Wind power cluster power prediction method, system and device considering space-time correlation and medium

Abstract

The invention discloses a wind power cluster power prediction method, system and device considering space-time correlation and a medium. The method comprises the steps of obtaining power grid coverage area information; grid division is carried out on the power grid coverage area based on the power grid coverage area information, the sum of power of all wind power plants in a grid is calculated, and a historical power time sequence is generated; obtaining an initialized adjacency matrix according to the historical power time sequence in combination with the grid division result; and inputting the historical power time sequence and the initialized adjacency matrix into a wind power cluster power prediction model, obtaining aggregation characteristics through the wind power cluster power prediction model, and obtaining wind power cluster power prediction data based on the aggregation characteristics. According to the method, the regional overall trend and local difference in the power change of the wind power plant can be effectively captured, the problem that the prediction precision is insufficient due to the fact that a traditional bottom-to-top accumulation or single-station model does not fully consider the space-time coupling characteristic is solved, and the accuracy of large power grid wind power cluster power prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power prediction, and in particular to a wind power cluster power prediction method considering space-time correlation, a system, a device and a medium. BACKGROUND

[0002] With the continuous increase of new energy installed capacity in China, renewable energy represented by wind power is experiencing unprecedented scale and cluster development. However, this rapid development has also brought challenges. The inherent randomness and volatility of wind energy resources mean that large-scale, high-density wind power clusters connected to the power grid will inevitably pose a potential threat to the safe and stable operation of regional power systems, increasing the complexity and uncertainty of power grid dispatching. Therefore, achieving accurate and efficient short-term prediction of large power grid wind power cluster power is not only the cornerstone of ensuring power grid safety and improving dispatching economy, but also the key technical bottleneck to solving the problem of large-scale wind power grid integration and improving renewable energy utilization.

[0003] At present, there are many studies on power prediction technology for a single wind farm. For the prediction of the total power of the regional large power grid, either a "bottom-up" accumulation method is used to simply superimpose the prediction results of each station, or a total power historical data modeling method similar to a single station model is used. In actual application, the prediction accuracy is not high. Different wind farms in a large area are scattered in different locations, and the wind condition changes have both regional overall trends and significant local differences. The distribution of power changes of different wind farms indirectly reflects the spatial movement and evolution process of weather systems (such as wind belts and pressure fields) and the key "spatial-temporal shift" information. If the complex space-time coupling characteristics of the power of each wind farm can be more effectively modeled and mined in the prediction of the power of large-area wind farms, it will help to improve the accuracy of the prediction of the power of regional wind power clusters. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a wind power cluster power prediction method considering space-time correlation to solve the problem in the prior art that the prediction accuracy is not high when the power of a regional large power grid wind power cluster is predicted because the complex space-time coupling characteristics of the power of different wind farms are not effectively modeled and mined.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a wind power cluster power prediction method considering space-time correlation, comprising:

[0008] obtaining power grid coverage area information;

[0009] grid the power grid coverage area based on the power grid coverage area information, calculate the sum of power of all wind farms in the grid, and generate a historical power time series;

[0010] According to the historical power time series, combined with the grid division result, an initialization adjacency matrix is obtained.

[0011] The historical power time series and the initialization adjacency matrix are input into a wind power cluster power prediction model, and the aggregation features are obtained through the wind power cluster power prediction model, and the wind power cluster power prediction data is obtained based on the aggregation features.

[0012] As a preferred scheme of the wind power cluster power prediction method considering space-time correlation, wherein: calculating the sum of power of all wind farms in the grid to generate a historical power time series comprises:

[0013] Collecting real-time active power data of all wind farms in each grid at each time, superimposing the active power of all wind farms in the grid at the same time to obtain the total power of each grid at each time;

[0014] Arranging the total power at each time according to the set condition, and setting a time interval to obtain the historical power time series of each grid.

[0015] The beneficial effects of the preferred technical scheme are: by summing the power of wind farms in the grid and organizing the time series, the dispersed wind farm power data can be converted into space-time distribution data with geographical area attributes, which not only retains the local characteristics of wind power in different regions, but also provides structured input for subsequent mining of space-time correlation between grids, avoiding the fragmentation problem of single station data.

