Wind farm group power prediction method and system based on multi-scale spatiotemporal correlation

Abstract

The present invention relates to the technical field of wind power prediction, and discloses a wind farm group power prediction method and system based on multi-scale spatiotemporal correlation, the method comprising obtaining the operation data, environmental data and geographic location data of each wind turbine, and performing preprocessing to generate a basic feature matrix; constructing a geographic location matrix, a historical wind speed similarity correlation matrix and a wake effect dynamic relationship matrix; performing multi-scale spatiotemporal correlation modeling on the basic feature matrix through a spatial feature fusion unit, and fusing it with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix; performing information aggregation on the in-degree neighbor nodes and out-degree neighbor nodes of the wind turbine to generate a node embedding representation of each wind turbine; inputting the node embedding representation of each wind turbine into a multi-scale time convolutional neural network prediction model for processing to obtain a wind power prediction value. The present invention solves the problem that the traditional method is not comprehensive in modeling the spatiotemporal correlation between wind turbines, and improves the power prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and more specifically, to a method and system for predicting wind farm power based on multi-scale spatiotemporal correlation. Background Art

[0002] As the global energy structure accelerates its transformation to clean energy, wind power, as a core component of renewable energy, has an increasingly significant impact on the dispatch optimization and safe operation of power systems. Wind farms are usually composed of dozens to hundreds of wind turbines. Their output power is affected by multi-dimensional meteorological factors such as wind speed, wind direction, temperature, and the coupling of wake effects and geographical distribution differences between units, showing complex multi-scale spatiotemporal correlation characteristics. However, most current prediction models still focus on single machines or local clusters, and do not adequately describe the global dynamic characteristics of wind farms and the correlation mechanisms across spatiotemporal scales, making it difficult to meet the stringent requirements for prediction accuracy and robustness in scenarios with a high proportion of new energy grid-connected.

[0003] In the existing technology, wind farm power prediction methods are mainly based on two paradigms: physical modeling and data-driven. Physical models (such as computational fluid dynamics models) rely on high-precision meteorological data and wind turbine parameters to simulate power output, but their computational complexity is high and their adaptability to complex terrain and dynamic wake effects is poor; data-driven models (such as long short-term memory neural networks, Transformer, etc.) achieve prediction by mining the temporal laws of historical data, but usually use fixed time windows or single spatial scale modeling, ignoring multi-time scale characteristics such as short-term turbulent fluctuations (seconds), medium- and long-term meteorological trends (hours) between wind turbine groups, as well as dynamic correlations across spatial scales such as wake interference between units and geographical distribution differences. For example, although traditional time series models can capture the time-varying laws of single-machine power, it is difficult to quantify the synergy and competition relationship of wind turbine clusters; although spatial modeling methods (such as convolutional neural networks in the figure) introduce topological structures, they do not effectively distinguish the contribution differences of different spatiotemporal scale features. The above method fails to fully integrate the multi-scale spatiotemporal dynamic correlation characteristics of wind farm fleets, resulting in incomplete spatiotemporal correlation modeling and low prediction accuracy. Summary of the invention

[0004] In order to overcome the defect of incomplete spatiotemporal correlation modeling in the prior art, which leads to low power prediction accuracy of wind farm groups, the present invention proposes the following technical solutions:

[0005] In a first aspect, the present invention proposes a wind farm group power prediction method based on multi-scale spatiotemporal correlation, comprising:

[0006] Obtain the operating data, environmental data and geographic location data of each wind turbine in the target wind farm, perform preprocessing, and generate a basic feature matrix;

[0007] Construct a geographical location matrix to characterize the spatial correlation between wind turbines, a historical wind speed similarity correlation matrix to characterize the similarity of meteorological evolution, and a wake effect dynamic relationship matrix to characterize the dynamic wake effect;

[0008] Based on the geographical location matrix and the historical wind speed similarity correlation matrix, the basic feature matrix is ​​modeled with multi-scale spatiotemporal correlation through the spatial feature fusion unit, and is fused with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix.

[0009] According to the multi-scale dynamic relationship matrix, respectively aggregate the information of the in-degree neighbor nodes and the out-degree neighbor nodes of the wind turbine to generate a node embedding representation of each wind turbine;

[0010] The node embedding of each wind turbine represents the input of the trained multi-scale time convolutional neural network prediction model for prediction processing to obtain the wind power prediction value.

[0011] As a preferred technical solution, the operation data, environmental data and geographical location data of each wind turbine are cleaned, normalized and feature extracted in turn to generate a basic feature matrix;

[0012] The expression of the basic feature matrix is ​​as follows:

[0013]

[0014] in, t is the time step, For fans i Real-time power output, For fans i The rotation speed, For fans i The blade angle, is the wind speed, For wind direction, For temperature, For humidity, For fans i The longitude coordinates of For fans i For the wind farm N Typhoon, overall characteristic matrix N × F The matrix of F is the number of all the above characteristics of a single fan.

[0015] As a preferred technical solution, a geographical location matrix is ​​constructed, including:

[0016] Based on the longitude and latitude coordinates of the wind turbines, the Euclidean distance between the wind turbines is calculated, and the distance map is generated through threshold screening to obtain the geographic location matrix , whose expression is as follows:

[0017]

[0018] in, For fans i The longitude coordinates of For fans i The latitude coordinate of For fans j The longitude coordinates of For fans j The latitude coordinate of is the set threshold.

