution of the true sample x, p z (z) represents the probability distribution of the generated sample z, G represents the generator neural network, and D represents the discriminator neural network;

[0129] Based on the above equation, the objective function of generative adversarial networks can be derived:

[0130]

[0131] Cold wave weather defines the loss function L of the embedded model. cold It consists of two parts:

[0132] L cold1 =|T max -T min -10| 2

[0133] L cold2 =|T min -4| 2

[0134] Among them, T 降温 To determine the temperature drop conditions for a cold wave, this embodiment selects 10℃, T 低温 In this example, the minimum temperature condition for determining cold wave weather is 4℃. max The highest temperature during the cold wave was T. min The lowest temperature during the cold wave.

[0135] Therefore, the goal of the mechanism-embedded model is not only to minimize the model loss of the generative adversarial network, but also, guided by physical mechanisms, to ensure that the model-generated scenario data considers the actual physical meaning of cold wave weather. Its objective function can be written as:

[0136]

[0137] Step 4: Based on graph theory, construct a cold wave feature information map by combining cold wave wind power operation scenario samples.

[0138] As an example of this embodiment, step 4 includes the following steps:

[0139] A graph data representation in graph theory is G(A,X), where A represents the connectivity matrix and X represents the node features. The feature data D = {D...} from the cold wave weather sample database... 1，n D 2，n ...D 96，nEach time segment in the graph is treated as a graph node, thus forming a graph data node. The graph nodes are connected in a fully connected manner to form a cold wave characteristic information graph.

[0140] Step 5: Based on graph theory, construct a cold wave feature information map by combining cold wave wind power operation scenario samples.

[0141] As an example of this embodiment, step 5 includes the following steps:

[0142] A graph attention network is built based on the graph convolutional neural network framework to learn and extract features from the graph data of cold wave feature information. A nonlinear mapping is established from graph data to wind power output values. The short-term prediction results of wind power output under cold wave weather are output. The short-term prediction model of wind power output is trained to achieve short-term prediction of wind power output under cold wave weather.

[0143] The construction of a graph attention network includes:

[0144] In the constructed cold wave feature information map, each node represents a different time segment. Multi-layer graph attention layers are used to aggregate information between nodes in the cold wave feature information map. Then, a fully connected layer is used to perform dimensionality transformation on the features, achieving a non-linear mapping from graph data to wind power output. The method for aggregating node information in the cold wave feature information map is as follows:

[0145]

[0146]

[0147] Here, parameter W is used to perform feature dimension transformation for each node, parameter a is used to calculate the correlation weights between nodes, and || represents vector concatenation. i，j Let σ represent the weights between nodes i and j calculated under a, and h represent the nonlinear activation function. i The updated feature information for node i is given, where k is an integer.

[0148] The method of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0149] Example 1:

[0150] like Figure 1 As shown in the figure, this embodiment provides a short-term wind power output prediction method, including the following steps:

[0151] Step 1: To construct a historical characteristic database of wind farms, data collection is required. Historical power output data of wind farms needs to be obtained from the metering system, and meteorological data of the same time and granularity needs to be obtained from the weather station, including wind speed, wind direction, humidity, temperature, air pressure, and air density. All relevant data are then fused together based on time information. D = {D} 1，nD 2，n ...D t，n}, where D t,n ={D t,P D t,W ... D t,T} represents the data characteristics such as power, wind speed, wind direction, humidity, temperature, air pressure, and air density at time t, where t is the data length, typically ranging from one to three years, thus forming a historical feature database of the wind farm.

[0152] Step 2: Construct a small sample scenario set for cold wave weather. First, based on the national cold wave weather standard: a temperature drop of more than 10℃ within 24 hours, or a temperature drop of more than 12℃ within 48 hours, with the minimum temperature below 4℃. Therefore, using a 24-hour time window, temperature data is extracted from the historical characteristic database of wind farms. The granularity of the data used in the example is 15 minutes, resulting in 96 time segments per day. Each temperature segment is defined as: T = {T1, T2, ... T...} 96 Based on cold wave weather standards, filter the time period within the time window for periods of temperature drop: If during that time period, T max -T min ≥10℃ and T min If the temperature is ≤4℃, it is determined that a cold wave occurred on that day, and the relevant wind farm operation data for that day, D={D 1，n D 2，n ...D 96，n The extracted data serves as the operating scenario for wind farms during cold waves, thus forming a small sample scenario set for cold wave weather.