[0016] As a preferred scheme of the wind power cluster power prediction method considering space-time correlation, wherein: according to the historical power time series, combined with the grid division result, an initialization adjacency matrix is obtained, comprising:

[0017] Based on the number of divided grids, an initial matrix is constructed.

[0018] Using the historical power time series of each grid, the correlation value of power change between different grids is calculated, and the correlation value is generated as an initial element value in the initial matrix.

[0019] The initial matrix is initialized to obtain an initialization adjacency matrix.

[0020] The beneficial effects of the preferred technical solution are: the initial adjacency matrix is constructed by the correlation of the historical power sequence, which can preliminarily quantify the correlation strength of different grid wind power, provide reliable initial parameters for the dynamic position learning module, make the subsequent learning of the dynamic adjacency matrix more in line with the actual spatio-temporal correlation law, and avoid the deviation caused by random initialization.

[0021] As a preferred scheme of the wind power cluster power prediction method considering spatio-temporal correlation provided by the application, the wind power cluster power prediction model comprises a dynamic position learning module and a spatio-temporal attention parallel feature aggregation module.

[0022] The dynamic position learning module adopts the KAN mechanism to perform embedding processing on the input data, and obtains embedding features through a full connection layer and an activation function.

[0023] The initial adjacency matrix is converted into a dynamic adjacency matrix based on the KAN mechanism, and a feature matrix with fused spatio-temporal correlation is obtained through fusion calculation of the dynamic adjacency matrix and the embedding features.

[0024] The spatio-temporal attention parallel feature aggregation module performs spatio-temporal dimension conversion on the feature matrix, generates attention scores in the spatial and temporal dimensions through self-attention mechanisms respectively, and aggregates features to obtain aggregated features.

[0025] The beneficial effects of the preferred technical solution are: the dynamic position learning module realizes autonomous learning of nonlinear and dynamic spatio-temporal correlation between grids through the KAN mechanism, breaking through the limitations of traditional fixed adjacency matrices; the spatio-temporal attention parallel aggregation module simultaneously excavates spatial distribution features and time series trends, realizes efficient fusion of spatio-temporal features, and the combination of the two enables the model to accurately capture the spatio-temporal coupling characteristics of wind power, thereby improving the prediction accuracy of regional wind power clusters.

[0026] As a preferred scheme of the wind power cluster power prediction method considering spatio-temporal correlation provided by the application, the spatio-temporal attention parallel feature aggregation module comprises:

[0027] The dimension of the feature matrix with fused spatio-temporal correlation output by the dynamic position learning module is converted to obtain a spatial dimension feature matrix and a temporal dimension feature matrix.

[0028] The spatial dimension feature matrix and the temporal dimension feature matrix are respectively mapped to high-dimensional features to generate spatial parameters of a spatial self-attention mechanism and temporal parameters of a temporal self-attention mechanism.

[0029] The spatial attention score is obtained by transposing the spatial parameters and calculating through a softmax function, and the temporal attention score is obtained by transposing the temporal parameters and calculating through a softmax function.

[0030] The spatial attention score is operated with a spatial parameter, and the time attention score is operated with a time parameter, to obtain a spatial attention output and a time attention output, and the spatial attention output and the time attention output are dimensionally adjusted and algebraically added to obtain the aggregated feature.

[0031] As a preferred scheme of the wind power cluster power prediction method considering space-time correlation, the calculation of the spatial attention score and the time attention score is represented as:

[0032]

[0033] wherein M t is the spatial attention score, M s is the time attention score, Q s and is the spatial parameter, Q t and is the time parameter, T is the number of time points, N is the number of grids, and D is the number of days to be predicted.

[0034] As a preferred scheme of the wind power cluster power prediction method considering space-time correlation, the wind power cluster power prediction data is obtained based on the aggregated feature, including:

[0035] The aggregated feature is input into a fully connected layer, the aggregated feature is dimensionally converted and mapped through the fully connected layer, and after being processed by an activation function, a time series of total wind power of the whole network on the prediction day is generated;

[0036] The time series contains wind power cluster power prediction data at each time point within a day, and the time interval is consistent with the time interval of the historical power time series.

[0037] In a second aspect, the present application provides a wind power cluster power prediction system considering space-time correlation, including: a data acquisition module for acquiring power grid coverage area information;

[0038] A division module is configured to divide the power grid coverage area into grids based on the power grid coverage area information.