[0019] As a preferred technical solution, a historical wind speed similarity correlation matrix is ​​constructed, including:

[0020] Based on the historical wind speed sequence of the wind turbines, the wind speed time series cumulative alignment distance between the wind turbines is calculated and normalized to obtain the historical wind speed similarity association matrix. , whose expression is as follows:

[0021]

[0022] in, D ( i,j ) is the fan calculated by the dynamic time warping algorithm i and fan j The cumulative alignment distance of wind speed time series is A preset parameter that controls the sensitivity of distance to similarity.

[0023] As a preferred technical solution, a wake effect dynamic relationship matrix is ​​constructed, including:

[0024] Based on the coordinate position, wind direction and wind speed information of the wind turbine, the Jensen wake model is used to generate the dynamic relationship matrix of the wake effect. , whose expression is as follows:

[0025]

[0026] in, α is the adjustment factor used to determine the impact strength of the upstream fan, Indicates fan i and fan j The wind direction angle between k represents the wake expansion coefficient, r i Indicates fan i The blade diameter, Indicates fan i and fanj The distance between is the axial flow interference factor of the upstream fan.

[0027] As a preferred technical solution, a dynamic time warping algorithm is used to calculate the cumulative alignment distance of wind speed time series between wind turbines, and its expression is as follows:

[0028]

[0029] in, Indicates the first i Line j The value of the column represents the minimum cumulative distance from the starting point of the sequence to the current point. For fans i Time Series At the moment i The value of For fans j Time Series At the moment j The value of N and M The time series X and time series Y Length, For time series X Middle i Points and time series Y Middle j The Euclidean distance between points.

[0030] As a preferred technical solution, based on the geographic location matrix and the historical wind speed similarity association matrix, the basic feature matrix is ​​modeled with multi-scale spatiotemporal association through the spatial feature fusion unit, and is fused with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix, including:

[0031] Based on the geographic location matrix Similarity correlation matrix with historical wind speed , using graph convolutional neural network to transform the feature matrix of basic feature data K Perform graph convolution to obtain intermediate feature representation H , whose expression is as follows:

[0032]

[0033] in, and are the normalized geographical location matrix and the historical wind speed similarity association matrix, I is the identity matrix, D is the degree matrix, W P andW S are the weight matrices of similarity between geographical location and historical wind speed, σ is the activation function;

[0034] The intermediate features are represented H After being input into the gated fusion unit for processing and linear transformation, the dynamic relationship matrix with the wake effect Weighted summation to obtain the multi-scale dynamic relationship matrix , whose expression is as follows:

[0035]

[0036] in, is the output of the update gate, U z is the weight parameter, b z is the bias parameter, r is the output of the overlap gate, U r is the weight parameter, b r is the bias parameter, represents a hyperbolic tangent transformation, It is the intermediate feature representation after being processed by the reset gate and hyperbolic tangent transform.

[0037] As a preferred technical solution, according to the multi-scale dynamic relationship matrix, information aggregation is performed on the in-degree neighbor nodes and out-degree neighbor nodes of the wind turbine respectively to generate a node embedding representation of each wind turbine, including:

[0038] According to the multi-scale dynamic relationship matrix, the attention weights are calculated for the out-degree neighbor nodes and in-degree neighbor nodes of the wind turbine respectively, and the expressions are as follows:

[0039]

[0040] in, and are the attention weights of the out-degree neighbor nodes and the in-degree neighbor nodes respectively, and The attention vectors that can be learned are: and are the linear transformation matrices, and Fan i and fan j The feature matrix of the basic feature data, and are independent learning parameters respectively;

[0041] The attention weights of out-degree neighbor nodes and in-degree neighbor nodes are normalized, and the expressions are as follows:

[0042]

[0043] in, and are the normalized attention weights of the out-degree neighbor nodes and the in-degree neighbor nodes, and Fan i The set of out-degree neighbor nodes and in-degree neighbor nodes;

[0044] Use the normalized attention weights to perform weighted aggregation on the information of out-degree neighbor nodes and in-degree neighbor nodes, and obtain the information aggregation representation of the out-degree neighbor nodes and the information aggregation representation of in-degree neighbor nodes , whose expression is as follows:

[0045]

[0046] in, σ is the activation function, Indicates fan i For outgoing neighbor nodes j The attention weight, Indicates fan i For outgoing neighbor nodes j The attention weight of

[0047] Information aggregation representation and information aggregation representation Perform summation to obtain the final node embedding representation , whose expression is as follows:

[0048] .

[0049] As a preferred technical solution, the node of each wind turbine is embedded into the input trained multi-scale time convolutional neural network prediction model for prediction processing to obtain the wind power prediction value, including:

[0050] Using node embedding representation Constructing the graph information matrix ;

[0051] Using the Graph Information Matrix Constructing the input sequence , where the input sequence The graph information matrix of T consecutive time steps constitute, N is the number of fans, Represents the feature dimension of each wind turbine after passing through the graph convolutional neural network;

[0052] In the multi-scale temporal convolutional neural network prediction model, several convolutional branch units with different expansion rates are used to transform the input feature sequence. Perform convolution operations to extract feature representations at different time scales. The expressions are as follows:

[0053]

[0054] in, Indicates that at time step t At that time, l The feature representation of the output of the convolutional branch unit, Indicates l The weight parameters of the convolutional branch units, Represents a graph with expansion rate d l The one-dimensional convolution operation, Represents the time step To time step t The input feature sequence is For the l The bias parameters of the convolutional branch units;

[0055] Will l The feature representation of the output of the convolutional branch units is concatenated to obtain t The final multi-scale feature representation Z t , whose expression is as follows:

[0056]

[0057] Constructing time series features , using the fully connected layer The time series characteristics are used for forecasting and processing to obtain the wind power forecast value, and its expression is as follows:

[0058]

[0059] in, For the future The wind power forecast value set for each time step, Time step The predicted value of wind power.

