Wind power cluster power day-ahead prediction method and device, equipment and storage medium

Abstract

The invention provides a day-ahead prediction method and device for wind power cluster power, equipment and a storage medium. Relates to the technical field of wind power generation prediction. The method comprises the steps that clustering is completed based on historical multivariate information of all stations in a wind power cluster, cluster virtual nodes are set, and the visual angle of global information is expanded for cluster modeling; and information propagation characteristics between nodes are explored based on a cluster prevailing wind direction. Constructing an adjacent matrix of the dynamic graph, and quantifying the dynamic evolution process of the cluster topology; and finally, mining cluster dynamic information by using a frequency domain channel attention mechanism and STGCN, and completing short-term power prediction of the wind power cluster. The method is a prediction method which considers multivariate information space-time diversity and is suitable for the day-ahead wind power cluster power, and is simple in calculation, high in prediction performance, clear in physical significance, effective in prediction result and high in practicability.

Description

Technical Field

[0001] This application relates to the field of wind power generation prediction technology, and in particular to a method, apparatus, equipment and storage medium for predicting the day-ahead power of a wind power cluster. Background Technology

[0002] With the continuous growth of global energy demand and increasing global attention to environmental protection, new energy sources have become one of the important development directions in the world's energy sector. As wind power installed capacity continues to increase, the impact of its volatility and uncertainty is becoming increasingly prominent, posing a series of challenges to grid operation. Under these circumstances, accurate and reliable wind power forecasting has become one of the key technologies for improving the absorption capacity of new energy sources and grid stability. Short-term wind power forecasting is an indispensable part of this. This paper focuses on wind power forecasting on the day-ahead timescale, which refers to forecasts within 24-48 hours, and applies it to optimizing daily power generation plans and cold standby, as well as adjusting maintenance plans.

[0003] Currently, numerous studies have been conducted on short-term power prediction for wind power clusters, resulting in a series of mainstream prediction methods. These include the overall method, the cumulative method, and the cluster partitioning method. However, clustering methods that only consider a single characteristic lose sight of the comprehensive spatiotemporal information of the clusters, leading to insufficient data mining. Therefore, current cluster partitioning methods are inadequate in comprehensively considering spatiotemporal correlations, and further research is needed in clustering methods and modeling approaches. Summary of the Invention

[0004] This application provides a method, apparatus, equipment, and storage medium for predicting the power of wind power clusters before the day, in order to overcome the shortcomings of existing prediction methods and provide a short-term power prediction scheme for wind power clusters that is physically clear, scientifically sound and effective, reliable and stable, and capable of utilizing the spatiotemporal heterogeneity of multi-source information.

[0005] In a first aspect, this application provides a method for predicting the power day-ahead of a wind power cluster, including:

[0006] The subgraph partitioning target is determined based on the tiling method, and the correlation matrix constructed in a progressive manner is used as the affinity matrix to partition the wind power cluster, resulting in multiple sub-clusters.

[0007] The latitude and longitude of the stations within the sub-cluster are projected onto a two-dimensional plane, and virtual information nodes are established through the computational geometric center;

[0008] Based on the sub-clusters obtained from the division, an NWP sliding time window is established, and the prevailing wind direction of the sub-clusters is identified by combining the stable state and abrupt changes of the wind direction data within the window.

[0009] Based on the prevailing wind direction of the sub-cluster, the directed edge connection relationship between the stations is determined. The adjacency matrix of the dynamic weighted directed graph is constructed by combining the distance matrix and the wind speed correlation, and virtual information nodes are integrated to construct the dynamic weighted directed graph.

[0010] A day-ahead power prediction model is constructed, which takes the dynamic weighted directed graph as input and outputs the day-ahead power prediction results of the wind power cluster.

[0011] In one possible design, the subgraph partitioning objective is determined based on the tiling method, including:

[0012] The goal of subgraph partitioning is to minimize the sum of the edge weights of all subgraphs; wherein the formula for calculating the sum of the edge weights of the subgraphs is:

[0013]

[0014] In the formula, Ncut(A1,A2,...,A) K A represents the sum of edge weights across all subgraphs, where k is the number of subgraphs, and A i and A j Let A and B represent the sets of nodes within the i-th and j-th subgraphs, respectively, and satisfy any A i ∩A j =φ, A1∪A2∪…∪A k =V, where V is the set of all nodes. For A i The supplement, For A i The sum of the edge weights of its complement and its complement, vol(A) i ) is A i The sum of the weights of all edges in the equation.

[0015] In one possible design, the affinity matrix is ​​constructed as follows:

[0016] Based on the geographic location information, the distance weight matrix is ​​calculated using the following formula:

[0017]

[0018] In the formula, and Let λ be the latitude of the i-th and j-th stations. i and λ j Let be the longitudes of the i-th and j-th stations, h be the semi-versus of the central angle between the two points, r be the radius of the Earth, and D be the longitude of the station. ij H is the great circle distance between the two stations; H is the variance of all distances between stations; G is the variance of all distances between stations. DW (i,j) is the distance weight matrix G DWFor each element at each position in the array, exp is an exponential function with the natural constant as its base;

[0019] Based on the aforementioned distance weight matrix, a distance matrix G ignoring long-distance station connections is constructed using the following formula. DE :

[0020]

[0021] In the formula, G DE (i,j) is the distance matrix G DE The elements at each position within the range, where ε is the set correlation threshold;

[0022] Based on the distance matrix G DE The affinity matrix S is constructed using the following formula:

[0023] S = a * G DW +b*G DE W DC +g*G DE W PC (6)

[0024] In the formula, a, b, and g are the fusion weights, and W DC For the wind direction correlation matrix, W PC This is the power correlation matrix.

[0025] In one possible design, the latitude and longitude of the stations within the sub-cluster are projected onto a two-dimensional plane, and virtual information nodes are established through computational geometric centers, including:

[0026] The latitude and longitude coordinates of the stations within the sub-cluster are projected onto a two-dimensional plane using the following formula:

[0027]

[0028] In the formula, g is the station number, λ g and Let x be the longitude and latitude of station g within the sub-cluster. g With y g Let g be the coordinates of the latitude and longitude of station g in the plane projection, and r be the radius of the Earth;

[0029] The location of the virtual information node is determined by the following formula:

[0030]

[0031] In the formula, N is the number of wind farms in a certain sub-cluster, (X n ,Y n ) represents the coordinates of the virtual information node in the nth cluster.

[0032] In one possible design, based on the divided sub-clusters, a sliding time window for NWP (Non-Wave Dynamics and Wind) is established. The prevailing wind direction of the sub-clusters is identified by combining the stable state and abrupt changes of wind direction data within the window, including:

[0033] Establish a sliding time window for NWP, and represent the wind direction window sequence as w = [w0, w1, ..., w K-1 ], where K is the time window length, w0 and w1 represent the wind direction data at the first and second sampling times, respectively, w K-1 The wind direction data is for the time to be predicted;

[0034] A mutation threshold is set. When the difference between the wind direction data at the time to be predicted and the wind direction at the previous time is less than or equal to the mutation threshold, the wind direction data within the time window is determined to be in a stable state. When the difference between the wind direction data at the time to be predicted and the wind direction at the previous time is greater than the mutation threshold, the wind direction data within the time window is determined to be in a mutation state.

