Wind power ultra-short-term prediction method and device based on graph convolutional neural network

Abstract

The invention discloses a wind power ultra-short-term prediction method and device based on a graph convolutional neural network, and relates to the field of wind power prediction.The method comprises the steps that wind power and NWP wind speed are obtained by collecting data of a wind power cluster; whether the trend of the wind power and the trend of the NWP wind speed are the same or not is judged, and processing is conducted according to the judgment result; if the trends are the same, carrying out fluctuation correlation analysis calculation on the wind power between the stations according to the fluctuation information between the stations to obtain the fluctuation correlation between the stations; constructing a relevance adjacency matrix based on the fluctuation relevance between stations; if the trends are different, the spacing distance between the stations is obtained through calculation; constructing a geographic position adjacency matrix based on the spacing distance between the stations; and according to the relevance adjacency matrix or the geographic position adjacency matrix, performing prediction processing through a dual-channel graph convolutional neural network, and outputting a wind power prediction result. According to the method, more refined modeling can be realized, and the ultra-short-term wind power cluster power prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a method and device for ultra-short-term wind power prediction based on a graph convolutional neural network. Background Art

[0002] In recent years, issues such as energy security, climate change, and environmental pollution have increasingly become constraints on sustainable, green economic and social development. Countries around the world are viewing the development of new energy sources, such as wind power and photovoltaics, as a key path to achieving green, low-carbon development. Wind power is a crucial source of clean energy today. The International Energy Agency (IEA) reports that wind power generation increased by a record 265TWh (14%) in 2022, exceeding 2,100TWh. To achieve net zero emissions by 2050, annual capacity additions will need to increase from approximately 75GW in 2022 to 350GW by 2030.

[0003] However, due to the inherent randomness, volatility, and intermittency of wind energy, current wind power generation suffers from a high degree of uncertainty. This shortcoming of wind energy poses a serious challenge to attempts to integrate large-scale wind power into modern power systems. Accurate wind power forecasting is one of the foundations for ensuring stable power system operation and promoting wind energy absorption. Ultra-short-term wind power forecasting (0-4 hours) is crucial for the safe and economic operation of power systems, such as real-time scheduling. With the increasing proportion of installed wind power capacity, there is an urgent need for accurate regional or cluster wind power forecasting to facilitate wind power grid integration and scientific scheduling planning.

[0004] Graph convolutional neural networks (GCNs) are currently widely used in ultra-short-term wind power forecasting. However, current GCN models typically consider wind farms as graph nodes, constructing an adjacency matrix based on geographic location or NWP correlations to represent spatiotemporal characteristics. This model fails to fully reflect the impact of wind power fluctuations on wind power output. Furthermore, current research still fails to fully explore the physical changes in wind power, ignoring the impact of wind power fluctuations on power forecasting.

[0005] Therefore, how to invent an ultra-short-term wind power forecasting method that measures the fluctuation characteristics of the power series while taking into account the spatiotemporal characteristics of each wind farm to achieve more refined modeling and improve the accuracy of ultra-short-term wind power cluster power forecasting has become an urgent problem to be solved. Summary of the Invention

[0006] To this end, the present invention provides a method and device for ultra-short-term wind power forecasting based on a graph convolutional neural network. Starting from the fluctuation information of each station in the cluster area, an adjacency matrix is constructed based on the fluctuation correlation. Based on the accuracy of NWP information, a trend switching mechanism is proposed. When the NWP trend information does not match the power information, a graph structure is constructed using an adjacency matrix based on geographic location, thereby achieving more accurate wind power forecasting and improving the accuracy of ultra-short-term wind power cluster power forecasting.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for ultra-short-term wind power prediction based on a graph convolutional neural network, comprising: By collecting data from wind power clusters, wind power and NWP wind speed are obtained; determining whether the trends of the wind power and the NWP wind speed are the same, and performing subsequent processing based on the determination result; If the trends are the same, the fluctuation correlation analysis and calculation of wind power between stations is performed based on the fluctuation information between stations to obtain the fluctuation correlation between stations; Constructing a correlation adjacency matrix based on the fluctuation correlation between the stations; If the trends are different, the separation distance between the stations is obtained by calculation; constructing a geographical location adjacency matrix based on the separation distances between stations; According to the correlation adjacency matrix or the geographical location adjacency matrix, prediction processing is performed through a dual-channel graph convolutional neural network to output a wind power prediction result.