[0060] In a second aspect, the present invention further proposes a wind farm group power prediction system based on multi-scale spatiotemporal correlation, which is applied to the wind farm group power prediction method based on multi-scale spatiotemporal correlation as described in any solution of the first aspect, comprising:

[0061] The data preprocessing module is used to obtain the operating data, environmental data and geographical location data of each wind turbine in the target wind farm, and perform preprocessing to generate a basic feature matrix;

[0062] A multi-scale relationship matrix construction module is used to construct a geographical location matrix to characterize the spatial correlation between wind turbines, a historical wind speed similarity correlation matrix to characterize the similarity of meteorological evolution, and a wake effect dynamic relationship matrix to characterize the dynamic wake effect;

[0063] The spatiotemporal feature fusion module is used to perform multi-scale spatiotemporal correlation modeling on the basic feature matrix based on the geographic location matrix and the historical wind speed similarity correlation matrix through the spatial feature fusion unit, and fuse it with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix;

[0064] A node embedding generation module, used to aggregate information of the in-degree neighbor nodes and out-degree neighbor nodes of the wind turbine according to the multi-scale dynamic relationship matrix, and generate a node embedding representation of each wind turbine;

[0065] The prediction module is used to embed the nodes of each wind turbine into the input trained multi-scale time convolutional neural network prediction model for prediction processing to obtain the wind power prediction value.

[0066] The beneficial effects of the present invention include at least:

[0067] The present invention constructs a geographical location matrix, a historical wind speed similarity association matrix and a wake effect dynamic relationship matrix to respectively characterize the spatial distribution characteristics of wind farm groups, meteorological evolution correlation and dynamic wake interaction, and uses a graph convolutional neural network and a gated fusion unit to perform feature fusion on the first two types of matrices, and then superimposes them with the wake effect dynamic relationship matrix to form a multi-scale dynamic relationship matrix, which can explicitly model the spatiotemporal coupling mechanism of long-period geographical correlation, medium-period meteorological similarity and short-period dynamic wake influence between wind turbines; further, by distinguishing the weighted aggregation of in-degree neighbor nodes and out-degree neighbor nodes, the information transmission direction between wind turbines (such as the difference in the dynamic effect of the wake of the upwind wind turbine on the downwind wind turbine) can be accurately captured to generate a node embedding representation with clear physical meaning; finally, the multi-scale time convolutional neural network prediction model effectively takes into account the second-level turbulent fluctuations and hour-level trend evolution laws by extracting time series features at different time resolutions in parallel. Through the hierarchical modeling and fusion of multi-scale spatiotemporal dynamic relationships, the problem of incomplete modeling of complex spatiotemporal correlations between wind turbines by traditional methods has been solved, and the power prediction accuracy has been significantly improved. At the same time, due to the efficient use of the model's characteristics at different spatiotemporal scales, the consumption of redundant computing resources has been reduced, and the adaptability of the prediction system to complex scenarios has been enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1A schematic flow chart of a method for predicting wind farm group power based on multi-scale spatiotemporal correlation provided by an embodiment of the present invention.

[0069] Figure 2 (a) is a comparison diagram between the wind power prediction and actual situation of a wind turbine in area 1 provided by an embodiment of the present invention.

[0070] Figure 2 (b) is a comparison diagram between the wind power prediction and actual situation of a wind turbine in area 2 provided by an embodiment of the present invention.

[0071] Figure 2 (c) is a comparison chart between the wind power prediction and actual situation of a wind turbine in area 3 provided by an embodiment of the present invention.

[0072] Figure 3 This is an architecture diagram of a wind farm fleet power prediction system based on multi-scale spatiotemporal correlation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0073] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred technical solutions. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred technical solutions are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0074] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0075] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0076] Example 1

[0077] This embodiment proposes a wind farm group power prediction method based on multi-scale spatiotemporal correlation, such as Figure 1 As shown, Figure 1A schematic flow chart of a wind farm group power prediction method based on multi-scale spatiotemporal correlation provided by an embodiment of the present invention, the method comprising the following steps:

[0078] S1: Obtain the operating data, environmental data and geographic location data of each wind turbine in the target wind farm, and perform preprocessing to generate a basic feature matrix.

[0079] S2: Construct a geographical location matrix to characterize the spatial correlation between wind turbines, a historical wind speed similarity correlation matrix to characterize the similarity of meteorological evolution, and a wake effect dynamic relationship matrix to characterize the dynamic wake effect.

[0080] In this embodiment, three relationship matrices are used to capture the long-term, medium-term and short-term correlation characteristics between different wind turbines in a wind farm.