[0039] A data processing module is configured to calculate the sum of powers of all wind farms in the grid to generate a historical power time series, and obtain an initial adjacency matrix based on the historical power time series and the grid division result.

[0040] A data prediction module is configured to input the historical power time series and the initial adjacency matrix into a wind power cluster power prediction model, obtain an aggregated feature through the wind power cluster power prediction model, and obtain wind power cluster power prediction data based on the aggregated feature.

[0041] In a third aspect, the present application provides an electronic device comprising:

[0042] a memory and a processor;

[0043] The memory is configured to store computer-executable instructions, and the processor is configured to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the wind power cluster power prediction method considering spatio-temporal correlation.

[0044] In a fourth aspect, the present application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the wind power cluster power prediction method considering spatio-temporal correlation.

[0045] Compared with the prior art, the present application has the following beneficial effects: by dividing a large area into non-uniform grids and generating grid historical power time series, combining a dynamic position learning module with a KAN mechanism to autonomously mine the complex spatio-temporal correlation of wind power in different grids, and a spatio-temporal attention parallel feature aggregation module to aggregate features in time and space dimensions in parallel, the present application can effectively capture regional overall trends and local differences in wind farm power changes, overcome the problem of insufficient prediction accuracy caused by traditional bottom-up accumulation or single station models due to insufficient consideration of spatio-temporal coupling characteristics, and improve the accuracy of large power grid wind power cluster power prediction, thereby providing a more reliable decision basis for power grid dispatching, wind power consumption, and safe and stable operation of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 The overall flowchart of the wind power cluster power prediction method considering spatio-temporal correlation according to an embodiment of the present application.

[0048] Figure 2 The wind power cluster power prediction model diagram of the wind power cluster power prediction method considering spatio-temporal correlation according to an embodiment of the present application.

[0049] Figure 3 The wind power cluster power prediction model training loss function diagram of the wind power cluster power prediction method considering spatio-temporal correlation according to an embodiment of the present application.

[0050] Figure 4A wind power cluster power prediction model prediction result comparison diagram of the wind power cluster power prediction method considering space-time correlation for an embodiment of the present application is shown. DETAILED DESCRIPTION

[0051] In order to make the above-mentioned objects, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0052] Embodiment 1, refer to Figure 1 For an embodiment of the present application, a wind power cluster power prediction method considering space-time correlation is provided, comprising:

[0053] S100: Obtain power grid coverage area information;

[0054] S101: Based on the power grid coverage area information, divide the power grid coverage area into grids, calculate the sum of the power of all wind farms in the grid, and generate a historical power time series;

[0055] S102: According to the historical power time series, combined with the grid division result, obtain an initialization adjacency matrix;

[0056] S103: Input the historical power time series and the initialization adjacency matrix into a wind power cluster power prediction model, obtain aggregated features through the wind power cluster power prediction model, and obtain wind power cluster power prediction data based on the aggregated features.

[0057] It should be noted that the power changes of different wind farms in a large area are jointly affected by regional weather trends and local wind conditions, and there are complex space-time coupling characteristics. The traditional bottom-up accumulation or single station model does not effectively mine these characteristics, resulting in low prediction accuracy. The present application can integrate the dispersed wind farm power into a grid power sequence with geographical attributes by obtaining the power grid coverage area information for grid division, and can lay a foundation for capturing the correlation between grids by constructing an initialization adjacency matrix combined with the historical power time series. Then, by inputting the prediction model containing a dynamic position learning module and a space-time attention parallel feature aggregation module, the space-time correlation can be autonomously extracted and the features can be aggregated, and finally the precise prediction can be realized. The present application effectively mines the space-time correlation of the power of different regional wind farms, improves the prediction accuracy of the wind power cluster power of the large power grid, and provides a reliable basis for power grid dispatching, wind power consumption and safe and stable operation of the power system.

[0058] Embodiment 2, refer to Figures 1-2And Table 1, for an embodiment of the application, based on the above embodiment, a wind power cluster power prediction method considering space-time correlation is provided.