[0035] Under steady-state conditions, the prevailing wind direction DIR of the sub-cluster is determined by the following formula:

[0036]

[0037] In the formula, w i Let w be the i-th element in the wind direction window; max[w] and min[w] are the maximum and minimum values ​​of the elements in the wind direction window, respectively.

[0038] In the event of a mutation, the prevailing wind direction (DIR) of the sub-cluster is determined using the following formula:

[0039]

[0040] In the formula, w q ∈{w q |min[w]≤w q ≤w K-1} and w p ∈{w p |w K-1 ≤w p ≤max[w]}, representing the first set and the second set respectively. The first set and the second set are used to find the elements between the latest wind direction and the sudden change in wind direction. M1 and M2 are the number of elements in the first set and the second set respectively. q and p are the indices of the elements in the first set and the second set respectively.

[0041] In one possible design, based on the prevailing wind direction of the sub-cluster, the directed edge connections between stations are determined. An adjacency matrix of a dynamic weighted directed graph is constructed by combining the distance matrix and wind speed correlation. Virtual information nodes are then integrated to construct the dynamic weighted directed graph, including:

[0042] Based on the prevailing wind direction of the sub-cluster, the evolution characteristics of the information propagation direction between stations are represented as a directed graph; wherein, the directed graph is traversed by the set of directed edges E. t The set of directed edges is characterized by the asymmetric wind direction matrix W. D Quantization; the asymmetric wind direction matrix W D The calculation formula is:

[0043]

[0044] In the formula, W D (i,j) represent the elements in the asymmetric wind direction matrix, e ab Let e ​​be the unit vector pointing from station a to station b. DIR This is the unit vector for the prevailing wind direction, where ||·|| represents the calculation of the modulus.

[0045] Based on the aforementioned asymmetric wind direction matrix W D The adjacency matrix of the dynamically weighted directed graph is constructed using the following formula:

[0046] M W =[G DE ⊙(W D +I n )]⊙(αM person +βM spearman (12)

[0047] In the formula, G DE To ignore the distance matrix of long-distance station connections, ⊙ represents the Hadamard product operation of the matrix, I n It is an identity matrix, where α and β are weighting coefficients, and M... person and M spearman A correlation matrix was established to calculate the Pearson and Spearman coefficients for the wind speed sequences at the stations within the window.

[0048] In one possible design, the power day-ahead prediction model includes a first network layer, a spatiotemporal information aggregation module, and a second network layer;

[0049] The input data of the first network layer is a wind speed sequence with a fixed time window L, where each column represents a feature channel and each feature channel is considered an independent vector. The first network layer is used to: perform a one-dimensional discrete cosine transform on each independent vector to convert the original time series data into a frequency domain vector; process the frequency domain vector through a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function to obtain a frequency domain attention vector; concatenate all the frequency domain attention vectors horizontally to form a frequency domain attention matrix, and perform a Hadamard product with the original input data to obtain the output features of the first network layer.

[0050] The spatiotemporal information aggregation module takes the output features of the first network layer and the adjacency matrix of the dynamically weighted directed graph as input, and uses one-dimensional dilated causal convolution to obtain temporal features; and based on the temporal features, obtains the spatiotemporally aggregated features using the following formula:

[0051]

[0052] In the formula, t is the time step; A t Given the input adjacency matrix, It is a dynamic adjacency matrix with autocorrelation; for The degree matrix; I n X is the identity matrix; (l) With X (l+1) These are the spatiotemporal aggregation features of the l-th and l+1-th layers, respectively; σ g W is the sigmoid activation function for the nonlinear model. (l) Here is the weight matrix for layer l;

[0053] The second network layer takes the high-dimensional features obtained after spatiotemporal aggregation as input and obtains the predicted power of each node using the following formula:

[0054] v freq-Ⅱi =σ(W 2-Ⅱ δ(W 1-Ⅱ F DCT (f flatten (H i )))) (twenty one)

[0055] Pred i =σ[W fci (v freq-Ⅱi ⊙f flatten (H i ))] (twenty two)

[0056] In the formula, H i f is a high-dimensional representation of the wind speed at the p-th station. flatten (·) represents the flattening and fully connected operations, which transform high-dimensional information back into 1-dimensional information; W 1-Ⅱ With W 2-Ⅱ For a fully connected layer, v freq- Ⅱi is the frequency domain attention vector in the frequency domain attention process, W fci Pred is an independent fully connected layer for each station. i F is the predicted power corresponding to each node. DCT (·) represents the one-dimensional discrete cosine transform operation.

[0057] Secondly, this application provides a wind power cluster power day-ahead prediction device, the device comprising:

[0058] The sub-cluster partitioning module is configured to determine the sub-graph partitioning target based on the tiling method, and to partition the wind power cluster using a progressively constructed correlation matrix as the affinity matrix to obtain multiple sub-clusters.

[0059] The virtual node construction module is configured to project the latitude and longitude of the stations within the sub-cluster onto a two-dimensional plane and establish virtual information nodes through the computational geometric center;

[0060] The prevailing wind direction identification module is configured to establish an NWP sliding time window based on the divided sub-clusters, and identify the prevailing wind direction of the sub-clusters by combining the stable state and sudden changes of the wind direction data within the window.

[0061] The directed graph construction module is configured to determine the directed edge connection relationship between the stations based on the prevailing wind direction of the sub-cluster, construct the adjacency matrix of the dynamic weighted directed graph by combining the distance matrix and wind speed correlation, and integrate virtual information nodes to construct the dynamic weighted directed graph.

[0062] The day-ahead power prediction module is configured to construct a day-ahead power prediction model, which takes the dynamic weighted directed graph as input and outputs the day-ahead power prediction results of the wind power cluster.

[0063] Thirdly, embodiments of this application provide an electronic device, including: at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the wind power cluster power day-ahead prediction method as described in the first aspect and various possible designs of the first aspect.

[0064] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the wind power cluster power day-ahead prediction method described in the first aspect and various possible designs of the first aspect.

[0065] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the wind power cluster power day-ahead prediction method as described in the first aspect and various possible designs of the first aspect.