[0153] Step 3: Build a generative adversarial network to generate a large number of cold wave wind power operation scenario samples. The key to its function is the construction of the network loss function. In addition to the model's own loss function, in order to guide the model to generate operation samples that are more in line with the real cold wave scenario, the temperature change conditions described in the definition of cold wave weather are embedded into the loss function in the form of mathematical expression. The generative adversarial network is trained in combination with the extracted small sample scenario set of cold wave weather to realize the generation of cold wave wind power operation scenario samples.

[0154] Step 4: Based on graph theory, construct a cold wave feature information map by combining cold wave wind power operation scenario samples.

[0155] A graph data representation in graph theory is G(A,X), where A represents the connectivity matrix and X represents the node features. The feature data D = {D...} from the cold wave weather sample database... 1，n D 2，n ...D 96，nEach time segment in the graph is treated as a graph node, thus forming a graph data node. The graph nodes are connected in a fully connected manner to form a cold wave characteristic information graph.

[0156] Step 5: Based on graph theory, construct a cold wave feature information map by combining cold wave wind power operation scenario samples.

[0157] A graph attention network is built based on the graph convolutional neural network framework to learn and extract features from the graph data of cold wave feature information. A nonlinear mapping is established from graph data to wind power output values. The short-term prediction results of wind power output under cold wave weather are output. The short-term prediction model of wind power output is trained to achieve short-term prediction of wind power output under cold wave weather.

[0158] The parameters selected for the Generative Adversarial Network are as follows: both the generator G and the discriminator D consist of fully connected layers. The generator G has three fully connected layers with 500, 250, and 96 nodes respectively, and the discriminator D has three fully connected layers with 500, 250, and 96 nodes respectively. The parameters selected for the Graph Attention Network are: two graph attention layers with 15 hidden nodes, and three fully connected layers with 500, 250, and 96 nodes respectively.

[0159] Model training and testing: Based on the constructed generative adversarial network and graph attention network model, the generative adversarial network is first trained to generate cold wave wind power operation scenario samples, and the graph attention network model is trained based on the generated samples. The model learns and extracts features from the cold wave feature map, establishes a nonlinear mapping from graph data to wind power output values, and divides the training samples in the training set into several batches (set to 50 in this embodiment) for training to obtain a model with good parameters, and finally achieves high-precision short-term prediction of wind power output under cold wave weather.

[0160] Example 2:

[0161] like Figure 1 As shown in the figure, this embodiment provides a short-term wind power output prediction method, including the following steps:

[0162] Step 1: To construct a historical characteristic database of wind farms, data collection is required. Historical power output data of wind farms needs to be obtained from the metering system, and meteorological data of the same time and granularity needs to be obtained from the weather station, including wind speed, wind direction, humidity, temperature, air pressure, and air density. All relevant data are then fused together based on time information. D = {D} 1，n D 2，n ...D t，n}, where D t,n ={D t,P D t,W ... D t,T} represents the data characteristics such as power, wind speed, wind direction, humidity, temperature, air pressure, and air density at time t, where t is the data length, typically ranging from one to three years, thus forming a historical feature database of the wind farm.

[0163] Step 2: Construct a small sample scenario set for cold wave weather. First, based on the national cold wave weather standard: a temperature drop of more than 10℃ within 24 hours, or a temperature drop of more than 12℃ within 48 hours, with the minimum temperature below 4℃. Therefore, using a 24-hour time window, temperature data is extracted from the historical characteristic database of wind farms. The granularity of the data used in the example is 30 minutes, meaning a day contains 48 time segments. Each temperature segment is then defined as: T = {T1, T2, ... T...} 48 Based on cold wave weather standards, filter the time period within the time window for periods of temperature drop: If during that time period, T max -T min ≥10℃ and T min If the temperature is ≤4℃, it is determined that a cold wave occurred on that day, and the relevant wind farm operation data for that day, D={D 1，n D 2，n ...D 48，n The extracted data serves as the operating scenario for wind farms during cold waves, thus forming a small sample scenario set for cold wave weather.

[0164] Step 3: Build a generative adversarial network to generate a large number of cold wave wind power operation scenario samples. The specific steps are the same as step 3 in Example 1.