[0008] As a preferred solution of the ultra-short-term wind power forecasting method based on graph convolutional neural network, the steps of judging whether the trends of the wind power and the NWP wind speed are the same are: Extract the power trend for the set hour before the forecast and the NWP wind speed trend for the set hour before the forecast; Connect the starting point, maximum point and end point of wind speed and power to form power trend triangle and NWP wind speed trend triangle; The slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle are respectively extracted; and it is determined whether the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle have the same positive and negative relationship; if the positive and negative relationships of the three sides are all the same, it is determined that the NWP wind speed trend effectively tracks the power trend and the trends are the same; if the positive and negative relationships of the three sides are not all the same, it is determined that the NWP wind speed trend cannot effectively track the power trend and the trends are different.

[0009] As a preferred solution of the ultra-short-term wind power forecasting method based on graph convolutional neural network, the calculation formula of the fluctuation correlation between stations is: ; Where, P a and P b are the power sequences of the two wind farms; c is the defined threshold; JUDGE is used to determine whether there is fluctuation correlation between the two sequences; FC is the calculation result of the fluctuation correlation; SHIFT is the characterization of its lag or lead phenomenon.

[0010] As a preferred solution of the ultra-short-term wind power forecasting method based on graph convolutional neural network, the expression of the correlation adjacency matrix is: ; Where a and b are two different wind farms; FC is the result of the fluctuation correlation calculation; std is the variance between the two series; c is the defined threshold; e is a natural constant; The expression of the geographical location adjacency matrix is: ; Where dist(i, j) is the geographical distance between two wind farms; ε is the defined threshold.

[0011] As a preferred solution for the ultra-short-term wind power forecasting method based on graph convolutional neural networks, the dual-channel graph convolutional neural network includes two GCN network modules, and uses each wind farm as a node to extract the power characteristics and NWP wind speed characteristics of the station respectively; The GCN network module expression for extracting power features is: ; Where, F power is the GCN network module for extracting power features; L is the length of the extracted time series of wind power; ∈R N*N is the adjacency matrix formed by the locations of wind farms; N is the number of wind farms; ∈R N*N for The degree matrix of d*h is the learnable convolution kernel parameter; is the activation function; x is the time node; y is the station node; V is the number of stations; i is the number of edges; j is the feature dimension; is the power sequence; The GCN network module expression for extracting NWP wind speed characteristics is: ; Where, F NWPis the GCN network module for extracting NWP wind speed characteristics; K is the length of the extracted time series of NWP wind speed; is the wind speed sequence.

[0012] As a preferred solution for the ultra-short-term wind power forecasting method based on graph convolutional neural network, the prediction accuracy of the dual-channel graph convolutional neural network is evaluated by the normalized root mean square error; the calculation formula of the normalized root mean square error is: ; Where, P Mi is the actual average power in period i; P Pi is the predicted average power in period i; N is the total daily assessment period; CAP is the wind farm station startup capacity.

[0013] The present invention also provides a wind power ultra-short-term prediction device based on a graph convolutional neural network, which is based on the above wind power ultra-short-term prediction method based on a graph convolutional neural network, including: The data acquisition module is used to obtain wind power and NWP wind speed by collecting data from the wind power cluster; a trend judgment module, configured to judge whether the trends of the wind power and the NWP wind speed are the same, and perform subsequent processing according to the judgment result; Inter-station fluctuation correlation calculation module is used to perform fluctuation correlation analysis and calculation on the wind power between stations based on the fluctuation information between stations if the trends are the same, and obtain the fluctuation correlation between stations; A correlation adjacency matrix construction module, configured to construct a correlation adjacency matrix based on the fluctuation correlation between the stations; A station interval distance calculation module is used to obtain the interval distance between stations by calculation if the trends are different; A geographic location adjacency matrix construction module, configured to construct a geographic location adjacency matrix based on the interval distances between stations; A dual-channel graph convolutional neural network prediction module is used to perform prediction processing through a dual-channel graph convolutional neural network based on the correlation adjacency matrix or the geographic location adjacency matrix, and output a wind power prediction result.