[0066] The wind power cluster power day-ahead prediction method, apparatus, equipment, and storage medium provided in this application have at least the following beneficial effects:

[0067] This application addresses the problem of day-ahead wind power cluster prediction by proposing a wind power cluster day-ahead prediction model (FECAM-DSTGCN) based on frequency domain information gain and dynamic trend perception. First, it integrates multi-dimensional feature information from wind farms with proposed virtual cluster nodes to improve the spatiotemporal differentiation analysis of the cluster. Second, based on prevailing wind direction identification, wind direction information is rationally integrated into a dynamic weighted directed graph structure, comprehensively revealing the dynamic spatiotemporal correlation characteristics of the wind power cluster. Finally, the FECAM-DSTGCN model is designed to achieve in-depth analysis of cluster data at multiple levels—time, space, and frequency—improving the accuracy and reliability of day-ahead wind power prediction. Attached Figure Description

[0068] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0069] Figure 1 A flowchart of a wind power cluster power day-ahead prediction method provided in this application embodiment;

[0070] Figure 2 This is a schematic diagram of the wind farm sub-cluster and virtual node projection provided in the embodiments of this application;

[0071] Figure 3 The dynamic weighted directed graph structure based on wind direction information provided in the embodiments of this application;

[0072] Figure 4 The structural diagram of the FECAM-DSTGCN model provided in the embodiments of this application;

[0073] Figure 5 A flowchart illustrating the overall prediction process of a wind power cluster day-ahead prediction method provided in this application embodiment;

[0074] Figure 6 This is a comparison chart of prediction results from different clustering methods provided in the embodiments of this application;

[0075] Figure 7 A comparison chart of results from different day-ahead power prediction methods provided in the embodiments of this application;

[0076] Figure 8 A structural diagram of the wind power cluster power day-ahead prediction device provided in the embodiments of this application.

[0077] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0078] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0079] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0080] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0081] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0082] This application provides a method for day-ahead power prediction of wind power clusters. First, the method performs clustering based on historical multivariate information from each wind farm within the cluster and establishes virtual nodes to expand the global information perspective for cluster modeling. Second, it explores the information propagation characteristics between nodes based on the prevailing wind direction within the cluster. An adjacency matrix of a dynamic graph is constructed to quantify the dynamic evolution of the cluster topology. Finally, a frequency domain channel attention mechanism and STGCN are used to mine dynamic information from the cluster to complete short-term power prediction of the wind power cluster. Figure 5As shown, this method, based on frequency domain information gain and dynamic trend perception, performs day-ahead power prediction for wind power clusters. It consists of three parts: cluster partitioning and virtual nodes, wind direction identification and dynamic graph construction, and short-term power modeling and prediction for wind power clusters. Specifically: Cluster partitioning and virtual node construction utilizes the site's geographic information to construct a distance matrix, combines NWP (Non-Wide Potential Width) and power data to construct a correlation matrix, and then uses spectral clustering and virtual node construction to partition clusters A, B, and C and assign virtual nodes. Wind direction identification and dynamic graph construction involves identifying the prevailing wind direction for each sub-cluster based on the NWP sliding window, combining distance and wind speed correlation to construct a directed edge weight matrix, forming dynamic graphs of clusters A, B, and C, including wind direction identification examples (distribution of different wind directions within a time window). Short-term power modeling and prediction for wind power clusters involves inputting the dynamic adjacency matrices A, B, and C of each cluster into the spatiotemporal modeling module, processing them through the FECAM-DSTGCN model, incorporating prediction corrections, and outputting the power prediction results.

[0083] In summary, the day-ahead power forecasting method for this wind power cluster includes the following steps:

[0084] 1) Cluster division is performed based on the progressive fusion of multi-source information, and virtual information nodes of sub-clusters are constructed; 2) Cluster wind direction information is integrated into the construction of a dynamic weighted directed graph to reveal the spatiotemporal characteristics of wind power clusters from multiple perspectives; 3) Frequency domain information is integrated into an improved dynamic spatiotemporal graph convolution model, so that the model can make full use of multi-level feature information in time, space and frequency domains to obtain cluster power prediction results and improve the accuracy of cluster power prediction results.

[0085] Specifically, such as Figure 1 The diagram shown is a flowchart of a wind power cluster power day-ahead prediction method provided in an embodiment of this application. The method includes the following steps S100-S500.

[0086] S100: Based on the tiling method, the subgraph partitioning target is determined, and the correlation matrix constructed in a progressive manner is used as the affinity matrix to partition the wind power cluster, resulting in multiple sub-clusters.

[0087] In this embodiment, spectral clustering can be used to divide the wind power cluster. This involves two key technical points: first, which slicing method to use; and second, how to construct the affinity matrix. For the first technical point, this embodiment chooses the more universal Ncut slicing method. Its core formula is as follows:

[0088]

[0089] In the formula, k represents the number of subgraphs. A i Let A represent the set of nodes within the i-th subgraph, satisfying any A i ∩A j =φ, A1∪A2∪…∪Ak =V, where V is the set of all nodes. For A i The supplement, For A i The sum of the edge weights connecting A to other subsets. i ) is A i The sum of the weights of all edges in the array. Ncut(A1,A2,...,A) K The sum of edge weights for all subgraphs is represented by ). Therefore, the objective of Ncut can be expressed as min Ncut(A1,A2,...,A) K ).

[0090] Regarding the second technical point, this embodiment proposes a progressive correlation matrix construction method, using an affinity matrix as a medium for multi-source information fusion, adding multi-level considerations to the cluster partitioning process. Simultaneously, this embodiment eliminates the need for multiple clustering operations on the multi-source information, simplifying the data preprocessing process. Specifically, it includes the following steps S101-S103.

[0091] S101: Obtain the distance weight matrix based on geographical location information. Calculate the great-circle distance between stations using the Haversine formula. Great-circle distance is the shortest path distance between two points on the Earth's surface, and it more accurately reflects the true geographical location than Euclidean distance (straight-line distance). Great-circle distance D between two stations. ij The calculation formula is as follows:

[0092]

[0093] In the formula, and Let λ be the latitude of station i and station j. i and λ j Let be the longitudes of stations i and j. Let h be the semi-versus of the central angle between the two points, and r be the radius of the Earth. Further, a formula can be used to transform the distance between stations into an expression of correlation coefficients, yielding the distance weight matrix G. DW In other words, the closer the distance, the larger the distance coefficient. The formula is as follows:

[0094]

[0095] In the formula, S is the variance of the distance between all stations, and D ij G represents the calculated great circle distance. DW (i,j) is the distance weight matrix G DW Elements at each position within.

[0096] S102: Filter out long-distance interference information based on the distance matrix and establish wind direction correlation matrix and power correlation matrix. First, in G... DW Based on this, a correlation threshold ε is set, with weights above the threshold remaining unchanged and those below the threshold set to 0. This yields the distance matrix G, ignoring long-distance station connections. DE To ensure that each station node has at least one edge connection, this invention sets the distance correlation threshold ε to 0.6. The matrix formula is as follows:

[0097]

[0098] Then, based on the Pearson correlation of historical data from each station, the wind direction correlation matrix W is obtained. DC And the power correlation matrix W PC The specific calculation formula is shown in (5).

[0099]

[0100] In the formula, x i and y i Let represent the i-th element in different feature sequences, and n represent the sequence length. Substitute the wind direction and power characteristics of each station into formula (5) for calculation to obtain the wind direction correlation matrix W. DC And the power correlation matrix W PC .

[0101] S103: W DC and W PC respectively with G DE The Hadamard product is performed to filter out information from distant locations. Distance, wind direction, and power information are given equal importance. The affinity matrix S, which integrates multi-source information, is used as the basis for spectral clustering. The formula is as follows:

[0102] S = a * G DW +b*G DE W DC +g*G DE W PC (6)

[0103] In the formula, a, b, and g are the fusion weights. When their importance is the same, the three values ​​are equal and their sum is 1. DW W DC and W PC The element values ​​are all in [0, 1].