[0081] S3: Based on the geographic location matrix and the historical wind speed similarity correlation matrix, the basic feature matrix is ​​modeled with multi-scale spatiotemporal correlation through the spatial feature fusion unit, and is fused with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix.

[0082] In this embodiment, the multi-scale dynamic relationship matrix is ​​used to capture the spatiotemporal information of the wind turbine in various association modes.

[0083] S4: According to the multi-scale dynamic relationship matrix, information is aggregated on the in-degree neighbor nodes and out-degree neighbor nodes of the wind turbine respectively to generate a node embedding representation of each wind turbine.

[0084] S5: The node of each wind turbine is embedded into the trained multi-scale time convolutional neural network prediction model for prediction processing to obtain the wind power prediction value.

[0085] It can be understood that by constructing the geographical location matrix, the historical wind speed similarity correlation matrix and the wake effect dynamic relationship matrix, the spatial distribution characteristics of the wind farm fleet, the meteorological evolution correlation and the dynamic wake interaction are respectively characterized, and the graph convolutional neural network and the gated fusion unit are used to fuse the features of the first two types of matrices, and then superimposed with the wake effect dynamic relationship matrix to form a multi-scale dynamic relationship matrix, which can explicitly model the spatiotemporal coupling mechanism of long-period geographical correlation, medium-period meteorological similarity and short-period dynamic wake influence between wind turbines; further, by distinguishing the weighted aggregation of in-degree neighbor nodes and out-degree neighbor nodes, the information transmission direction between wind turbines can be accurately captured (such as the difference in the dynamic effect of the wake of the upwind wind turbine on the downwind wind turbine), and a node embedding representation with clear physical meaning can be generated; finally, the multi-scale time convolutional neural network prediction model effectively takes into account the second-level turbulent fluctuations and hour-level trend evolution laws by extracting time series features at different time resolutions in parallel. Through the hierarchical modeling and fusion of multi-scale spatiotemporal dynamic relationships, the problem of incomplete modeling of complex spatiotemporal correlations between wind turbines by traditional methods has been solved, and the power prediction accuracy has been significantly improved. At the same time, due to the efficient use of the model's characteristics at different spatiotemporal scales, the consumption of redundant computing resources has been reduced, and the adaptability of the prediction system to complex scenarios has been enhanced.

[0086] Example 2

[0087] This embodiment makes improvements on the wind farm fleet power prediction method based on multi-scale spatiotemporal correlation proposed in Embodiment 1.

[0088] In this embodiment, the operation data, environmental data and geographical location data of each wind turbine are cleaned, normalized and feature extracted in sequence to generate a basic feature matrix;

[0089] The expression of the basic feature matrix is ​​as follows:

[0090]

[0091] in, t is the time step, For fans i Real-time power output, For fans i The rotation speed, For fans i The blade angle, is the wind speed, For wind direction, For temperature, For humidity, For fans i The longitude coordinates of For fans i For the wind farm N Typhoon, overall characteristic matrixN × F The matrix of F is the number of all the above characteristics of a single fan.

[0092] In the specific implementation process, outlier detection is performed on the real-time power, speed and other operating data of the wind turbine (such as the 3σ principle to eliminate outliers), and sliding window mean filtering is used to reduce noise on environmental data such as wind speed, temperature and humidity. The geographical location coordinates are converted into relative coordinates with the center of the wind farm as the origin. Each feature is mapped to the [0,1] interval through Min-Max normalization, and finally a basic feature matrix containing 9-dimensional features is generated.

[0093] In this embodiment, constructing a geographical location matrix includes:

[0094] Based on the longitude and latitude coordinates of the wind turbines, the Euclidean distance between the wind turbines is calculated, and the distance map is generated through threshold screening to obtain the geographic location matrix , whose expression is as follows:

[0095]

[0096] in, For fans i The longitude coordinates of For fans i The latitude coordinate of For fans j The longitude coordinates of For fans j The latitude coordinate of is the set threshold.

[0097] In this embodiment, constructing a historical wind speed similarity association matrix includes:

[0098] Based on the historical wind speed sequence of the wind turbines, the wind speed time series cumulative alignment distance between the wind turbines is calculated and normalized to obtain the historical wind speed similarity association matrix. , whose expression is as follows:

[0099]

[0100] in, D ( i,j ) is the fan calculated by the dynamic time warping algorithm i and fan j The cumulative alignment distance of wind speed time series is A preset parameter that controls the sensitivity of distance to similarity.

[0101] In this embodiment, the wake effect dynamic relationship matrix is ​​constructed, including:

[0102] Based on the coordinate position, wind direction and wind speed information of the wind turbine, the Jensen wake model is used to generate the dynamic relationship matrix of the wake effect. , whose expression is as follows:

[0103]

[0104] in, α is the adjustment factor used to determine the impact strength of the upstream fan, Indicates fan i and fan j The wind direction angle between k represents the wake expansion coefficient, r i Indicates fan i The blade diameter, Indicates fan i and fan j The distance between is the axial flow interference factor of the upstream fan.

[0105] It can be understood that the dynamic relationship matrix determines whether the adjacent wind turbine is located in the wake area of ​​the upstream wind turbine based on the coordinate position, wind direction and wind speed information of each wind turbine. The Jensen wake model is used to calculate the effect of wake attenuation on wind speed to determine the wake transfer relationship between adjacent wind turbines, thereby generating a wake effect mapping matrix between each wind turbine.