[0059] In the embodiment of the application, the power grid coverage area information is obtained by linking the power grid geographic information system with the wind farm account database in step S100, and the power grid coverage area information includes administrative division boundaries, wind farm geographic location coordinates, power transmission line distribution ranges, topographic features, and the like.

[0060] In an optional embodiment, the power grid coverage area information is obtained by calling the power grid geographic information platform interface of the provincial power dispatch center, and the power grid coverage area information includes municipal administrative boundary vector data, wind farm administrative region codes, geographic reference coordinate systems required for grid division, and main weather station distribution locations within the region.

[0061] In another optional embodiment, the power grid coverage area information is obtained by integrating remote sensing image data and wind farm real-time monitoring system data, and the power grid coverage area information includes geographic ranges of regional remote sensing images, latitude and longitude boundary boxes of wind farms, division results of different wind resource areas, total areas and shape parameters of power grid coverage areas, and the like.

[0062] In the embodiment of the application, the power grid coverage area is divided based on the power grid coverage area information in step S101, the sum of the powers of all wind farms within the grid is calculated, and a historical power time series is generated, further including sub-steps A1-A2:

[0063] A1: Collecting real-time active power data of all wind farms within each grid at each time, superimposing and calculating the active powers of all wind farms within the grid at the same time to obtain the total power of each grid at each time;

[0064] A2: Arranging the total power at each time according to the set conditions, and setting a time interval to obtain a historical power time series of each grid.

[0065] Specifically, the geographic area covered by the large power grid is divided into non-uniform irregular grids, which are further merged and split based on administrative division.

[0066] For example, a provincial power grid can be based on regions, and adjacent cities with small differences in wind resources or small geographic areas can be merged, and cities with large differences in wind resources or large geographic areas can be subdivided into multiple geographic grids. After grid division, it will remain fixed in long-term operation.

[0067] Let the grid number be [1, N], for any grid i∈[1, N], calculate the sum of the wind farm powers within the grid at each time t to form a historical power time series P i (t).

[0068] In an alternative embodiment, the total power of each time can be arranged in chronological order;

[0069] For example, the total power of each time is recorded in 15-minute intervals starting from 00:00, forming an ordered sequence from t=1 (00:00-00:15) to t=96 (23:45-24:00), ensuring that the historical power time series P i (t) follows the logic of time passing, intuitively reflecting the timing change law of wind power within a day.

[0070] In another alternative embodiment, the total power of each time can be arranged according to the power change trend characteristics;

[0071] For example, the slope of the total power of each time within a day is calculated, and the time of the power rising stage, the stable stage and the falling stage is classified respectively, arranged in the order of rising stage-stable stage-falling stage, and the time sequence is maintained within each stage, so that the historical power time series P i (t) highlights the dynamic change characteristics of wind power, facilitating the subsequent model to capture the trend of power fluctuation.

[0072] In the embodiment of the present application, the initialization adjacency matrix obtained in step S102 according to the historical power time series combined with the grid division result further includes sub-steps B1-B3:

[0073] B1: based on the number of divided grids, an initial matrix is constructed;

[0074] B2: using the historical power time series of each grid, the correlation value of power change between different grids is calculated, and the correlation value is generated as the initial element value in the initial matrix;

[0075] B3: the initial matrix is initialized to obtain the initialization adjacency matrix.

[0076] Specifically, based on the number of divided grids N, an initial matrix with dimension N×N is constructed, each element in the matrix corresponds to the correlation between two different grids; using the historical power time series P i (t) and P j (t) of each grid (i, j∈[1, N] and i≠j);

[0077] The correlation value of power change between different grids is obtained by calculating the Pearson correlation coefficient or cosine similarity, and the value is assigned to the element in the i-th row and the j-th column of the initial matrix to generate all element values of the initial matrix; the initial matrix is randomly initialized by using the Kaiming initialization function in the pytorch library to obtain an initialized adjacency matrix A0eR N×N , which is the initial input of the dynamic position learning module and lays a foundation for subsequent learning of the dynamic spatial and temporal correlation between grids through the KAN mechanism.

[0078] In an optional embodiment, the initial matrix is randomly initialized by the Xavier initialization method. Specifically, based on the dimension N*N of the initial matrix, the random value of the element in the initial matrix satisfies the normal distribution with a mean of 0 and a variance of 2 / (N+N) according to the Xavier initialization strategy. The element value scale of the matrix initialized in this way is adapted to the input and output dimensions, which can avoid the problems of gradient vanishing or explosion in the model training process caused by excessively large or small initial values, and provides stable initial parameters for subsequent learning of the dynamic adjacency matrix.