[0104] S200: Project the latitude and longitude of the stations within the sub-cluster onto a two-dimensional plane, and establish virtual information nodes through the computational geometry center.

[0105] Based on the cluster partitioning results, to provide richer cluster information and mapping methods for cluster prediction, this embodiment proposes a virtual node representing the overall information of the sub-cluster. Within each sub-cluster, the stations are adjacent to each other, and the distance between stations can be represented by Euclidean distance. Similarly, after partitioning, the surface undulations within a small area are approximate. The station latitude and longitude information can be further converted into coordinates on a two-dimensional plane. This embodiment uses Mercator projection, the principle of which is as follows:

[0106]

[0107] In the formula, g is the station number. λ g and x represents the longitude and latitude of station g within the sub-cluster. g With y g These are the coordinates of latitude and longitude projected onto a plane.

[0108] Based on the planar equivalent cluster distribution, this invention proposes a virtual node. For a given sub-cluster, the geometric center of the shape formed by all stations is identified and determined as the location of the virtual information node. The calculation formula is as follows:

[0109]

[0110] In the formula, N represents the number of wind farms within a sub-cluster. n represents the number of sub-clusters. (X n ,Y n The coordinates of the virtual nodes in the nth cluster are shown below. These virtual node coordinates can be adapted to the subsequent dynamic graph structure modeling, which is beneficial for completing the dynamic interaction between single-field and cluster information based on wind direction.

[0111] S300: Based on the sub-clusters obtained from the division, establish a sliding time window for NWP, and identify the prevailing wind direction of the sub-clusters by combining the stable state and abrupt changes of the wind direction data within the window.

[0112] To address the issue of information propagation direction between wind farms, this embodiment extends the identification of prevailing wind direction within a wind farm's window to a method for identifying the prevailing wind direction across the wind farm cluster. To mine the temporal characteristics of NWP data, this embodiment establishes a sliding time window for NWP data and establishes a many-to-one mapping relationship with power data. The wind direction window sequence is denoted as w = [w0, w1, ..., w...]. K-1 K is the width of the time window, w K-1 This refers to the wind direction data for the predicted time, expressed in degrees. Prevalent wind direction identification methods within a certain station window can be categorized into three types:

[0113] Scenario 1: Stable wind direction within the time window. Referring to the eight-directional division method, this embodiment sets the abrupt change boundary at 45°. During periods of stable wind direction, the prevailing wind direction can be calculated using the averaging method. The calculation formula is as follows:

[0114]

[0115] In the formula, w i Let be the i-th element within the wind direction window, and K be the window width. max[w] and min[w] are the maximum and minimum values ​​of the elements within the window, respectively.

[0116] Scenario 2: The wind direction changes abruptly at the prediction time. A change is considered abrupt if the difference between the wind direction at the prediction time and the previous time is greater than 45°. When the wind direction changes abruptly, the directed edge connections between stations will change significantly. In this case, the abruptly changed wind direction (w) should be used. K-1 As the mainstream trend within the window.

[0117] Scenario 3: The abrupt change has passed, but the extreme differences in wind direction within the window are still greater than 45°. In this case, the new wind direction within the window remains near the abruptly changed wind direction or continues to evolve in a slightly fluctuating manner. The prevailing wind direction should be based primarily on the current wind direction information, while also considering nearby historical wind direction trends.

[0118] Cases two and three can be analyzed together. When max[w] - min[w] ≥ 45°, the formula is as follows:

[0119]

[0120] In the formula, w K-1 It shows the wind direction at the latest moment within the window. Among them, w q ∈{w q |min[w]≤w q ≤w K-1} and w p ∈{w p |w K-1 ≤w p Let ≤max[w]} represent the sets of elements used to calculate the prevailing wind direction under different judgment conditions, and let the first set w be the set of elements used to calculate the prevailing wind direction. q The second set w p Both are used to find elements between the latest wind direction and the sudden change in wind direction. M1 and M2 are the number of elements in the first set and the second set, respectively.

[0121] S400: Based on the prevailing wind direction of the sub-cluster, the directed edge connection relationship between the stations is determined. The adjacency matrix of the dynamic weighted directed graph is constructed by combining the distance matrix and the wind speed correlation, and virtual information nodes are integrated to construct the dynamic weighted directed graph.

[0122] Step S400 mainly quantifies the NWP information transmission process, which mainly includes two points: (1) Based on the prevailing wind direction of the cluster, the evolution characteristics of the information propagation direction between stations are represented as a directed graph (set of directed edges E). t (2) Quantify the correlation characteristics and time delay availability of wind speeds between stations, and construct a dynamic weighted directed graph structure (including the weighted adjacency matrix M). W At the same time, by combining sub-cluster virtual nodes, the information representation capability of the graph structure is enhanced from multiple perspectives, including individual sites and the cluster as a whole.

[0123] Regarding the first point, in this embodiment, wind direction is the biggest factor affecting the structure of the directed graph. (The rest of the text appears to be a fragment and doesn't translate directly.) Figure 2 It can be seen that the directed edge connection relationship E t It is established using the unit vector of the prevailing wind direction of the cluster and the direction vectors between the station nodes. And E t The asymmetric wind direction matrix W can be used D Further quantification is performed. When the angle θ between these two vectors is acute, it is assumed that stations a and b have an upstream-downstream relationship under the prevailing wind direction, and the corresponding elements in the wind direction matrix are set to 1. Elements with non-wind direction connections are set to 0. The corresponding formula is as follows:

[0124]

[0125] In the formula, W D It is an asymmetric wind direction matrix, which is an n-order asymmetric matrix, where n is the number of field stations in the sub-cluster. W D (i,j) represent the elements in the asymmetric wind direction matrix, whose values ​​change over time. ab Let e ​​be the unit vector pointing from station a to station b. DIR This represents the unit vector of the prevailing wind direction. ||·|| represents the magnitude calculation; the magnitude of the unit vector is 1.

[0126] The process of constructing the adjacency matrix of a dynamically weighted directed graph can be represented by formula (11).

[0127] M W =[G DE ⊙(W D +I n )]⊙(αM person +βM spearman (12)

[0128] In the formula, G DE To ignore the distance matrix of long-distance station connections, the distance matrix is ​​represented by 0s and 1s, I n It is an identity matrix. The Pearson and Spearman coefficients are calculated for the wind speed sequences at various stations within the window, and a correlation matrix M is established. person Mspearman ∈R n×n α and β are weighting coefficients; in this embodiment, they are given the same importance weight. ⊙ represents the Hadamard product operation of the matrix, establishing wind direction and distance constraints on the weight matrix. (W) D +I n This ensures that all nodes possess autocorrelation. When a certain field station cannot receive valid wind speed information from other field stations, autocorrelation still ensures that the nodes at that field station can complete the modeling using their own NWP (Non-Wave Dynamics Predicted by the Wind Power).