[0106] In this embodiment, there are two time series and , the distance between each pair of points is expressed as Euclidean distance , the dynamic time warping algorithm is used to calculate the cumulative alignment distance of wind speed time series between wind turbines, and its expression is as follows:

[0107]

[0108] in, Indicates the first i Line j The value of the column represents the minimum cumulative distance from the starting point of the sequence to the current point. For fans i Time Series At the moment i The value of For fans j Time Series At the moment j The value of N and M The time series X and time series Y Length, For time series X Middle i Points and time series Y Middle j The Euclidean distance between points.

[0109] In this embodiment, based on the geographic location matrix and the historical wind speed similarity association matrix, the basic feature matrix is ​​modeled with multi-scale spatiotemporal association through the spatial feature fusion unit, and is fused with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix, including:

[0110] Based on the geographic location matrix Similarity correlation matrix with historical wind speed , using graph convolutional neural network to transform the feature matrix of basic feature data K Perform graph convolution to obtain intermediate feature representation H , whose expression is as follows:

[0111]

[0112] in, and are the normalized geographical location matrix and the historical wind speed similarity association matrix, I is the identity matrix, D is the degree matrix, W P and W S are the weight matrices of similarity between geographical location and historical wind speed, σ is the activation function;

[0113] The intermediate features are represented H After being input into the gated fusion unit for processing and linear transformation, the dynamic relationship matrix with the wake effect Weighted summation to obtain the multi-scale dynamic relationship matrix , whose expression is as follows:

[0114]

[0115] in, is the output of the update gate, U z is the weight parameter, b z is the bias parameter, r is the output of the overlap gate, U r is the weight parameter, b r is the bias parameter, represents a hyperbolic tangent transformation, It is the intermediate feature representation after being processed by the reset gate and hyperbolic tangent transform.

[0116] In this embodiment, a node embedding representation of each wind turbine is generated by constructing a dynamic direction-aware graph convolutional neural network. In order to distinguish the influence of outgoing edges and incoming edges, this embodiment divides the attention heads into two groups: one group is used to aggregate the out-degree neighbor information, and the other group is used to aggregate the in-degree neighbor information. Such multi-head attention aggregation results can combine the in-degree and out-degree neighbor information, and finally assign independent learning parameters to each node to obtain a comprehensive node embedding. Taking into account the flow characteristics of wind in wind farms, this embodiment defines the out-degree and in-degree relationship of wind turbines according to wind direction, where the out-degree neighbor nodes refer to other wind turbine nodes that the wind blows from the wind turbine under the current wind direction conditions; and the in-degree neighbor nodes refer to other wind turbine nodes that the wind comes from under the current wind direction conditions.

[0117] Furthermore, according to the multi-scale dynamic relationship matrix, information aggregation is performed on the in-degree neighbor nodes and out-degree neighbor nodes of the wind turbine respectively to generate a node embedding representation of each wind turbine, including:

[0118] According to the multi-scale dynamic relationship matrix, the attention weights are calculated for the out-degree neighbor nodes and in-degree neighbor nodes of the wind turbine respectively, and the expressions are as follows:

[0119]

[0120] in, and are the attention weights of the out-degree neighbor nodes and the in-degree neighbor nodes respectively, and The attention vectors that can be learned are: and are the linear transformation matrices, and Fan i and fan j The feature matrix of the basic feature data, and are independent learning parameters respectively;

[0121] The attention weights of out-degree neighbor nodes and in-degree neighbor nodes are normalized, and the expressions are as follows:

[0122]

[0123] in, and are the normalized attention weights of the out-degree neighbor nodes and the in-degree neighbor nodes, and Fani The set of out-degree neighbor nodes and in-degree neighbor nodes;

[0124] Use the normalized attention weights to perform weighted aggregation on the information of out-degree neighbor nodes and in-degree neighbor nodes, and obtain the information aggregation representation of the out-degree neighbor nodes and the information aggregation representation of in-degree neighbor nodes , whose expression is as follows:

[0125]

[0126] in, σ is the activation function, Indicates fan i For outgoing neighbor nodes j The attention weight, Indicates fan i For outgoing neighbor nodes j The attention weight of

[0127] Information aggregation representation and information aggregation representation Perform summation to obtain the final node embedding representation , whose expression is as follows:

[0128] .

[0129] In this embodiment, the node of each wind turbine is embedded into the input trained multi-scale time convolutional neural network prediction model for prediction processing to obtain the wind power prediction value, including:

[0130] Using node embedding representation Constructing the graph information matrix ; Each graph information matrix consists of the embedding vectors of multiple nodes obtained in requirement 5, where N is the number of nodes, i.e. the number of wind turbines.

[0131] Using the Graph Information Matrix Constructing the input sequence , where the input sequence The graph information matrix of T consecutive time steps constitute, N is the number of fans, Represents the feature dimension of each wind turbine after passing through the graph convolutional neural network;

[0132] In the multi-scale temporal convolutional neural network prediction model, several convolutional branch units with different expansion rates are used to transform the input feature sequence. Perform convolution operations to extract feature representations at different time scales. The expressions are as follows:

[0133]

[0134] in, Indicates that at time step t At that time, l The feature representation of the output of the convolutional branch unit, Indicates l The weight parameters of the convolutional branch units, Represents a graph with expansion rate d l The one-dimensional convolution operation, Represents the time step To time step t The input feature sequence is For the l Each convolution branch unit initially uses the same size of convolution kernel to process the input data, and then applies multiple different dilated convolutions to operate. The effective receptive field size corresponding to different dilation rates has different time spans in the data dimension, which can extract high-level features contained in non-adjacent time at different time intervals in the time series.