[0014] As a preferred solution of the ultra-short-term wind power forecasting device based on graph convolutional neural network, in the trend judgment module, the submodule for judging whether the trends of the wind power and the NWP wind speed are the same includes: The power and wind speed trend extraction submodule is used to extract the power trend of the set hour before the forecast and the NWP wind speed trend of the set hour before the forecast; The trend triangle generation submodule is used to connect the starting point, maximum point and end point of wind speed and power to form the power trend triangle and NWP wind speed trend triangle; The trend judgment submodule is used to respectively extract the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle; and judge whether the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle have the same positive and negative relationship; if the positive and negative relationships of the three sides are all the same, it is determined that the NWP wind speed trend effectively tracks the power trend and the trends are the same; if the positive and negative relationships of the three sides are not all the same, it is determined that the NWP wind speed trend cannot effectively track the power trend and the trends are different.

[0015] As a preferred solution of the ultra-short-term wind power prediction device based on graph convolutional neural network, in the inter-station fluctuation correlation calculation module, the calculation formula of the inter-station fluctuation correlation is: ; Where, P a and P b are the power sequences of the two wind farms; c is the defined threshold; JUDGE is used to determine whether there is fluctuation correlation between the two sequences; FC is the calculation result of the fluctuation correlation; SHIFT is the characterization of its lag or lead phenomenon.

[0016] As a preferred solution of the ultra-short-term wind power forecasting device based on graph convolutional neural network, in the correlation adjacency matrix construction module, the expression of the correlation adjacency matrix is: ; Where a and b are two different wind farms; FC is the result of the fluctuation correlation calculation; std is the variance between the two series; c is the defined threshold; e is a natural constant; In the geographic location adjacency matrix construction module, the expression of the geographic location adjacency matrix is: ; Where dist(i, j) is the geographical distance between two wind farms; ε is the defined threshold.

[0017] As a preferred solution of the ultra-short-term wind power prediction device based on graph convolutional neural network, in the dual-channel graph convolutional neural network prediction module, the dual-channel graph convolutional neural network includes two GCN network modules, and uses each wind farm as a node to extract the power characteristics and NWP wind speed characteristics of the station respectively; The GCN network module expression for extracting power features is: ; Where, Fpower is the GCN network module for extracting power features; L is the length of the extracted time series of wind power; ∈R N*N is the adjacency matrix formed by the locations of wind farms; N is the number of wind farms; ∈R N*N for The degree matrix of d*h is the learnable convolution kernel parameter; is the activation function; x is the time node; y is the station node; V is the number of stations; i is the number of edges; j is the feature dimension; is the power sequence; The GCN network module expression for extracting NWP wind speed characteristics is: ; Where, F NWP is the GCN network module for extracting NWP wind speed characteristics; K is the length of the extracted time series of NWP wind speed; is the wind speed sequence.

[0018] As a preferred solution of the ultra-short-term wind power prediction device based on graph convolutional neural network, in the dual-channel graph convolutional neural network prediction module, the prediction accuracy of the dual-channel graph convolutional neural network is evaluated by the normalized root mean square error; the calculation formula of the normalized root mean square error is: ; Where, P Mi is the actual average power in period i; P Pi is the predicted average power in period i; N is the total daily assessment period; CAP is the wind farm station startup capacity.

[0019] The present invention has the following advantages: the present invention obtains wind power and NWP wind speed by collecting data from a wind power cluster; determines whether the trends of the wind power and the NWP wind speed are the same, and performs subsequent processing based on the determination result; if the trends are the same, the fluctuation correlation analysis and calculation of the wind power between the stations is performed based on the fluctuation information between the stations to obtain the fluctuation correlation between the stations; a correlation adjacency matrix is constructed based on the fluctuation correlation between the stations; if the trends are different, the interval distance between the stations is obtained by calculation; a geographic location adjacency matrix is constructed based on the interval distance between the stations; and prediction processing is performed using a dual-channel graph convolutional neural network based on the correlation adjacency matrix or the geographic location adjacency matrix to output a wind power prediction result. The present invention takes the fluctuation information of each station in the cluster area into consideration, constructs an adjacency matrix based on fluctuation correlation, and proposes a trend switching mechanism based on the accuracy of NWP information. When the NWP trend information does not match the power information, a graph structure is constructed using an adjacency matrix based on geographic location, thereby achieving more accurate wind power prediction and improving the accuracy of ultra-short-term wind power cluster power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can, without inventive effort, derive other implementation drawings based on the provided drawings.

[0021] The structures, proportions, sizes, etc. illustrated in this specification are intended solely to complement the contents disclosed herein and to facilitate understanding and reading by persons skilled in the art. They are not intended to limit the conditions under which the present invention may be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportions, or adjustments in sizes, without affecting the efficacy and objectives of the present invention, shall remain within the scope of the technical contents disclosed herein.