[0129] It should be noted that when merging virtual information nodes to construct a dynamic weighted directed graph, the position coordinates of the virtual information nodes have already been determined in the aforementioned steps. The node information of the virtual information nodes includes the cluster average wind speed. Within a given cluster, the node positions and information of the other nodes are determined based on actual data. For example, each node corresponds to a wind farm, and its position is determined based on the actual location of the wind farm. The node information it carries includes the actual wind speed sequence of that wind farm. The virtual information nodes need to act as global information nodes for the sub-cluster, establishing connections with all wind farm nodes to supplement the global perspective.

[0130] In this embodiment, the constructed dynamic weighted directed graph structure based on wind direction information is as follows: Figure 3 As shown. Figure 3 In the example of cluster A, the directed graph t0 represents the directed edge connections between stations at time t0, and the weighted adjacency matrix t0 presents the connection weights between stations at time t0 as a grid matrix. Here, n is the time step, representing the time interval from t0 to subsequent times. n Indicates t n The directed edges between stations change over time, and the weighted adjacency matrix t n Corresponding to t n The adjacency matrix of the directed graph and its weights changes after n time steps, with the matrix square distribution differing from that at time t0. n+k Indicates t n+k The new directed edge connection state at time t, the weight adjacency matrix n+k Corresponding to t n+k The connection weight matrix at time t, where k represents the connection weight matrix at time step t. n Based on this, after k time steps, the results of the dynamic evolution of the directed graph structure and weighted adjacency matrix over time reflect the time-varying characteristics of the relationship between wind power cluster stations at different time steps.

[0131] Therefore, this embodiment further explores the transmission mechanism of wind speed information between wind farms based on the prevailing wind direction, and constructs a dynamically weighted directed graph model. Virtual nodes established by S200 are introduced into this graph structure to quantify the dynamic characteristics of the connections between wind farms and effectively utilize information during the propagation process to achieve efficient information interaction between the overall characteristics of the cluster and the local characteristics of the wind farm.

[0132] S500: Construct a day-ahead power prediction model. The day-ahead power prediction model takes a dynamic weighted directed graph as input and outputs the day-ahead power prediction results of the wind power cluster.

[0133] In this embodiment, the power day-ahead prediction model is named the FECAM-DSTGCN model, such as... Figure 4 As shown, the FECAM-DSTGCN model integrates two frequency domain attention processes. Frequency domain attention process I initially captures the frequency domain features of the input data, while frequency domain attention process II recaptures the key frequency domain information of the data after temporal and spatial aggregation. The FECAM-DSTGCN model consists of three network layers: the first network layer FECAMⅠ, the spatiotemporal information aggregation module DSTGCNN, and the second network layer FECAMⅡ.

[0134] For the first network layer FECAMⅠ, it processes time series tasks, and for time series tasks, a one-dimensional DCT transform is sufficient. The formulas for the one-dimensional discrete cosine transform and inverse transform are shown in (13), (14) and (15).

[0135]

[0136]

[0137] In the formula, L is the length of the original sequence, i is the coordinate of the original sequence, u is the coordinate of the transformed sequence, i, u∈{0,1,…,L-1}, f(i) is the original sequence, C(u) is the value in the frequency domain after DCT transformation, and α(u) is a compensation coefficient that can make the DCT transformation matrix an orthogonal matrix.

[0138] In the first network layer, FECAMⅠ, the frequency domain analysis process is integrated into a module that can be updated as the model trains. Initially, the input data has a fixed time window L, and each column represents a feature channel, i.e., the wind speed sequence for each station. During frequency domain analysis, these are treated as independent vectors v = [v0, v1, ..., v...]. n-1 ], where v is the original data matrix and n is the number of independent vectors. Then, each independent vector is transformed into a vector representation in the frequency domain according to formula (15):

[0139]

[0140] In the formula, p represents the index of an independent vector, p∈{0,1,…,n-1}, and n is the number of independent vectors. q and u are the coordinates of the sequences before and after the transformation, respectively, and q,u∈{0,1,…,L-1}. Freq represents the q-th element of the p-th vector in the original time series representation. p (u) represents the u-th element of the p-th vector in the frequency domain representation. F DCT (·) denotes the one-dimensional discrete cosine transform operation. The vector lengths before and after the DCT transformation are equal. The resulting Freq... p It can also be called the attention weight vector in the frequency domain, and the attention weights can be continuously learned through a neural network structure. The DCT forward transform process can also be written in matrix form Freq. p =Gv p , where G is the DCT transform matrix. The entire process of the frequency enhancement channel attention mechanism I can be expressed as the following formulas (16) and (17).

[0141] v freqⅠ-p =σ(W 2-Ⅰ δ(W 1-Ⅰ Freq p (17)

[0142] v freq-Ⅰ =f stack ([v freqⅠ-1 ,v freqⅠ-2 …,v freqⅠ-n (18)

[0143] In the formula, W 1-Ⅰ With W 2-Ⅰ For a fully connected layer, σ and δ are ReLU and sigmoid activation functions, respectively. stack (·) indicates a horizontal concatenation operation of vectors. freqⅠ-p The frequency domain attention vectors learned by the model are concatenated to obtain the frequency domain attention matrix v. freq-Ⅰ The frequency domain representation, when multiplied by the original features, can be scaled back to resize all feature channels.

[0144] The spatiotemporal information aggregation module in this embodiment comprises three parts: two temporal convolutional networks and one dynamic graph convolutional network. Based on the original model's "gated temporal convolutional layer," a TCN is used to implement one-dimensional dilated causal convolution, expanding the convolutional field of view to complete temporal feature extraction. The core formula of the TCN is:

[0145]

[0146] In the formula, x(td·k) is the value of the input sequence at time t, F(t) is the output after the convolution operation, k is the size of the convolution kernel, w(k) is the weight of the convolution kernel, and d is the dilation rate.

[0147] The core of dynamic graph convolutional networks is the symmetric normalized Laplacian matrix. The graph convolution formula that accepts a dynamic adjacency matrix is ​​as follows:

[0148]

[0149] In the formula, t is the time step, indicating the update of the graph structure. for The degree matrix; A t Given the input adjacency matrix, It is a dynamic adjacency matrix with autocorrelation, and the actual input is M. W ;I n X is the identity matrix; (l) With X (l+1) are the aggregated features of the l-th and l+1-th layers, respectively; σ(·) is the sigmoid activation function of the nonlinear model; W (l) Let be the weight matrix of layer l.

[0150] The wind speed data after dynamic spatiotemporal graph convolution, in addition to its own time and feature dimensions, also has an added channel dimension. Therefore, in the second network layer FECAMⅡ, the high-dimensional information of each station is followed by an independent MLP layer to achieve dimensionality reduction and aggregation in the time domain. The dimensionality-reduced sequence is then weighted with its own frequency domain weight vector, and the final prediction result is obtained after passing through the MLP layer and the ReLU activation function. The process is represented by formulas (20) and (21):

[0151] v freq-Ⅱi =σ(W 2-Ⅱ δ(W 1-Ⅱ F DCT (f flatten (H i )))) (twenty one)

[0152] Pred i =σ[W fci (v freq-Ⅱi ⊙f flatten (H i ))] (twenty two)

[0153] In the formula, H i f is a high-dimensional representation of the wind speed at the p-th station. flatten (·) represents the flattening and fully connected operations, which transform high-dimensional information back into 1-dimensional information; W 1-Ⅱ With W 2-Ⅱ For a fully connected layer, v freq-Ⅱi is the frequency domain attention vector in the frequency domain attention process, W fci Pred is an independent fully connected layer for each station. i F is the predicted power corresponding to each node. DCT (·) represents the one-dimensional discrete cosine transform operation.