[0135] Will l The feature representation of the output of the convolutional branch units is concatenated to obtain t The final multi-scale feature representation Z t , whose expression is as follows:

[0136]

[0137] Constructing time series features , using the fully connected layer The time series characteristics are used for forecasting and processing to obtain the wind power forecast value, and its expression is as follows:

[0138]

[0139] in, For the future The wind power forecast value set for each time step, Time step The predicted value of wind power.

[0140] Example 3

[0141] To verify the effectiveness of the present invention, this embodiment first obtains the operating data of 134 wind turbines in a wind farm from 00:00 on January 1, 2022 to 23:00 on July 29, 2022, including wind power-related data such as power output, wind speed, wind direction and precipitation. Subsequently, combined with the geographical location, meteorological characteristics and operating data of the wind turbines, a geographical location matrix characterizing the spatial correlation between wind turbines, a historical wind speed similarity correlation matrix characterizing the similarity of meteorological evolution, and a wake effect dynamic relationship matrix characterizing the dynamic wake effect are constructed to comprehensively capture the spatial and temporal correlation characteristics between wind turbines. On this basis, the dynamic direction perception graph attention module and the multi-scale time convolution module are used to extract the spatiotemporal correlation characteristics of the wind farm, and finally a multi-scale dynamic spatiotemporal matrix is ​​generated. In the experiment, the dynamic direction perception graph convolutional neural network is set to 2 attention heads, and the temporal convolutional network uses convolution kernels with expansion rates of 1, 3 and 5 for feature extraction. The wind power is predicted by the method of the present invention, and the following is obtained: Figure 2 The prediction results shown verify the superior performance of this method in wind power prediction.

[0142] like Figure 2 As shown in (a), Figure 2 (a) is a comparison chart of the wind power prediction and actual situation of a wind turbine in region 1 provided by an embodiment of the present invention, with the horizontal axis being time (h) and the vertical axis being wind power (kW). Among them, the solid line represents the actual value of wind power, reflecting the actual power generation level of the wind turbine at different times, and its fluctuation reflects the power change under the influence of complex factors such as wind speed and wind direction in actual operation. The dotted line is the wind power prediction output value, which is generated by the method of the present invention. It can be seen from the figure that in the time period of 0-200h, the predicted value is closely matched with the actual value, indicating that the method of the present invention has an accurate grasp of the wind turbine power change trend during this period; between 200-400h, although there are some fluctuations, the overall trend is still consistent, reflecting the adaptability of the method under complex working conditions; in the range of 400-1000h, the predicted value can better follow the large fluctuations of the actual value, indicating that the method can effectively capture the power change characteristics of the wind turbine on a long time scale, with high prediction accuracy, and can provide a reliable basis for wind power management in region 1.

[0143] like Figure 2 (b) Figure 2(b) is a comparison chart between the wind power prediction and actual situation of a wind turbine in region 2 provided by an embodiment of the present invention. The horizontal axis is time (h), and the vertical axis is wind power (kW). The solid line is the actual value of the wind power of the wind turbine, which records the real-time power generation of the wind turbine during operation. The dotted line is the predicted output value. In 0-200h, the deviation between the predicted value and the actual value is small, showing the good initial prediction performance of the method; during 200-600h, in the face of multiple fluctuations in actual power, the predicted value can respond in time and maintain a similar change trend, indicating that the method of the present invention can effectively cope with the dynamic changes of wind turbine power; 600-1000h, although there are some small-scale differences, the overall fitting effect is good, which proves that the method has high reliability in predicting the wind power of the wind turbine in region 2, and can provide effective reference for decisions such as wind power scheduling in region 2.

[0144] like Figure 2 (c) Figure 2 (c) is a comparison chart between the wind power prediction and actual situation of a wind turbine in region 3 provided by an embodiment of the present invention. The horizontal axis is time (h), and the vertical axis is wind power (kW). The solid line is the actual value of wind power, which represents the power output of the wind turbine in actual operation. The dotted line is the predicted output value. In 0-200h, the predicted value quickly follows the change of the actual value, showing the rapid response ability of the method; in 200-600h, the actual power fluctuates to a large extent. Although the predicted value has a certain deviation, it can still roughly reflect its change trend, reflecting the ability of the method to deal with complex power fluctuations; in 600-1000h, the predicted value and the actual value are consistent in the overall trend, which further verifies the effectiveness of the method of the present invention in predicting wind power of wind turbines in region 3, and can more accurately predict wind turbine power, which is conducive to the stable operation and planning of the wind power system in region 3.

[0145] Example 4

[0146] like Figure 3 As shown, this embodiment proposes a wind farm cluster power prediction system based on multi-scale spatiotemporal correlation, which is applied to the wind farm cluster power prediction method based on multi-scale spatiotemporal correlation as described in the above embodiment, including: a data preprocessing module 100, a multi-scale relationship matrix construction module 200, a spatiotemporal feature fusion module 300, a node embedding generation module 400 and a prediction module 500.