[0022] Figure 1 This is a flow chart of the ultra-short-term wind power prediction method based on graph convolutional neural network provided in Example 1 of the present invention; Figure 2 This is a schematic diagram of a specific implementation process of the method for ultra-short-term wind power prediction based on graph convolutional neural network provided in Example 1 of the present invention; Figure 3 This is a schematic diagram of different cluster prediction curves in a possible embodiment provided in Example 1 of the present invention; Figure 4This is a prediction diagram of various prediction methods for wind power cluster A in a possible embodiment provided in Example 1 of the present invention; Figure 5 This is a schematic diagram of RMSE trend changes under various prediction steps in a possible embodiment provided in Example 1 of the present invention; Figure 6 A schematic diagram of prediction curves of various graph construction methods at different output levels in a possible embodiment provided in Example 1 of the present invention; Figure 7 A schematic diagram of a prediction curve for determining whether a trend switching mechanism is adopted in a possible embodiment provided in Example 1 of the present invention; Figure 8 This is a schematic diagram of the architecture of the ultra-short-term wind power prediction device based on graph convolutional neural network provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0023] The following describes the implementation of the present invention using specific embodiments. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. Obviously, the embodiments described are only a portion of the present invention, not all of it. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0024] Example 1

[0025] See also Figure 1 and Figure 2 Embodiment 1 of the present invention provides a method for ultra-short-term wind power prediction based on a graph convolutional neural network, comprising the following steps: S1. Obtain wind power and NWP wind speed by collecting data from wind power clusters; S2. Determine whether the trends of the wind power and the NWP wind speed are the same, and perform subsequent processing based on the determination result; S3. If the trends are the same, then the fluctuation correlation analysis and calculation of the wind power between the stations is performed based on the fluctuation information between the stations to obtain the fluctuation correlation between the stations; S4. constructing a correlation adjacency matrix based on the fluctuation correlation between the stations; S5. If the trends are different, the interval distance between the stations is obtained by calculation; S6. Constructing a geographical location adjacency matrix based on the interval distances between stations; S7. Perform prediction processing using a dual-channel graph convolutional neural network based on the correlation adjacency matrix or the geographic location adjacency matrix, and output a wind power prediction result.

[0026] In this embodiment, in step S1, wind power and NWP wind speed are obtained by collecting data of the wind power cluster; Specifically, wind power data is obtained from the data acquisition and monitoring control system; NWP wind speed is obtained through the numerical weather forecast provided by the meteorological center.

[0027] In this embodiment, in step S2, it is determined whether the trends of the wind power and the NWP wind speed are the same, and subsequent processing is performed according to the determination result; Specifically, the steps of determining whether the trends of the wind power and the NWP wind speed are the same are as follows: S21, extracting the power trend for a set hour before the forecast and the NWP wind speed trend for the set hour before the forecast; Specifically, the power and NWP wind speed trends of the previous 8 hours were extracted; S22, connecting the starting point, maximum value, and end point of wind speed and power to form a power trend triangle and a NWP wind speed trend triangle; S23. Extract the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle respectively; determine whether the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle have the same positive and negative relationship; if the positive and negative relationships of the three sides are all the same, it is determined that the NWP wind speed trend effectively tracks the power trend and the trends are the same; if the positive and negative relationships of the three sides are not all the same, it is determined that the NWP wind speed trend cannot effectively track the power trend and the trends are different.

[0028] Specifically, the slopes of the three sides of the NWP wind speed trend triangle, W1, W2, and W3, and the three sides of the power trend triangle, P1, P2, and P3, are extracted. The slopes of W1, W2, and W3 are then determined to be in the same positive or negative relationship as the slopes of P1, P2, and P3. Only when all the positive and negative relationships are the same can the NWP wind speed trend be determined to be effectively tracking the power trend. In this case, an adjacency matrix based on fluctuation correlation is used for calculation. If the positive and negative relationships of any corresponding slopes are different, the NWP wind speed trend is determined to be ineffective in expressing the corresponding trend. In this case, an adjacency matrix based on geographic location information is used for calculation.