[0079] In another optional embodiment, the initial matrix is randomly initialized by uniform distribution random sampling. Specifically, uniform random sampling is performed in the interval , and each element of the initial matrix is assigned a value, where N is the number of grids, and the interval range is determined based on the input and output dimensions of the matrix, which ensures that the element values of the initial matrix are distributed within a reasonable range, retaining a certain randomness to support subsequent dynamic learning, and avoiding extreme values to interfere with the initial training stage of the model, thereby ensuring the stable start of the dynamic position learning module.

[0080] In the embodiments of the present application, the historical power time series and the initialized adjacency matrix are input into the wind power cluster power prediction model in step S103, the aggregated features are obtained through the wind power cluster power prediction model, and the wind power cluster power prediction data is obtained based on the aggregated features, which further includes sub-steps C1-C3:

[0081] C1: the wind power cluster power prediction model includes a dynamic position learning module and a spatio-temporal attention parallel feature aggregation module;

[0082] C2: the dynamic position learning module uses the KAN mechanism to embed the input data, and obtains embedded features through a full connection layer and an activation function; the initial adjacency matrix is converted into a dynamic adjacency matrix based on the KAN mechanism, and a feature matrix with fused spatio-temporal correlation is obtained through fusion calculation of the dynamic adjacency matrix and the embedded features;

[0083] C3: The spatio-temporal attention parallel feature aggregation module converts the feature matrix in the spatio-temporal dimension, respectively generates attention scores in the spatial and temporal dimensions through the self-attention mechanism, and aggregates the features to obtain aggregated features.

[0084] In the embodiment of the application, the wind power cluster power prediction model is as shown in the figure. Figure 2 The wind power cluster power prediction model mines the change relationship between each wind power station in the region through the KAN-containing dynamic position learning module, the edges of the KAN-containing dynamic position learning module are continuous functions, which match the continuous change nature of the wind speed in the region, represents the dot product of two matrices.

[0085] The adjacency matrix and the B-spline function to be fitted in the KAN-containing dynamic position learning module and other parameters are updated in the overall training process of the model.

[0086] For example, the output of the KAN-containing dynamic position learning module is denoted as X, and is represented as:

[0087] X=A ′ ·X ′ emb

[0088] X ′ emb =Relu(WP h +b)

[0089] A ′ =f1(A)+f2(A)+…+f n (A)

[0090] Where, Relu() represents an activation function; W represents a fully connected layer weight; b represents a fully connected layer bias; f i (·) represents a B-spline function to be fitted in the KAN module, P h is the input historical data, A ′ is the adjacency matrix.

[0091] In the embodiment of the application, the spatio-temporal attention parallel feature aggregation module can realize the feature aggregation of wind power in time and space in parallel through the self-attention model.

[0092] Let X=X s ∈R T×N×c , where T represents the number of time points, N represents the number of spatial grids, and c represents the feature dimension; exchange the spatio-temporal dimension of X to obtain X t ∈R N×T×D .

[0093] Respectively, the spatial dimension feature matrix X s and the time dimension feature matrix Xt High-dimensional feature mapping is performed to obtain the input X′ of the temporal and spatial self-attention mechanisms, respectively. t ∈R N×T×D and X′ s ∈R T×N×D The input data is multiplied by different coefficient matrices to obtain different Q, K, and V values, which serve as the inputs for the temporal and spatial self-attention mechanisms. The outputs of the two self-attention mechanisms are then adjusted in dimensionality and algebraically summed to obtain the final output.

[0094] For example, let the input be X, and the output of the self-attention mechanism be Y, represented as:

[0095] Q = XW Q K = XW K Q = XW K

[0096]

[0097] Y = M·V

[0098] Where D represents the model dimension and M represents the attention score.

[0099] When the input contains both spatiotemporal information, the input X∈R T×N×D Where L represents the time length, N represents the number of nodes, and D represents the number of features.