[0165] Step 4: Based on graph theory, construct a cold wave feature information map by combining cold wave wind power operation scenario samples. The specific steps are the same as step 4 in Example 1.

[0166] Step 5: Based on graph theory, construct a cold wave feature information map using samples of cold wave wind power operation scenarios. The specific steps are the same as step 5 in Example 1. The differences are as follows:

[0167] The parameters used in the Generative Adversarial Network (GAN) are as follows: both the generator G and the discriminator D consist of fully connected layers. The generator G has three fully connected layers with 500, 250, and 48 nodes respectively, and the discriminator D has three fully connected layers with 500, 250, and 48 nodes respectively. The parameters used in the Graph Attention Network (GAN) are as follows: two graph attention layers with 15 hidden nodes, and three fully connected layers with 500, 250, and 48 nodes respectively.

[0168] The training samples in the training set are divided into several batches (50 in this embodiment) for training to obtain a model with good parameters, and finally achieve high-precision short-term prediction of wind power output under cold wave weather.

[0169] Example 3:

[0170] like Figure 1 As shown in the figure, this embodiment provides a short-term wind power output prediction method, including the following steps:

[0171] Step 1: To construct a historical characteristic database of wind farms, data collection is required. Historical power output data of wind farms needs to be obtained from the metering system, and meteorological data of the same time and granularity needs to be obtained from the weather station, including wind speed, wind direction, humidity, temperature, air pressure, and air density. All relevant data are then fused together based on time information. D = {D} 1，n D 2，n ...D t，n}, where D t,n ={D t,P D t,W ... D t,T} represents the data characteristics such as power, wind speed, wind direction, humidity, temperature, air pressure, and air density at time t, where t is the data length, typically ranging from one to three years, thus forming a historical feature database of the wind farm.

[0172] Step 2: Construct a small sample scenario set for cold wave weather. First, based on the national cold wave weather standard: a temperature drop of more than 10℃ within 24 hours, or a temperature drop of more than 12℃ within 48 hours, with the minimum temperature below 4℃. Therefore, using a 24-hour time window, temperature data is extracted from the historical characteristic database of wind farms. The granularity of the data used in the example is 60 minutes, resulting in 24 time segments per day. Each temperature segment is defined as: T = {T1, T2, ... T...} 24 Based on cold wave weather standards, filter the time period within the time window for periods of temperature drop: If during that time period, T max -T min ≥10℃ and T min If the temperature is ≤4℃, it is determined that a cold wave occurred on that day, and the relevant wind farm operation data for that day, D={D 1，n D 2，n ...D 24，n The extracted data serves as the operating scenario for wind farms during cold waves, thus forming a small sample scenario set for cold wave weather.

[0173] Step 3: Build a generative adversarial network to generate a large number of cold wave wind power operation scenario samples. The specific steps are the same as step 3 in Example 1.

[0174] Step 4: Based on graph theory, construct a cold wave feature information map by combining cold wave wind power operation scenario samples. The specific steps are the same as step 4 in Example 1.

[0175] Step 5: Based on graph theory, construct a cold wave feature information map using samples of cold wave wind power operation scenarios. The specific steps are the same as step 5 in Example 1. The differences are as follows:

[0176] The parameters used in the Generative Adversarial Network (GAN) are as follows: both the generator G and the discriminator D consist of fully connected layers. The generator G has three fully connected layers with 500, 250, and 24 nodes respectively, and the discriminator D has three fully connected layers with 500, 250, and 24 nodes respectively. The parameters used in the Graph Attention Network (GAN) are as follows: two graph attention layers with 15 hidden nodes, and three fully connected layers with 500, 250, and 24 nodes respectively.

[0177] The training samples in the training set are divided into several batches (50 in this embodiment) for training to obtain a model with good parameters, and finally achieve high-precision short-term prediction of wind power output under cold wave weather.

[0178] The second objective of this invention is to provide a short-term wind power output prediction method, comprising:

[0179] By integrating wind farm output data and meteorological data for the predicted date, a sample of wind power operation scenarios for the predicted cold wave is obtained.

[0180] Based on graph theory, a feature information graph of the cold wave to be predicted is constructed by combining the wind power operation scenario samples of the cold wave to be predicted.

[0181] The cold wave feature information map to be predicted is input into the short-term wind power output prediction model trained by the method described above to obtain the short-term prediction result of wind power output under cold wave weather.