[0147] Among them, the data preprocessing module 100 is used to obtain the operating data, environmental data and geographical location data of each wind turbine in the target wind farm, and perform preprocessing to generate a basic feature matrix; the multi-scale relationship matrix construction module 200 is used to construct a geographical location matrix that characterizes the spatial correlation between wind turbines, a historical wind speed similarity correlation matrix that characterizes the similarity of meteorological evolution, and a wake effect dynamic relationship matrix that characterizes the dynamic wake effect; the spatiotemporal feature fusion module 300 is used to perform multi-scale spatiotemporal correlation modeling on the basic feature matrix through a spatial feature fusion unit based on the geographical location matrix and the historical wind speed similarity correlation matrix, and fuse it with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix; the node embedding generation module 400 is used to aggregate the information of the in-degree neighbor nodes and out-degree neighbor nodes of the wind turbine according to the multi-scale dynamic relationship matrix, and generate a node embedding representation of each wind turbine; the prediction module 500 is used to input the node embedding representation of each wind turbine into the trained multi-scale time convolutional neural network prediction model for prediction processing to obtain a wind power prediction value.

[0148] It should be noted that the aforementioned explanation of the embodiment of the wind farm group power prediction method based on multi-scale spatiotemporal correlation is also applicable to the wind farm group power prediction system based on multi-scale spatiotemporal correlation of this embodiment, and will not be repeated here.

[0149] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0150] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0151] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

[0152] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiment, the N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0153] A person skilled in the art may understand that all or part of the steps in the above-mentioned embodiment method may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0154] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A wind farm power prediction method based on multi-scale spatiotemporal correlation, characterized in that: include: Obtain the operating data, environmental data and geographic location data of each wind turbine in the target wind farm, perform preprocessing, and generate a basic feature matrix; Construct a geographical location matrix to characterize the spatial correlation between wind turbines, a historical wind speed similarity correlation matrix to characterize the similarity of meteorological evolution, and a wake effect dynamic relationship matrix to characterize the dynamic wake effect; Based on the geographic location matrix and the historical wind speed similarity correlation matrix, the basic feature matrix is ​​modeled with multi-scale spatiotemporal correlation through the spatial feature fusion unit, and is fused with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix, including: Based on the geographic location matrix Similarity correlation matrix with historical wind speed , using graph convolutional neural network to transform the feature matrix of basic feature data K Perform graph convolution to obtain intermediate feature representation H , whose expression is as follows: in, and are the normalized geographical location matrix and the historical wind speed similarity association matrix, I is the identity matrix, D is the degree matrix, W P and W S are the weight matrices of similarity between geographical location and historical wind speed, σ is the activation function; The intermediate features are represented H After being input into the gated fusion unit for processing and linear transformation, the dynamic relationship matrix with the wake effect Weighted summation to obtain the multi-scale dynamic relationship matrix , whose expression is as follows: in, is the output of the update gate, U z is the weight parameter, b z is the bias parameter, r is the output of the overlap gate, U r is the weight parameter, b r is the bias parameter, represents a hyperbolic tangent transformation, It is the intermediate feature representation after being processed by the reset gate and hyperbolic tangent transform; According to the multi-scale dynamic relationship matrix, respectively aggregate the information of the in-degree neighbor nodes and the out-degree neighbor nodes of the wind turbine to generate a node embedding representation of each wind turbine; The node embedding of each wind turbine represents the input of the trained multi-scale time convolutional neural network prediction model for prediction processing to obtain the wind power prediction value.

2. The wind farm group power prediction method based on multi-scale spatiotemporal correlation according to claim 1 is characterized in that: The operation data, environmental data and geographical location data of each wind turbine are cleaned, normalized and feature extracted in turn to generate a basic feature matrix; The expression of the basic feature matrix is ​​as follows: in, t is the time step, For fans i Real-time power output, For fans i The rotation speed, For fans i The blade angle, is the wind speed, For wind direction, For temperature, For humidity, For fans i The longitude coordinates of For fans i The latitude coordinate of .

3. The wind farm group power prediction method based on multi-scale spatiotemporal correlation according to claim 1 is characterized in that: Construct a geographic location matrix, including: Based on the longitude and latitude coordinates of the wind turbines, the Euclidean distance between the wind turbines is calculated, and the distance map is generated through threshold screening to obtain the geographic location matrix , whose expression is as follows: in, For fans i The longitude coordinates of For fans i The latitude coordinate of For fans j The longitude coordinates of For fans j The latitude coordinate of is the set threshold.

4. The wind farm group power prediction method based on multi-scale spatiotemporal correlation according to claim 3 is characterized in that: Construct the historical wind speed similarity correlation matrix, including: Based on the historical wind speed sequence of the wind turbines, the wind speed time series cumulative alignment distance between the wind turbines is calculated and normalized to obtain the historical wind speed similarity association matrix. , whose expression is as follows: in, D ( i,j ) is the fan calculated by the dynamic time warping algorithm i and fan j The cumulative alignment distance of wind speed time series is A preset parameter that controls the sensitivity of distance to similarity.