[0029] In this embodiment, in step S3, if the trends are the same, the fluctuation correlation analysis and calculation of the wind power between the stations is performed based on the fluctuation information between the stations to obtain the fluctuation correlation between the stations; Specifically, within the same cluster, wind power output exhibits similar fluctuation trends. Wind power output at each site also fluctuates to similar degrees over different time periods, with adjacent sites exhibiting leading or lagging fluctuations. To characterize the correlation between fluctuations across sites, this paper proposes a calculation method for fluctuation correlation that effectively characterizes these correlations.

[0030] First, the two power series (or NWP wind speed series) are defined as: P a =[P a0 , P a1 ,……, P an ]; P b =[P b0 , P b1 ,…… P bn ]; Where n is the sequence length; P a and P b is the power sequence of the two wind farms.

[0031] To measure its correlation, the sequence P b Keep it fixed, and slide the other sequence along the time axis with s as the sliding window, where s∈[-n,n], and the vector after displacement is defined as P as , its expression: ; Then calculate P as With P b The sum of the inner products G(P as , P b ) and the expression of correlation CCV is: ; ; By enumerating all possible values of s, we can obtain a cross-correlation value vector with a sequence length of 2n−1. Let the minimum and maximum values of the vector be minCCV and maxCCV, corresponding to the shift values k1 and k2 respectively. Then P a and P b The final cross-correlation value, namely the volatility correlation, can be defined as follows: ; ; ; From the above formula, we can find that the value of FC is between [-1,1]. The closer FC is to 1 or -1, the stronger the correlation between the fluctuations of the two power series. When FC is a positive value, it means that they fluctuate at the same time. When FC is a negative value, it means that there is an advance or lag in the volatility between the power series.

[0032] According to the above formula, the output vector of the fluctuation correlation calculation (FCC) is defined as: ; Where, P a and P b is the power sequence of the two wind farms; c is the defined threshold; JUDGE is used to judge whether there is fluctuation correlation between the two sequences. The calculated FC value is compared with the threshold c. If the FC value is greater than the c value, it is determined that there is fluctuation correlation, otherwise there is no fluctuation correlation; FC is the result of the fluctuation correlation calculation; SHIFT is the representation of its lag or lead phenomenon. The time sequence can be determined by P a The shift value is determined by the value of P; if it is zero (ie no shift), then P a ↔P b ; If it is negative (i.e. X is offset to the left), then P a →P b Otherwise, P b →P a .

[0033] In this embodiment, in step S4, a correlation adjacency matrix is constructed based on the fluctuation correlation between the stations; Specifically, the expression of the correlation adjacency matrix is: ; Where a and b are two different wind farms; FC is the result of the fluctuation correlation calculation; std is the variance between the two sequences; c is a defined threshold, which is 0.5 in this embodiment; and e is a natural constant. If FC is less than the threshold, it is assessed that there is no fluctuation correlation, or link, between the two wind farms to ensure the sparsity of the adjacency matrix.

[0034] In this embodiment, in step S5, if the trends are different, the interval distance between the stations is obtained by calculation; In this embodiment, in step S6, a geographical location adjacency matrix is constructed based on the interval distances between the stations; Specifically, the expression of the geographic location adjacency matrix is: ; Where dist(i, j) is the geographical distance between the two wind farms. The elements in the adjacency matrix are defined using a Gaussian kernel. To maintain the sparsity of the adjacency matrix, elements less than a threshold ε are set to 0. ε is a defined threshold, defined as less than 1 / 2 of the average distance between all sites.

[0035] In this embodiment, in step S7, prediction processing is performed using a dual-channel graph convolutional neural network based on the correlation adjacency matrix or the geographic location adjacency matrix to output a wind power prediction result.

[0036] In order to effectively extract power features and NWP features, a dual-channel graph convolutional neural network with two GCN network modules is designed; The dual-channel graph convolutional neural network uses each wind farm as a node to extract the power characteristics and NWP wind speed characteristics of the station respectively; The GCN network module expression for extracting power features is: ; Where, F power is the GCN network module for extracting power features; L is the length of the extracted time series of wind power; ∈R N*N is the adjacency matrix formed by the locations of wind farms; N is the number of wind farms; ∈R N*N for The degree matrix of d*h is the learnable convolution kernel parameter; is the activation function; x is the time node; y is the station node; V is the number of stations; i is the number of edges; j is the feature dimension; is the power sequence; The GCN network module expression for extracting NWP wind speed characteristics is: ; Where, F NWP is the GCN network module for extracting NWP wind speed characteristics; K is the length of the extracted time series of NWP wind speed; is the wind speed sequence; the spatial characteristics and fluctuation correlation are given by Adjustment is performed and W is used to extract timing characteristics.