[0182] Based on the models obtained in Examples 1 to 3, a short-term wind power output prediction method is proposed. Compared with other existing prediction methods, the sample of wind power operation during cold waves generated by this method has higher similarity and higher short-term wind power output prediction under cold wave weather.

[0183] like Figure 4 As shown, a short-term wind power output prediction model training system includes:

[0184] The filtering module is used to filter wind farm operation scenarios under cold wave weather from a pre-built historical feature database of wind farms, forming a small sample scenario set for cold wave weather.

[0185] The generation module is used to input the small sample scene set of cold wave weather into a pre-trained cold wave sample scene generator to generate cold wave wind power operation scene samples.

[0186] The module is used to construct a cold wave feature information map based on graph theory knowledge and combined with the cold wave wind power operation scenario samples;

[0187] The training module is used to train the basic model by combining the constructed cold wave feature information map, establish a nonlinear mapping from the cold wave feature information map to the wind power output value, and obtain a short-term wind power output prediction model. The basic model is based on a graph attention network model built on a graph convolutional neural network framework.

[0188] In some instances, the pre-built historical feature database of wind farms in the filtering module is obtained by fusing historical wind farm output data and meteorological data. Specific methods include:

[0189] 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 meteorological station. All relevant data are then fused together based on the time information, resulting in D = {D...} 1，n D 2，n ...D t，n}, where D t,n ={D t,P D t,W ... D t,T 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.

[0190] The process involves selecting wind farm operation scenarios under cold wave weather from a pre-built historical feature database of wind farms to form a small sample scenario set for cold wave weather, including:

[0191] Using M as the time window, temperature data is extracted from a pre-built historical feature database of wind farms. A day contains m time segments, and each temperature segment is: T = {T1, T2, ... T} k}, k = M / m;

[0192] Based on the requirements for cold wave weather, select the time period within the time window where the temperature will drop: t i When the temperature drops, if, during this period, the weather is determined to be a cold wave based on the cold wave weather criteria, then the relevant wind farm operation data for that day, D = {D}, is used. 1，n D 2，n ...D k，n The extracted data serves as the operating scenario for wind farms during cold waves, thus forming a small sample scenario set for cold wave weather.

[0193] In some instances, the training method for the pre-trained cold wave sample scene generator in the generation module includes:

[0194] The definition of cold wave weather is transformed into a mathematical expression and embedded into the loss function of a generative adversarial network model. A cold wave sample scene generator is obtained by training on a small sample set of cold wave weather scenarios.

[0195] The generative adversarial network model includes a generator G and a discriminator D. Generator G generates a set of scenarios with a similar distribution to real cold-wave wind power operation scenarios, while discriminator D distinguishes the scenario set generated by generator G from the real sample set. Both generator G and discriminator D are constructed from fully connected layers, and the calculation formula for a fully connected layer is as follows:

[0196]

[0197] Where, x (k) W represents the feature vector of the k-th layer. (k) Let b be the weight vector of the k-th layer. (k) This is the bias of the k-th layer.

[0198] Optionally, the loss function of the generative adversarial network is:

[0199]

[0200] Among them, L cold Define the loss function for the embedding model for cold wave weather; E represents the expected distribution of the samples, p data (x) represents the probability distribution of the true sample x, p z (x) represents the probability distribution of the generated sample z;

[0201] L cold It consists of two parts:

[0202] L cold1 =|T max -T min -T 降温 | 2

[0203] L cold2 =|T min -T 低温 | 2

[0204] Among them, T 降温 To determine the temperature drop conditions for a cold wave, T 低温 Determining the lowest temperature conditions during a cold wave.

[0205] In some instances, the construction module, which involves constructing a cold wave feature information graph based on graph theory knowledge and combining it with samples of cold wave wind power operation scenarios, includes:

[0206] A graph data representation in graph theory is G(A,X), where A represents the connectivity matrix and X represents the node features. The feature data D = {D...} from the cold wave weather sample database... 1，n D 2，n ...D 96，n Each time segment in the graph is treated as a graph node, thus forming a graph data node. The graph nodes are connected in a fully connected manner to form a cold wave characteristic information graph.