[0154] In some embodiments, simulation calculations can be performed based on steps S100-S500 above. Specifically, a sufficiently long amount of wind power cluster data is used as the simulation input, including historical power datasets, wind speed and direction datasets from the NWP (time resolution 15 min), and the latitude and longitude information of the wind farm; according to S100-S500, the short-term day-ahead forecast of the total power of the wind power cluster is obtained.

[0155] In some embodiments, the root mean square error E is selected. rmse (Root Mean Square Error, RMSE), Mean Absolute Error E MAE Mean Absolute Error (MAE) and correlation coefficient (r1) are used as evaluation metrics for the performance of the prediction model (FECAM-DSTGCN model), and their calculation formulas are as follows:

[0156]

[0157]

[0158] In the formula, n is the number of samples in the test set; P real P represents the actual power output of the wind power plant. pred To predict power; It is the average of the actual power output of wind power; To predict the mean power, C i This refers to the installed capacity.

[0159] The following embodiments of the present invention will further illustrate the feasibility and progressiveness of the method proposed in this application with specific data.

[0160] To maximize the accuracy of wind power prediction, this embodiment focuses on day-ahead wind power cluster prediction, and the proposed method requires extensive historical data support. Based on the National Key Research and Development Program project: Large-Scale Wind and Photovoltaic Multi-Timescale Power Supply Capacity Prediction Technology (2022YFB2403000), this embodiment uses a partial wind farm cluster in Inner Mongolia's western region for practical verification, including 30 wind farms with a total installed capacity of 4671MW. The data period is from June 1, 2020 to December 31, 2020, with a power and NWP sampling interval of 15 minutes, totaling 96 sampling points per day. The training set is from June to November 2020, the validation set is the first 4 days of December 2020, and the remaining 27 days are the test set. Specific implementation steps are as described in steps S100 to S500 above, and will not be repeated here.

[0161] This embodiment selects a wind farm cluster in a certain region for case analysis. The cluster has 30 wind farms with a total installed capacity of 4671MW; the data sampling interval is 15 minutes; according to steps 1 to 4, different clustering methods and different models are used to compare accuracy and obtain prediction results.

[0162] Table 1 Power prediction results of different clustering methods

[0163] Clustering methods NRMSE NMAE <![CDATA[r1]]> Method 1 0.0756 0.0584 0.9201 Method 2 0.0846 0.0660 0.8991 Method 3 0.0862 0.0662 0.9086 Method 4 0.0927 0.0746 0.8795 Method 5 0.0915 0.0729 0.8961 Method 6 0.0961 0.0805 0.9038 Method 7 0.0914 0.0721 0.8934

[0164] The clustering methods involved in Table 1 are explained as follows:

[0165] Method 1: The method proposed in this application uses the sum of the prediction results of virtual nodes as the final result;

[0166] Method 2: Based on geographical and power correlation, the proposed multivariate information fusion method and spectral clustering are used to divide the clusters;

[0167] Method 3: Spectral clustering based solely on geographic location;

[0168] Method 4: Spectral clustering based solely on wind direction correlation;

[0169] Method 5: Spectral clustering based solely on power correlation;

[0170] Method 6: No clustering;

[0171] Method 7: The prediction method is the same as Method 1, using the sum of the prediction results from each station as the final result.

[0172] Under the prediction model of this application, the prediction results obtained by using different clustering methods are as follows: Figure 6As shown in Table 1, the prediction results of each method are compared and analyzed. First, methods 3, 4, and 5 based on single-feature clustering are analyzed. Geographical location information is the most intuitive and typical clustering basis. However, this clustering result is not suitable for subsequent modeling processes. For method 4, there is no direct linear correlation between wind direction and power; it needs to be used in conjunction with other strongly correlated features. Method 5, due to the large fluctuations in the power sequence itself, leads to inaccurate clustering results. Therefore, single-feature clustering methods cannot effectively complete the clustering task. Compared with methods 3, 4, and 5, the NRMSE of the method in this application decreased by 1.06%, 1.71%, and 1.59%, respectively.

[0173] Further comparison with multivariate feature clustering and other methods shows that the proposed method outperforms other methods in both curve fitting and index prediction accuracy. Based on distance partitioning, the proposed method groups wind farms with similar power output and wind direction information into one category, thus accurately identifying the prevailing wind direction of sub-clusters. The inclusion of wind direction and power information further analyzes and corrects the results of geographical location clustering, achieving clustering results more conducive to modeling. To demonstrate the effectiveness of virtual information nodes, the embodiments of this application compare Method 1 and Method 7. It is evident that the prediction results using virtual nodes are superior to the results of single-farm summation, reducing NRMSE by 1.58%. This is due to the fact that virtual nodes can utilize both the convergence effect and fully integrate effective fluctuation information from single farms. The results show that virtual nodes not only supplement richer cluster information, but their own feature mapping relationship is also an important modeling method. Therefore, the subsequent methods proposed in this application all use the prediction results of virtual information nodes.

[0174] Table 2 Comparison of prediction results from different models

[0175] Prediction methods NRMSE NMAE <![CDATA[r1]]> Method A 0.0756 0.0584 0.9201 Method B 0.0801 0.0643 0.9025 Method C 0.0895 0.0738 0.8863 Method D 0.0862 0.0691 0.8941 Method E 0.0869 0.0672 0.8908 Method F 0.0941 0.0727 0.8657

[0176] The methods involved in Table 2 are explained as follows:

[0177] Method A: The method proposed in this invention is based on spectral clustering of geographic and wind direction correlation + graph (dynamic wind / correlation / distance) + FECAM-STGCN;

[0178] Method B: Spectral clustering based on geographic and wind direction correlation + graph (dynamic wind / correlation / distance) + STGCN. Based on the model originally used for traffic flow spatiotemporal prediction, only the format of the accepted data was changed to suit the spatiotemporal prediction task of wind power.

[0179] Method C: Spectral clustering based on geographic and wind direction correlation + graph (dynamic wind / correlation / distance) + GCN. GCN is used to capture the spatial correlation between wind farms to complete cluster power prediction.

[0180] Method D: The spectral clustering method of this application + Transformer. The Transformer model is used to extract the temporal global features of the virtual nodes in the sub-cluster, and the prediction results of the sub-cluster are summed to obtain the total power.

[0181] Method E: Comparison with the Adaptive Graph Convolutional Network (ENDCC-AGCN) using the entropy weighting method. Four typical static graph structures are established using wind speed correlation indices, and a graph switching mechanism is implemented based on NWP information matching. The input graph to the GCN is adaptively changed to complete the modeling and prediction.