[0100] For example, when mining spatial information, the spatial self-attention score M T The calculation formula is:

[0101]

[0102] For example, when mining time information, the time self-attention score M s The calculation formula is:

[0103]

[0104] Finally, the formula for calculating the attention output Y, which takes into account spatiotemporal relationships, is as follows:

[0105] Y = M t ·V+M s ·V

[0106] Among them, M t The score represents the temporal self-attention score, where V represents the temporal and spatial parameters, and M represents the temporal self-attention score. s The score is the spatial self-attention score.

[0107] In an alternative embodiment, the aggregated features are input into a fully connected layer, the aggregated features are dimensionally converted and mapped by the fully connected layer, and after being processed by an activation function, the time series of the total wind power of the entire network on the prediction day is generated; the time series contains wind power cluster power prediction data at each time of the day, and the time interval is consistent with the time interval of the historical power time series.

[0108] It should be noted that each grid wind power historical day time series input is first weighted based on geographical location influence by the <dynamic location learning module with KAN mechanism>, and then the spatial attention aggregation and the time attention aggregation of the power of different grid wind power stations are realized by the <spatio-temporal attention parallel feature aggregation module>. Finally, the next day time series prediction result of the total wind power of the power grid is output by the fully connected layer.

[0109] In an alternative embodiment, taking T=96 as an example, the wind power cluster power prediction model input is the historical wind power data P in ∈R 96×N×k of N grids in [D-1, …, D-k) days. out ∈R 96×1 The input and output dimensions of each module in the wind power cluster power prediction model are set as shown in Table 1.

[0110] Table 1 Input and output of each module of the wind power cluster power prediction model

[0111] Module number Network layer Input data dimension Output data dimension 1 Dynamic position learning (batch, 96, N, k) (batch, 96, N, k) 2 Spatio-temporal attention (batch, 96, N, k) (batch, 96, N, 72) 3 Fully connected layer (batch, 96, N, 72) (batch, 96, 1, 72) 4 Activation function (batch, 96, 1, 72) (batch, 96, 1, 72) 5 Fully connected layer (batch, 96, 1, 72) (batch, 96, 1)

[0112] In an alternative embodiment, a multi-scale fully connected layer stacking method is used to obtain wind power cluster power prediction data.

[0113] For example, the aggregated features are divided into multiple sub-sequences according to the time scale, each sub-sequence corresponds to a group of fully connected layers, and different sub-scale fully connected layers are used to extract local features and map dimensions of the aggregated features, respectively. After splicing the output results of each sub-scale, the global fully connected layer is input, and a nonlinear conversion is performed by combining a Sigmoid activation function to generate a time series of the total wind power of the entire network on the prediction day, wherein the weight parameters of each sub-scale fully connected layer are independently trained to adapt to the change rule of wind power at different time periods.

[0114] In another alternative embodiment, a time series convolution network combined with a fully connected layer is used to obtain wind power cluster power prediction data.

[0115] Exemplarily, the aggregated features are input into a time sequence convolution network, long-distance dependencies in the time dimension are captured through multiple convolution layers with expansion rates, time sequence features are extracted and converted into one-dimensional vectors through a flatten operation, and then input into a fully connected layer for dimension compression and mapping, and finally output the predicted daily total wind power time sequence of the whole network through a linear activation function, wherein the convolution kernel size of the time sequence convolution network is set according to the periodic characteristics of the historical power time sequence, so as to enhance the ability to capture power fluctuation cycles.

[0116] Embodiment 3, refer to Figures 3-4 For an embodiment of the present application, based on the above embodiment, a wind power cluster power prediction method considering spatio-temporal correlation is provided, and scientific demonstration is carried out through experiments to verify the beneficial effects of the present application.

[0117] The running data of a certain province from January 2022 to December 2023 is used to form the calculation example data, and modeling and testing are carried out. The data set has 15 minutes per day, T = 96.

[0118] The province is divided into 14 grids, and the wind power of the grid is summed up point by point. The maximum and minimum normalization is performed on all data according to the installed capacity of the wind power in the grid.

[0119] The test example takes k = 1, that is, the model input is the measured data of the wind power of each grid in D-1 days, and the model output is the predicted total wind power of the province on the Dth day.

[0120] The model loss function is MSE, and the optimizer is Adam.