[0037] In this example, after extracting spatial features using two GCN modules, multimodal learning is used to link the spatiotemporal features of wind power and NWP wind speed, leveraging the advantages of data-driven model prediction. A fully connected layer fuses the spatiotemporal features of wind power and NWP wind speed and then maps them, defining each wind farm in the cluster as a specific layer.

[0038] In this embodiment, the prediction accuracy of the dual-channel graph convolutional neural network is evaluated by the normalized root mean square error; RMSE can reflect the slight differences in prediction results between different targets by calculating the prediction deviation, which is particularly suitable for evaluation scenarios with multiple targets and small differences.

[0039] The calculation formula of the normalized root mean square error is: ; Where, P Mi is the actual average power in period i; P Pi is the predicted average power in period i; N is the total daily assessment period; CAP is the wind farm station startup capacity.

[0040] In a possible embodiment, an example of improving the power prediction of a specific wind power cluster is as follows:

[0041] In order to maximize the accuracy of wind power cluster power prediction, three typical wind power cluster data sets were used to verify the effectiveness of the present invention. The data set contains wind power measurement data and NWP data of three real wind power clusters, including meteorological factors (wind speed) (updated every 24 hours for a 24-hour advance forecast). The time interval is 15 minutes. The present invention uses wind power data and numerical forecast data to perform ultra-short-term multi-step prediction of wind power in the next 4 hours. The three selected wind power clusters are located in Jilin Province, Gansu Province, and Yunnan Province, China, with total installed capacities of 5820MW, 14286.6MW, and 8089.2MW, respectively. There are 180 days from 00:00 on January 1, 2021 to 23:45 on June 30, 2021, with a time interval of 15 minutes. The first five months are used as training sets, and the sixth month is used as a test set.

[0042] To verify the effectiveness of the present invention in cluster deterministic prediction, several methods were selected for comparison. These included the cumulative method (AC, which builds prediction models for all stations within a cluster and then adds the predicted power of each station to form the cluster power. This embodiment uses an LSTM model for cumulative cluster prediction), the holistic method (HC, which sums the power of all stations within the cluster and averages the NWP of each station to construct a feature, directly modeling the overall power of the entire site), the statistical upscaling method (SU, which selects one or more reference wind farms with power output similar to that of the entire region, and converts the regional power prediction result into a ratio of the prediction results of the reference wind farms), the cluster partitioning method (CD, which divides the cluster into multiple subclusters according to a certain rule, builds a prediction model for each subcluster, and sums the power prediction results of the subclusters to form the overall cluster power. This experiment uses the Birch clustering method for cluster partitioning), the standard GCN, and the dual-channel GCN (which only considers geographic distance).

[0043] Table 1 RMSE of the 16th step of power prediction for each wind power cluster

[0044] As shown in Table 1, the RMSE of each prediction method in the 16th step of the prediction of three wind power clusters shows the prediction effect of several cluster prediction methods. Among them, the RMSE of the prediction method proposed in the present invention is the smallest, which shows the effectiveness of the cluster prediction of the present invention. The prediction effect of the SU method is the worst, because the selection of the reference wind farm affects the prediction accuracy. The accumulation method also leads to a decrease in the overall prediction accuracy due to the poor prediction level of some wind farms. By clustering, better prediction results are obtained, but due to the distribution of sites, individual sites will always be treated as a sub-cluster alone, which will also affect the prediction accuracy. The overall method and GCN have better prediction effects than other methods, but because they use cluster power as the modeling target, they lose a certain amount of fluctuation information. Through DC-GCN and the proposed method, while taking into account the geographical location, the fluctuation information is considered to varying degrees, and it has the best prediction effect. In particular, the present invention comprehensively considers the fluctuation correlation and location distribution. Compared with other methods, the prediction accuracy in three stations is improved by an average of 1.34%, 1.62% and 2.07%, which verifies the effectiveness of the present invention.

[0045] The prediction results of each cluster are as follows Figure 3 As shown, the prediction curve of the present invention can better track the change trend of the actual power curve and has better prediction effect in different clusters.