[0182] Method F: Spectral clustering based on geographic and wind direction correlation + TPA-LSTM. Typical NWP features such as wind speed and direction are input into multiple parallel LSTM layers to extract temporal features. Then, an attention weight matrix is ​​used to weight and sum the outputs of different LSTMs to capture key information and output the prediction result. The prediction process is the same as Method 4, using a temporal attention mechanism to capture key information and output the prediction result.

[0183] Figure 7 This paper presents a comparison of the prediction results of the proposed method with other mainstream prediction models. Table 2 provides a comparative analysis of the prediction results under each method.

[0184] Method B is essentially an ablation experiment of the prediction model of this invention, used to demonstrate the effectiveness of considering the frequency domain attention module. It can be seen that Method A improves the NRMSE by 0.45% compared to Method B, and the predicted values ​​during the high-output phase of wind power are closer to the actual values. This indicates that the frequency domain module effectively captures the implicit frequency domain information, and its combination with the original time-series information achieves information gain, contributing to the improvement of wind power prediction accuracy.

[0185] This paper continues to compare the advantages of the method of this invention with other deep learning and machine learning models in the task of short-term wind power cluster prediction. Method C uses a basic GCN model to aggregate meteorological data at the spatial level. However, when applied to prediction, the model simply connects the MLP layer without considering temporal correlation processing. Method E completes the graph structure switching through future information matching. Since the four graph structures are fixed, this method still ignores temporal details. In methods D and F, Transformer and TPA-LSTM each use different attention mechanisms, but their focus is still on temporal sequence, rather than taking into account the entire spatiotemporal and frequency domain perspectives. FECAM-STGCN introduces frequency domain information gain. At the same time, through temporal convolution and graph convolution processes, it fully explores the spatiotemporal correlation between various wind farms. Therefore, FDGCA-DSTGCN has stronger data analysis capabilities and higher prediction accuracy.

[0186] The wind power day-ahead prediction method for wind power clusters based on frequency domain information gain and dynamic trend perception proposed in this application fully considers the spatiotemporal characteristics of the cluster sites and the important NWP of wind direction in short-term wind power prediction. It realizes in-depth mining and analysis of cluster data at multiple levels of time, space and frequency domain, and further improves the effectiveness and reliability of day-ahead wind power prediction.

[0187] This application also provides a wind power cluster power day-ahead prediction device, such as... Figure 8 As shown, the day-ahead power forecasting device for this wind power cluster includes:

[0188] The sub-cluster partitioning module 801 is configured to determine the sub-graph partitioning target based on the tiling method, and to partition the wind power cluster using the progressively constructed correlation matrix as the affinity matrix to obtain multiple sub-clusters.

[0189] The virtual node construction module 802 is configured to project the latitude and longitude of the stations within the sub-cluster onto a two-dimensional plane and establish virtual information nodes through the computational geometric center.

[0190] The prevailing wind direction identification module 803 is configured to establish an NWP sliding time window based on the divided sub-clusters, and identify the prevailing wind direction of the sub-clusters by combining the stable state and sudden changes of the wind direction data within the window.

[0191] The directed graph construction module 804 is configured to determine the directed edge connection relationship between the stations based on the prevailing wind direction of the sub-cluster, construct the adjacency matrix of the dynamic weighted directed graph by combining the distance matrix and wind speed correlation, and integrate virtual information nodes to construct the dynamic weighted directed graph.

[0192] The day-ahead power prediction module 805 is configured to construct a day-ahead power prediction model, which takes the dynamic weighted directed graph as input and outputs the day-ahead power prediction result of the wind power cluster.

[0193] This application provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0194] The processor executes computer execution instructions stored in memory, causing the processor to perform the schemes in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0195] Communication buses can be Peripheral Component Interconnect (PCI) buses or Extended Industry Standard Architecture (EISA) buses, etc. System buses can be divided into address buses, data buses, control buses, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0196] The electronic device provided in this application embodiment can be the terminal device described in the above embodiments.

[0197] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the wind power cluster power day-ahead prediction method described in the above embodiments.

[0198] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the wind power cluster power day-ahead prediction method in the above embodiments.

[0199] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0200] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0201] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0202] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0203] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0204] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0205] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0206] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0207] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0208] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0209] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting the day-ahead power output of a wind power cluster, characterized in that, The method includes: The subgraph partitioning target is determined based on the tiling method, and the correlation matrix constructed in a progressive manner is used as the affinity matrix to partition the wind power cluster, resulting in multiple sub-clusters. The latitude and longitude of the stations within the sub-cluster are projected onto a two-dimensional plane, and virtual information nodes are established through the computational geometric center; Based on the sub-clusters obtained from the division, an NWP sliding time window is established, and the prevailing wind direction of the sub-clusters is identified by combining the stable state and abrupt changes of the wind direction data within the window. Based on the prevailing wind direction of the sub-cluster, the directed edge connection relationship between the stations is determined. The adjacency matrix of the dynamic weighted directed graph is constructed by combining the distance matrix and the wind speed correlation, and virtual information nodes are integrated to construct the dynamic weighted directed graph. A day-ahead power prediction model is constructed, which takes the dynamic weighted directed graph as input and outputs the day-ahead power prediction results of the wind power cluster.

2. The wind power cluster day-ahead forecasting method according to claim 1, characterized in that, The subgraph partitioning objective is determined based on the tiling method, including: The goal of subgraph partitioning is to minimize the sum of the edge weights of all subgraphs; wherein the formula for calculating the sum of the edge weights of the subgraphs is: In the formula, Ncut(A1,A2,...,A) K A represents the sum of edge weights across all subgraphs, where k is the number of subgraphs, and A i and A j Let A and B represent the sets of nodes within the i-th and j-th subgraphs, respectively, and satisfy any A i ∩A j =φ, A1∪A2∪…∪A k =V, where V is the set of all nodes. For A i The supplement, For A i The sum of the edge weights of its complement and its complement, vol(A) i ) is A i The sum of the weights of all edges in the equation.

3. The wind power cluster day-ahead forecasting method according to claim 1, characterized in that, Construct the affinity matrix as follows: Based on the geographic location information, the distance weight matrix is ​​calculated using the following formula: In the formula, and Let λ be the latitude of the i-th and j-th stations. i and λ j Let be the longitudes of the i-th and j-th stations, h be the semi-versus of the central angle between the two points, r be the radius of the Earth, and D be the longitude of the station. ij H is the great circle distance between the two stations; H is the variance of all distances between stations; G is the variance of all distances between stations. DW (i,j) is the distance weight matrix G DW For each element at each position in the array, exp is an exponential function with the natural constant as its base; Based on the distance weight matrix, a distance matrix G that ignores connections of far distance stations is constructed by the following formula DE : where G DE (i,j) are the elements of the distance matrix G DE the elements at each position within the matrix, and ε is a set correlation threshold value; Based on the distance matrix G DE The affinity matrix S is constructed by the following equation: S=a*G DW +b*G DE W DC +g*G DE W PC (6) In the formula, a, b, and g are the fusion weights, and W DC For the wind direction correlation matrix, W PC This is the power correlation matrix.