[0121] The training period is set to 8, and the following is the loss function change process of the model during the training process as shown in Figure 3 During the training process of the wind power cluster power prediction model considering spatio-temporal correlation, the training loss (blue line), validation loss (orange line), and test loss (yellow line) gradually decrease with the cycle. The training loss decreases from 0.0257 to 0.0037, and the validation and test losses are continuously optimized and tend to be stable, reflecting that the model has more accurate spatio-temporal correlation mining of the wind power cluster power in the power grid coverage area, and the prediction performance gradually improves, verifying the ability of the model to effectively mine spatio-temporal correlation and improve prediction accuracy, and providing a reliable basis for power grid dispatching.

[0122] In addition, a test sample is randomly selected, and the historical data of the wind power of each grid in D-1 days is shown in the following table:

[0123]

[0124] The power prediction result of the second day D output by the model is:

[0125] P f | 96 = [0.249, 0.241, 0.234, 0.226, 0.225, …, 0.201, 0.188, 0.197, 0.195, 0.187]

[0126] The real result of the day is:

[0127] P h | 96 = [0.289, 0.278, 0.275, 0.275, 0.256, …, 0.198, 0.194, 0.199, 0.201, 0.221]

[0128] The 96-point sequence is long, so not all input and output results are listed above. The comparison of the prediction result and the second-day actual measurement sample is shown in the following figure. Figure 4

[0129] Figure 4 Among them, the green line is the total provincial wind power generation of D-1 day, and the blue line is the total provincial wind power generation of the next day D. It can be seen that there is a significant change in the power generation of the two days. The prediction result of the total provincial wind power generation of D day by the model in this paper is the red dotted line. The result has a very high consistency with the actual power generation, and the prediction accuracy is good.

[0130] In the above, a schematic scheme of a wind power cluster power prediction method considering time and space correlation is described. It should be noted that the technical scheme of the wind power cluster power prediction system considering time and space correlation belongs to the same concept as the technical scheme of the wind power cluster power prediction method described above. The technical scheme of the wind power cluster power prediction system considering time and space correlation in this embodiment is not described in detail. Please refer to the description of the technical scheme of the wind power cluster power prediction method considering time and space correlation described above.

[0131] The embodiment also provides a wind power cluster power prediction system considering time and space correlation, which comprises:

[0132] A data acquisition module is configured to acquire power grid coverage area information.

[0133] A division module is configured to divide the power grid coverage area into grids based on the power grid coverage area information.

[0134] A data processing module is configured to calculate the sum of the power of all wind farms in the grid, generate a historical power time sequence, and obtain an initialized adjacency matrix based on the historical power time sequence and the grid division result.

[0135] ​The data prediction module is configured to input the historical power time sequence and the initialized adjacency matrix into a wind power cluster power prediction model, acquire aggregated features through the wind power cluster power prediction model, and acquire wind power cluster power prediction data based on the aggregated features.

[0136] The embodiment also provides an electronic device suitable for the wind power cluster power prediction considering the space-time correlation, which comprises a memory and a processor.

[0137] The embodiment also provides a storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the method for the wind power cluster power prediction considering the space-time correlation.

[0138] The storage medium provided by the embodiment belongs to the same inventive concept as the method for the wind power cluster power prediction considering the space-time correlation, and the technical details not described in the embodiment can be referred to the above embodiments, and the embodiment has the same beneficial effects as the above embodiments.

[0139] From the above description about the embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software and necessary universal hardware, and of course can also be realized by hardware. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, and the computer software product can be stored in a computer readable storage medium, such as a floppy disk, a ROM, a RAM, a FLASH, a hard disk, or an optical disc, and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of various embodiments of the present application.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A method for predicting the power of wind power clusters considering spatiotemporal correlation, characterized in that, include: Obtain information on the power grid coverage area; Based on the grid coverage area information, the grid coverage area is divided into grids, the sum of the power of all wind farms in the grid is calculated, and a historical power time series is generated. Based on the historical power time series and the grid partitioning results, the initial adjacency matrix is ​​obtained; The historical power time series and the initial adjacency matrix are input into the wind power cluster power prediction model. Aggregation features are obtained through the wind power cluster power prediction model, and wind power cluster power prediction data are obtained based on the aggregation features.