[0046] like Figure 4 The figure shows the prediction results of various prediction methods for wind power cluster A. Compared with other methods, the prediction curve of the present invention can better track the changing trend of the actual power curve and more effectively fit the peak-to-valley changes, thereby improving prediction accuracy. At the peak value, the present invention better tracks the changing trend of the peak value and has a lower numerical error level, effectively improving the reliability and accuracy of cluster prediction.

[0047] Table 2 RMSE of the 16th step of each prediction method at different output levels

[0048] As shown in Table 2, the RMSE of this method in the region above 3000MW is 3.28% lower than that of other prediction methods. This method can more effectively predict high-output scenarios in the ultra-short term, reducing the harm caused by high output to the power system. Furthermore, during low-output phases, this method is more effective than other methods in predicting low-output trends, facilitating power system scheduling and promoting wind power consumption.

[0049] At the same time, this experiment also verified the prediction effect at different time steps, and its RMSE is shown in Table 3: Table 3 RMSE at each time step

[0050] As can be seen from Table 3, except for the first step, i.e., the 15th minute prediction, the present invention achieved the best prediction accuracy at each prediction step. In the 2-3h and 3-4h predictions, the present invention more effectively reduced the prediction error caused by the increase in the number of prediction steps, thereby effectively improving the prediction accuracy. AC and CD have almost the same effect, both of which suffer from prediction errors due to their inability to consider the predictions of individual stations. HC is unable to effectively extract spatial features and fluctuation information. In contrast, the present invention more effectively extracts fluctuation information, achieving the best prediction accuracy.

[0051] like Figure 5 The RMSE for each prediction method at each prediction time step in the three clusters is shown. As the number of prediction steps increases, the prediction RMSE of each method increases. Compared with other methods, the present invention effectively reduces the error caused by the increase in the number of prediction steps in the 3-4 hour time period, slowing the upward trend of the RMSE and effectively improving prediction accuracy.

[0052] To verify the effectiveness of the model that takes into account fluctuation correlation, we take wind farm cluster A as an example and compare several graph adjacency matrix (graph) construction methods: Method A: Single adjacency matrix method, only considering geographical distance.

[0053] Method B: Single adjacency matrix method, only considering correlation (Method 1: Correlation calculation of power between stations, using Pearson correlation coefficient; Method 2: Correlation calculation of NWP between stations, using Pearson correlation coefficient).

[0054] Method C: dual adjacency matrix approach, power correlation and NWP correlation.

[0055] Method D: Dual adjacency matrix approach, taking both geographic location and correlation into account (the power channel uses geographic location information, and the NWP channel uses the correlation between the NWPs of each station).

[0056] Method E: Double adjacency matrix method, considering fluctuation correlation and correlation (Method 1: fluctuation correlation and NWP correlation between the power of each station; Method 2: fluctuation correlation and power correlation between the NWP of each station).

[0057] Table 4 RMSE under different graph construction methods

[0058] The prediction result indicators are shown in Table 4, which shows the RMSE of predictions of various graph construction methods. Compared with the graph construction method that only uses a single adjacency matrix, the dual adjacency matrix method is more suitable for the GCN network with dual feature channels, and generally achieves relatively good prediction results. Compared with the graph construction method that only considers geographic location information, the construction method that considers the correlation between NWP and power effectively improves the prediction accuracy. Only considering geographic location information, only extracting spatial characteristics ignores the capture of NWP and power characteristics. Therefore, the construction method that takes into account both spatial characteristics and sequence characteristics is more conducive to the improvement of GCN prediction accuracy. The present invention more effectively improves the prediction accuracy based on the fluctuation correlation of power or NWP extraction between stations. Compared with the construction method that only considers correlation, it effectively takes into account the sequence fluctuation information, thereby improving the prediction accuracy. Among them, the present invention simultaneously considers the correlation of power and NWP fluctuations between stations to achieve the best prediction accuracy. Compared with other methods, the prediction RMSE is reduced by an average of 0.5%, which more effectively improves the prediction accuracy.

[0059] like Figure 6 The following are examples of prediction curves constructed using various adjacency matrix construction methods. The present invention uses a graph construction method that takes into account fluctuation correlation to more effectively track the changing trend of the actual curve, such as Figure 6 In the high and low output scenarios shown in , the present invention can better track the changing trends at the peaks and valleys in both high and low output scenarios, because the fluctuation information is more effectively extracted according to the FCC correlation coefficient, thereby better taking into account the physical change process, fitting the peak and valley trends, and improving the prediction accuracy.