[0207] Optionally, the step of training the basic model using the constructed cold wave feature information map to establish a nonlinear mapping from the cold wave feature information map to the wind power output value, thereby obtaining a short-term wind power output prediction model, includes:

[0208] A graph attention network is built based on the graph convolutional neural network framework to learn and extract features from the cold wave feature information graph data, establish a nonlinear mapping from graph data to wind power output values, and use short-term wind power output prediction under cold wave weather as the output result to train and obtain a short-term wind power output prediction model.

[0209] More specifically, the establishment of the nonlinear mapping from cold wave characteristic information map to wind power output value includes:

[0210] In the cold wave feature information map, each node represents a different time segment. The information between the nodes of the cold wave feature information map is aggregated by using a multi-layer graph attention layer. Then, a fully connected layer is used to transform the dimension of the features to obtain a non-linear mapping from the cold wave feature information map to wind power output.

[0211] The method for aggregating information between nodes in a cold wave feature graph using a multi-layer graph attention layer is as follows:

[0212]

[0213]

[0214] In the formula, parameter W is used to perform feature dimension transformation for each node, parameter a is used to calculate the correlation weights between nodes, || denotes vector concatenation, and a i，j Let σ represent the weights between nodes i and j calculated under a, and h represent the nonlinear activation function. i The updated feature information for node i is given, where k is an integer.

[0215] The fourth objective of this invention is to provide a short-term wind power output prediction system, comprising:

[0216] The fusion module is used to fuse wind farm output data and meteorological data for the predicted day to obtain a sample of wind power operation scenarios during the predicted cold wave.

[0217] The graph theory module is used to construct a feature information graph of the cold wave to be predicted based on graph theory knowledge and combined with the wind power operation scenario samples to be predicted.

[0218] The prediction module is used to input the cold wave feature information map to be predicted into the trained short-term wind power output prediction model to obtain the short-term prediction results of wind power output under cold wave weather.

[0219] A fifth 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, wherein the processor executes the computer program to implement the wind power output short-term prediction model training method or the wind power output short-term prediction method.

[0220] The sixth 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 power output short-term prediction model training method or the wind power output short-term prediction method.

[0221] 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.

[0222] 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.

[0223] 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 1The function specified in one or more boxes.

[0224] 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.

[0225] 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 protection scope of the claims of the present invention.

Claims

1. A training method for a short-term wind power output prediction model, characterized in that, include: Wind farm operation scenarios under cold wave weather were selected from a pre-built historical feature database of wind farms to form a small sample scenario set for cold wave weather. The small sample set of cold wave weather scenarios is input into a pre-trained cold wave sample scenario generator to generate cold wave wind power operation scenario samples. Based on graph theory, a cold wave feature information graph is constructed by combining the cold wave wind power operation scenario samples. By training the basic model with the constructed cold wave feature information map, a nonlinear mapping from the cold wave feature information map to the wind power output value is established, and a short-term prediction model for wind power output is obtained. The basic model is a graph attention network model built based on the graph convolutional neural network framework. The training method for the pre-trained cold wave sample scene generator includes: The definition of cold wave weather is transformed into a mathematical expression and embedded into the loss function of the generative adversarial network model. A cold wave sample scene generator is obtained by training on a small sample set of cold wave weather scenarios. The loss function of the generative adversarial network W ( D,G,L )for: in, L cold Define the loss function for the embedded model for cold wave weather; E z Indicates sample z The expected distribution of the distribution p data ( x ) represents the real sample x The probability distribution, p z ( z ) indicates the generation of samples z The probability distribution, G ( z ) represents a generator neural network. D ( x ) represents a discriminator neural network; L cold It consists of two parts: in, To determine the conditions for temperature drop during a cold wave, To determine the minimum temperature conditions for a cold wave, T max This is the highest temperature during the cold wave. T min The lowest temperature during the cold wave.

2. The method for training a short-term wind power output prediction model according to claim 1, characterized in that, The pre-built historical characteristic database of wind farms is obtained by integrating historical wind farm output data and meteorological data. Specific methods include: Historical power output data of wind farms are 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 meteorological station. The data characteristics of all meteorological data are then integrated based on time information to form a historical feature database of wind farms.