2. The wind power cluster power prediction method considering spatiotemporal correlation as described in claim 1, characterized in that, Calculate the sum of power from all wind farms within the grid to generate a historical power time series, including: Collect real-time active power data of all wind farms in each grid at each time, and calculate the total power of each grid at each time by superimposing the active power of all wind farms in the grid at the same time. The total power at each time point is arranged according to the set conditions, and the historical power time series of each grid is obtained by setting the time interval.

3. The wind power cluster power prediction method considering spatiotemporal correlation as described in claim 2, characterized in that, Based on the historical power time series and the grid partitioning results, an initial adjacency matrix is ​​obtained, including: Construct an initial matrix based on the number of grid cells after partitioning; Using the historical power time series of each grid, the correlation values ​​of power changes between different grids are calculated, and the correlation values ​​are used to generate the initial element values ​​in the initial matrix; The initial matrix is ​​initialized to obtain the initial adjacency matrix.

4. The wind power cluster power prediction method considering spatiotemporal correlation as described in claim 3, characterized in that, The wind power cluster power prediction model includes a dynamic location learning module and a spatiotemporal attention parallel feature aggregation module. The dynamic position learning module uses the KAN mechanism to embed the input data, and obtains the embedding features through a fully connected layer and an activation function. Based on the KAN mechanism, the initial adjacency matrix is ​​converted into a dynamic adjacency matrix. By fusing the dynamic adjacency matrix with the embedded features, a feature matrix with fused spatiotemporal correlation is obtained. The spatiotemporal attention parallel feature aggregation module performs spatiotemporal dimension transformation on the feature matrix, generates attention scores in the spatial and temporal dimensions through a self-attention mechanism, and aggregates the features to obtain aggregated features.

5. The wind power cluster power prediction method considering spatiotemporal correlation as described in claim 4, characterized in that, The spatiotemporal attention parallel feature aggregation module includes: The feature matrix of fusion spatiotemporal correlation output by the dynamic position learning module is transformed to obtain the spatial dimension feature matrix and the temporal dimension feature matrix. High-dimensional feature mapping is performed on the spatial and temporal feature matrices respectively to generate spatial parameters of the spatial self-attention mechanism and temporal parameters of the temporal self-attention mechanism. Spatial attention score is obtained by transposing the spatial parameters and calculating the result using the softmax function; temporal attention score is obtained by transposing the temporal parameters and calculating the result using the softmax function. The spatial attention score is calculated with spatial parameters, and the temporal attention score is calculated with temporal parameters to obtain spatial attention output and temporal attention output. The spatial attention output and temporal attention output are then adjusted in dimension and algebraically added to obtain aggregated features.

6. The wind power cluster power prediction method considering spatiotemporal correlation as described in claim 5, characterized in that, The spatial attention score and temporal attention score are calculated and represented as follows: Among them, M t For spatial attention scores, M s For time attention score, Q s and Q is a spatial parameter. t and Here, T represents the number of time points, N represents the number of grids, and D represents the number of days to be predicted.

7. The wind power cluster power prediction method considering spatiotemporal correlation as described in claim 1, characterized in that, Based on the aggregation features, wind power cluster power prediction data is obtained, including: The aggregated features are input into a fully connected layer, and the fully connected layer performs dimensional transformation and mapping on the aggregated features. After processing by an activation function, a time series of the total wind power of the entire network for the predicted day is generated. The time series includes wind power cluster power prediction data for each time point within a day, with time intervals consistent with the time intervals of historical power time series.

8. A wind power cluster power prediction system considering spatiotemporal correlation, using the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire information about the power grid coverage area; The grid division module is used to divide the power grid coverage area into grids based on the power grid coverage area information. The data processing module is used to calculate the sum of the power of all wind farms within the grid and generate a historical power time series; based on the historical power time series and the grid partitioning results, an initial adjacency matrix is ​​obtained. The data prediction module is used to input the historical power time series and the initial adjacency matrix into the wind power cluster power prediction model, obtain aggregated features through the wind power cluster power prediction model, and obtain wind power cluster power prediction data based on the aggregated features.

9. An electronic device, characterized in that, include: A memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the wind power cluster power prediction method considering spatiotemporal correlation as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the wind power cluster power prediction method considering spatiotemporal correlation as described in any one of claims 1 to 7.

Images