5. The wind farm group power prediction method based on multi-scale spatiotemporal correlation according to claim 4 is characterized in that: Construct the wake effect dynamic relationship matrix, including: Based on the coordinate position, wind direction and wind speed information of the wind turbine, the Jensen wake model is used to generate the dynamic relationship matrix of the wake effect. , whose expression is as follows: in, α is the adjustment factor used to determine the impact strength of the upstream fan, Indicates fan i and fan j The wind direction angle between k represents the wake expansion coefficient, r i Indicates fan i The blade diameter, Indicates fan i and fan j The distance between is the axial flow interference factor of the upstream fan.

6. The wind farm group power prediction method based on multi-scale spatiotemporal correlation according to claim 5 is characterized in that: The dynamic time warping algorithm is used to calculate the cumulative alignment distance of wind speed time series between wind turbines. The expression is as follows: in, Indicates the first i Line j The value of the column represents the minimum cumulative distance from the starting point of the sequence to the current point. For fans i Time Series At the moment i The value of For fans j Time Series At the moment j The value of N and M The time series X and time series Y Length, For time series X Middle i Points and time series Y Middle j The Euclidean distance between points.

7. The wind farm group power prediction method based on multi-scale spatiotemporal correlation according to claim 5 is characterized in that: According to the multi-scale dynamic relationship matrix, information aggregation is performed on the in-degree neighbor nodes and out-degree neighbor nodes of the wind turbine respectively to generate a node embedding representation of each wind turbine, including: According to the multi-scale dynamic relationship matrix, the attention weights are calculated for the out-degree neighbor nodes and in-degree neighbor nodes of the wind turbine respectively, and the expressions are as follows: in, and are the attention weights of the out-degree neighbor nodes and the in-degree neighbor nodes respectively, and The attention vectors that can be learned are: and are the linear transformation matrices, and Fan i and fan j The feature matrix of the basic feature data, and are independent learning parameters respectively; The attention weights of out-degree neighbor nodes and in-degree neighbor nodes are normalized, and the expressions are as follows: in, and are the normalized attention weights of the out-degree neighbor nodes and the in-degree neighbor nodes, and Fan i The set of out-degree neighbor nodes and in-degree neighbor nodes; Use the normalized attention weights to perform weighted aggregation on the information of out-degree neighbor nodes and in-degree neighbor nodes, and obtain the information aggregation representation of the out-degree neighbor nodes and the information aggregation representation of in-degree neighbor nodes , whose expression is as follows: in, σ is the activation function, Indicates fan i For outgoing neighbor nodes j The attention weight, Indicates fan i For outgoing neighbor nodes j The attention weight of Information aggregation representation and information aggregation representation Perform summation to obtain the final node embedding representation , whose expression is as follows: 。 8. The wind farm group power prediction method based on multi-scale spatiotemporal correlation according to claim 7 is characterized in that: The multi-scale temporal convolutional neural network prediction model includes an input layer, a hidden layer and an output layer; The node embedding of each wind turbine represents the input of the trained multi-scale time convolutional neural network prediction model for prediction processing to obtain the wind power prediction value, including: In the input layer, the nodes are embedded into the representation Constructing the graph information matrix , and using the graph information matrix Constructing input feature sequence , where the input feature sequence The graph information matrix of T consecutive time steps constitute, N is the number of fans, Represents the feature dimension of each wind turbine after passing through the graph convolutional neural network; In the hidden layer, several convolutional branch units with different expansion rates are used to transform the input feature sequence. Perform convolution operations to extract feature representations at different time scales. The expressions are as follows: in, Indicates that at time step t At that time, l The feature representation of the output of the convolutional branch unit, Indicates l The weight parameters of the convolutional branch units, Represents a graph with expansion rate d l The one-dimensional convolution operation, Represents the time step To time step t The input feature sequence is For the l The bias parameters of the convolutional branch units; Will l The feature representation of the output of the convolutional branch units is concatenated to obtain t The final multi-scale feature representation Z t , whose expression is as follows: Constructing time series features in the output layer , using the fully connected layer The time series characteristics are used for forecasting and processing to obtain the wind power forecast value, and its expression is as follows: in, For the future The wind power forecast value set for each time step, Time step The predicted value of wind power.

9. A wind farm group power prediction system based on multi-scale spatiotemporal correlation, applied to the wind farm group power prediction method based on multi-scale spatiotemporal correlation as claimed in any one of claims 1 to 8, characterized in that: include: The data preprocessing module is used to obtain the operating data, environmental data and geographical location data of each wind turbine in the target wind farm, and perform preprocessing to generate a basic feature matrix; A multi-scale relationship matrix construction module is used to construct a geographical location matrix to characterize the spatial correlation between wind turbines, a historical wind speed similarity correlation matrix to characterize the similarity of meteorological evolution, and a wake effect dynamic relationship matrix to characterize the dynamic wake effect; The spatiotemporal feature fusion module is used to perform multi-scale spatiotemporal correlation modeling on the basic feature matrix based on the geographic location matrix and the historical wind speed similarity correlation matrix through the spatial feature fusion unit, and fuse it with the wake effect dynamic relationship matrix to generate a multi-scale dynamic relationship matrix; A node embedding generation module, used to aggregate information of the in-degree neighbor nodes and out-degree neighbor nodes of the wind turbine according to the multi-scale dynamic relationship matrix, and generate a node embedding representation of each wind turbine; The prediction module is used to embed the nodes of each wind turbine into the input trained multi-scale time convolutional neural network prediction model for prediction processing to obtain the wind power prediction value.

Images