3. The method for training a short-term wind power output prediction model according to claim 1, characterized in that, The process involves selecting wind farm operation scenarios under cold wave weather from a pre-built historical feature database of wind farms to form a small sample scenario set for cold wave weather, including: Using M as the time window, temperature data is extracted from a pre-built historical feature database of wind farms. A day contains m time segments, and the temperature data for each segment is as follows: k=M / m; Based on the requirements for cold wave weather, select the time period within the time window where the temperature will drop: , t i If, during this period, the weather is determined to be a cold wave based on the criteria for cold wave weather, and the meteorological data related to the operation of the wind farm on that day is extracted, it will be used as the operating scenario of the wind farm under the cold wave weather, thus forming a small sample scenario set for cold wave weather.

4. The method for training a short-term wind power output prediction model according to claim 1, characterized in that, The generative adversarial network model includes a generator and a discriminator. The generator generates a set of scenarios with a similar distribution to real cold-wave wind power operation scenarios, while the discriminator distinguishes the generated scenario set from the real sample set. Both the generator G and the discriminator networks are built from fully connected layers, and the calculation formula for the fully connected layer is as follows: in, Indicates the first k +1 layer feature vectors, Indicates the first k Layer feature vectors, k It is an integer. For the first k Layer weight vector, For the first k Layer bias.

5. The method for training a short-term wind power output prediction model according to claim 1, characterized in that, The construction of a cold wave feature information map based on graph theory knowledge and combined with the cold wave wind power operation scenario samples includes: A graph data representation in graph theory is as follows ,in A Represents the connection matrix. X Representing node features, this involves using feature data from the cold wave weather sample database. Each time segment 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 cold wave characteristic information graph.

6. The method for training a short-term wind power output prediction model according to claim 1, characterized in that, The basic model is trained by combining the constructed cold wave feature information map, and a nonlinear mapping from the cold wave feature information map to the wind power output value is established to obtain a short-term wind power output prediction model, including: A graph attention network is built based on the graph convolutional neural network framework to learn and extract features from the cold wave feature information graph data, establish a nonlinear mapping from graph data to wind power output values, and use short-term wind power output prediction under cold wave weather as the output result to train and obtain a short-term wind power output prediction model.

7. The method for training a short-term wind power output prediction model according to claim 6, characterized in that, The establishment of the nonlinear mapping from cold wave characteristic information map to wind power output value includes: In the cold wave feature information map, each node represents a different time segment. The information between the nodes of the cold wave feature information map is aggregated by using a multi-layer graph attention layer. Then, a fully connected layer is used to transform the dimension of the features to obtain a non-linear mapping from the cold wave feature information map to wind power output. The method for aggregating information between nodes in a cold wave feature graph using a multi-layer graph attention layer is as follows: In the formula, the 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, k It is an integer.

8. A method for short-term prediction of wind power output, characterized in that, include: By integrating wind farm output data and meteorological data for the predicted date, a sample of wind power operation scenarios for the predicted cold wave is obtained. Based on graph theory, a feature information graph of the cold wave to be predicted is constructed by combining the wind power operation scenario samples of the cold wave to be predicted. The cold wave feature information map to be predicted is input into the short-term wind power output prediction model trained by the method described in any one of claims 1 to 7 to obtain the short-term prediction result of wind power output under cold wave weather.

9. A training system for a short-term wind power output prediction model, characterized in that, include: The filtering module is used to filter wind farm operation scenarios under cold wave weather from a pre-built historical feature database of wind farms, forming a small sample scenario set for cold wave weather. The generation module is used to input the small sample scene set of cold wave weather into a pre-trained cold wave sample scene generator to generate cold wave wind power operation scene samples. The module is used to construct a cold wave feature information map based on graph theory knowledge and combined with the cold wave wind power operation scenario samples; The training module is used to train the basic model by combining the constructed cold wave feature information map, establish a nonlinear mapping from the cold wave feature information map to the wind power output value, and obtain a short-term wind power output prediction model. The basic model is based on a graph attention network model built on a graph convolutional neural network framework. In the generation module, the training method for the pre-trained cold wave sample scene generator includes: The definition of cold wave weather is transformed into a mathematical expression and embedded into the loss function of the generative adversarial network model. A cold wave sample scene generator is obtained by training on a small sample set of cold wave weather scenarios. The loss function of the generative adversarial network W ( D,G,L )for: in, L cold Define the loss function for the embedded model for cold wave weather; E z Indicates sample z The expected distribution of the distribution p data ( x ) represents the real sample x The probability distribution, p z ( z ) indicates the generation of samples z The probability distribution, G ( z ) represents a generator neural network. D ( x ) represents a discriminator neural network; L cold It consists of two parts: in, To determine the conditions for temperature drop during a cold wave, To determine the minimum temperature conditions for a cold wave, T max This is the highest temperature during the cold wave. T min The lowest temperature during the cold wave.