4. The wind power cluster day-ahead forecasting method according to claim 1, characterized in that, The latitude and longitude of the stations within the sub-cluster are projected onto a two-dimensional plane, and virtual information nodes are established through computational geometry centers, including: The latitude and longitude coordinates of the stations within the sub-cluster are projected onto a two-dimensional plane using the following formula: In the formula, g is the station number, λ g and Let x be the longitude and latitude of station g within the sub-cluster. g With y g Let g be the coordinates of the latitude and longitude of station g in the plane projection, and r be the radius of the Earth; The location of the virtual information node is determined by the following formula: In the formula, N is the number of wind farms within a certain sub-cluster, (X n ,Y n ) represents the coordinates of the virtual information node in the nth cluster.

5. The wind power cluster day-ahead forecasting method according to claim 1, characterized in that, Based on the sub-clusters obtained from the partitioning, a sliding time window for NWP (Non-Wave Dynamics) is established. The prevailing wind direction of the sub-clusters is identified by combining the stable state and abrupt changes of wind direction data within the window, including: Establish a sliding time window for NWP, and represent the wind direction window sequence as w = [w0, w1, ..., w K-1 ], where K is the time window length, w0 and w1 represent the wind direction data at the first and second sampling times, respectively, w K-1 The wind direction data is for the time to be predicted; A mutation threshold is set. When the difference between the wind direction data at the time to be predicted and the wind direction at the previous time is less than or equal to the mutation threshold, the wind direction data within the time window is determined to be in a stable state. When the difference between the wind direction data at the time to be predicted and the wind direction at the previous time is greater than the mutation threshold, the wind direction data within the time window is determined to be in a mutation state. Under steady-state conditions, the prevailing wind direction DIR of the sub-cluster is determined by the following formula: In the formula, w i Let w be the i-th element in the wind direction window; max[w] and min[w] are the maximum and minimum values ​​of the elements in the wind direction window, respectively. In the event of a mutation, the prevailing wind direction (DIR) of the sub-cluster is determined using the following formula: In the formula, w q ∈{w q |min[w]≤w q ≤w K-1 } and w p ∈{w p |w K-1 ≤w p ≤max[w]}, representing the first set and the second set respectively. The first set and the second set are used to find the elements between the latest wind direction and the sudden change in wind direction. M1 and M2 are the number of elements in the first set and the second set respectively. q and p are the indices of the elements in the first set and the second set respectively.

6. The wind power cluster day-ahead forecasting method according to claim 1, characterized in that, Based on the prevailing wind direction of the sub-cluster, the directed edge connections between stations are determined. An adjacency matrix of a dynamic weighted directed graph is constructed by combining the distance matrix and wind speed correlation. Virtual information nodes are then integrated to construct the dynamic weighted directed graph, including: Based on the prevailing wind direction of the sub-cluster, the evolution characteristics of the information propagation direction between stations are represented as a directed graph; wherein, the directed graph is traversed by the set of directed edges E. t The set of directed edges is characterized by the asymmetric wind direction matrix W. D Quantization; the asymmetric wind direction matrix W D The calculation formula is: In the formula, W D (i,j) represent the elements in the asymmetric wind direction matrix, e ab Let e ​​be the unit vector pointing from station a to station b. DIR This is the unit vector for the prevailing wind direction, where ||·|| represents the calculation of the modulus. Based on the aforementioned asymmetric wind direction matrix W D The adjacency matrix of the dynamically weighted directed graph is constructed using the following formula: M W =[G DE ⊙(W D +I n )]⊙(αM person +βM spearman ) (12) In the formula, G DE To ignore the distance matrix of long-distance station connections, ⊙ represents the Hadamard product operation of the matrix, I n It is an identity matrix, where α and β are weighting coefficients, and M is the identity matrix. person and M spearman A correlation matrix was established to calculate the Pearson and Spearman coefficients for the wind speed sequences at the stations within the window.

7. The wind power cluster day-ahead forecasting method according to claim 1, characterized in that, The power day-ahead prediction model includes a first network layer, a spatiotemporal information aggregation module, and a second network layer. The input data of the first network layer is a wind speed sequence with a fixed time window L, where each column represents a feature channel and each feature channel is considered an independent vector. The first network layer is used to: perform a one-dimensional discrete cosine transform on each independent vector to convert the original time series data into a frequency domain vector; process the frequency domain vector through a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function to obtain a frequency domain attention vector; concatenate all the frequency domain attention vectors horizontally to form a frequency domain attention matrix, and perform a Hadamard product with the original input data to obtain the output features of the first network layer. The spatiotemporal information aggregation module takes the output features of the first network layer and the adjacency matrix of the dynamically weighted directed graph as input, and uses one-dimensional dilated causal convolution to obtain temporal features; and based on the temporal features, obtains the spatiotemporally aggregated features using the following formula: In the formula, t is the time step; A t Given the input adjacency matrix, It is a dynamic adjacency matrix with autocorrelation; for The degree matrix; I n X is the identity matrix; (l) With X (l+1) These are the spatiotemporal aggregation features of the l-th and l+1-th layers, respectively; σ g W is the sigmoid activation function for the nonlinear model. (l) Here is the weight matrix for layer l; The second network layer takes the high-dimensional features obtained after spatiotemporal aggregation as input and obtains the predicted power of each node using the following formula: v freq-Ⅱi =σ(W 2-Ⅱ δ(W 1-Ⅱ F DCT (f flatten (H i )))) (21) Pred i =σ[W fci (v freq-Ⅱi ⊙f flatten (H i ))] (22) In the formula, H i f is a high-dimensional representation of the wind speed at the p-th station. flatten (·) represents the flattening and fully connected operations, which transform high-dimensional information back into 1-dimensional information; W 1-Ⅱ With W 2-Ⅱ For a fully connected layer, v freq- Ⅱi is the frequency domain attention vector in the frequency domain attention process, W fci Pred is an independent fully connected layer for each station. i F is the predicted power corresponding to each node. DCT (·) represents the one-dimensional discrete cosine transform operation.

8. A wind power cluster day-ahead forecasting device, characterized in that, The device includes: The sub-cluster partitioning module is configured to determine the sub-graph partitioning target based on the tiling method, and to partition the wind power cluster using a progressively constructed correlation matrix as the affinity matrix to obtain multiple sub-clusters. The virtual node construction module is configured to project the latitude and longitude of the stations within the sub-cluster onto a two-dimensional plane and establish virtual information nodes through the computational geometric center; The prevailing wind direction identification module is configured to establish an NWP sliding time window based on the divided sub-clusters, and identify the prevailing wind direction of the sub-clusters by combining the stable state and sudden changes of the wind direction data within the window. The directed graph construction module is configured to determine the directed edge connection relationship between the stations based on the prevailing wind direction of the sub-cluster, construct the adjacency matrix of the dynamic weighted directed graph by combining the distance matrix and wind speed correlation, and integrate virtual information nodes to construct the dynamic weighted directed graph. The day-ahead power prediction module is configured to construct a day-ahead power prediction model, which takes the dynamic weighted directed graph as input and outputs the day-ahead power prediction results of the wind power cluster.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the wind power cluster day-ahead prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the wind power cluster day-ahead prediction method as described in any one of claims 1-7.

Images