10. A wind power output short-term prediction model training system according to claim 9, characterized in that, In the filtering module, the pre-built historical feature database of wind farms is obtained by fusing historical wind farm output data and meteorological data. Specific methods include: Historical power output data of wind farms are 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 meteorological station. The data characteristics of all meteorological data are then integrated based on time information to form a historical feature database of wind farms.

11. The wind power output short-term prediction model training system according to claim 9, characterized in that, In the filtering module, the step of filtering wind farm operation scenarios under cold wave weather from a pre-built historical feature database of wind farms to form a small sample scenario set for cold wave weather includes: Using M as the time window, temperature data is extracted from a pre-built historical feature database of wind farms. A day contains m time segments, and the temperature data for each segment is as follows: k=M / m; Based on the requirements for cold wave weather, select the time period within the time window where the temperature will drop: , t i If, during this period, the weather is determined to be a cold wave based on the criteria for cold wave weather, and the meteorological data related to the operation of the wind farm on that day is extracted, it will be used as the operating scenario of the wind farm under the cold wave weather, thus forming a small sample scenario set for cold wave weather.

12. The wind power output short-term prediction model training system according to claim 11, characterized in that, In the generation module, the generative adversarial network model includes a generator G and a discriminator D. Generator G generates a set of scenarios with a similar distribution to real cold-wave wind power operation scenarios, while discriminator D distinguishes the scenario set generated by generator G from the real sample set. Both generator G and discriminator D are constructed from fully connected layers, and the calculation formula for the fully connected layer is as follows: in, Indicates the first k +1 layer feature vectors, Indicates the first k Layer feature vectors, k It is an integer. For the first k Layer weight vector, For the first k Layer bias.

13. The wind power output short-term prediction model training system according to claim 9, characterized in that, In the construction module, the construction of a cold wave feature information graph based on graph theory knowledge and combined with the cold wave wind power operation scenario samples includes: A graph data representation in graph theory is as follows ,in A Represents the connection matrix. X Representing node features, this involves using feature data from the cold wave weather sample database. Each time segment 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 cold wave characteristic information graph.

14. The wind power output short-term prediction model training system according to claim 9, characterized in that, The basic model is trained by combining the constructed cold wave feature information map, and a nonlinear mapping from the cold wave feature information map to the wind power output value is established to obtain a short-term wind power output prediction model, including: A graph attention network is built based on the graph convolutional neural network framework to learn and extract features from the cold wave feature information graph data, establish a nonlinear mapping from graph data to wind power output values, and use short-term wind power output prediction under cold wave weather as the output result to train and obtain a short-term wind power output prediction model.

15. A wind power output short-term prediction model training system according to claim 14, characterized in that, The establishment of the nonlinear mapping from cold wave characteristic information map to wind power output value includes: In the cold wave feature information map, each node represents a different time segment. The information between the nodes of the cold wave feature information map is aggregated by using a multi-layer graph attention layer. Then, a fully connected layer is used to transform the dimension of the features to obtain a non-linear mapping from the cold wave feature information map to wind power output. The method for aggregating information between nodes in a cold wave feature graph using a multi-layer graph attention layer is as follows: In the formula, the 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, k It is an integer.

16. A short-term wind power output prediction system, characterized in that, include: The fusion module is used to fuse wind farm output data and meteorological data for the predicted day to obtain a sample of wind power operation scenarios during the predicted cold wave. The graph theory module is used to construct a feature information graph of the cold wave to be predicted based on graph theory knowledge and combined with the wind power operation scenario samples to be predicted. The prediction module is used to input the cold wave feature information map to be predicted into the short-term wind power output prediction model trained by the method described in any one of claims 1 to 7, so as to obtain the short-term prediction result of wind power output under cold wave weather.

17. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the wind power output short-term prediction model training method according to any one of claims 1-7 or the wind power output short-term prediction method according to claim 8.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the wind power output short-term prediction model training method according to any one of claims 1-7 or the wind power output short-term prediction method according to claim 8.

Images