[0060] Therefore, the graph structure established by the present invention based on fluctuation correlation is a better graph construction method, which enables GCN to fit fluctuation information more effectively, thereby improving prediction accuracy.

[0061] Due to the accuracy of NWP, the use of the present invention may cause the prediction model to learn incorrect trend information and establish an incorrect mapping relationship, which makes it difficult for the predicted power to effectively track the changing trend of the actual power. Therefore, in order to alleviate the problems caused by the accuracy of NWP, the present invention also adopts a trend-aware mechanism switching. When it is determined that the current predicted trend cannot be tracked, the feature information of NWP is not considered in the graph construction of the feature extraction channel of NWP, and the prediction is performed using the geographic location as the adjacency matrix. To verify the effectiveness of the trend switching mechanism of the present invention, taking wind power cluster A as an example, the prediction effects of using the trend switching mechanism and not using the trend switching mechanism are compared.

[0062] The overall prediction effect is as follows Figure 7As shown, it clearly shows that compared with the prediction method that does not consider the trend switching mechanism, the present invention can more effectively fit the peak and valley characteristics, alleviate the mismapping caused by the accuracy of NWP, and enable the prediction curve to better fit the changing trend of actual power. Based on the above analysis and comparison examples, it can be seen that the trend switching mechanism proposed by the present invention has certain advantages. It can effectively alleviate the problem that the prediction curve trend is difficult to track the actual power due to NWP inaccuracy, thereby improving the overall prediction accuracy.

[0063] In summary, the present invention acquires wind power and NWP wind speed by collecting data from a wind power cluster; determines whether the trends of the wind power and the NWP wind speed are the same, and performs subsequent processing based on the determination result; if the trends are the same, the fluctuation correlation analysis and calculation of the wind power between the stations is performed based on the fluctuation information between the stations to obtain the fluctuation correlation between the stations; a correlation adjacency matrix is constructed based on the fluctuation correlation between the stations; if the trends are different, the interval distance between the stations is obtained by calculation; a geographic location adjacency matrix is constructed based on the interval distance between the stations; and prediction processing is performed using a dual-channel graph convolutional neural network based on the correlation adjacency matrix or the geographic location adjacency matrix to output a wind power prediction result. The present invention approaches the problem from the perspective of fluctuation information of each station in the cluster area, constructs an adjacency matrix based on fluctuation correlation, and proposes a trend switching mechanism based on the accuracy of NWP information. When the NWP trend information does not match the power information, a graph structure is constructed based on the geographic location adjacency matrix, thereby achieving more accurate wind power prediction and improving the accuracy of ultra-short-term wind power cluster power prediction.

[0064] It should be noted that the method of the embodiments of the present disclosure can be performed by a single device, such as a computer or server. The method of the embodiments of the present disclosure can also be applied in a distributed scenario, where multiple devices cooperate to perform the method. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiments of the present disclosure, and the multiple devices will interact with each other to complete the method.

[0065] It should be noted that the above description is limited to some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0066] Example 2

[0067] See also Figure 8 Embodiment 2 of the present invention further provides an ultra-short-term wind power prediction device based on a graph convolutional neural network, comprising: Data acquisition module 001 is used to obtain wind power and NWP wind speed by collecting data from wind power clusters; The trend judgment module 002 is used to judge whether the trends of the wind power and the NWP wind speed are the same, and perform subsequent processing according to the judgment result; Inter-station fluctuation correlation calculation module 003 is used to perform fluctuation correlation analysis and calculation on the wind power between the stations based on the fluctuation information between the stations if the trends are the same, to obtain the fluctuation correlation between the stations; A correlation adjacency matrix construction module 004 is used to construct a correlation adjacency matrix based on the fluctuation correlation between the stations; The station interval distance calculation module 005 is used to obtain the interval distance between stations by calculation if the trends are different; A geographic location adjacency matrix construction module 006 is configured to construct a geographic location adjacency matrix based on the interval distances between stations; The dual-channel graph convolutional neural network prediction module 007 is used to perform prediction processing through a dual-channel graph convolutional neural network according to the correlation adjacency matrix or the geographical location adjacency matrix, and output a wind power prediction result.

[0068] In this embodiment, in the trend judgment module 002, the submodule for judging whether the trends of the wind power and the NWP wind speed are the same includes: The power and wind speed trend extraction submodule 021 is used to extract the power trend of the set hour before the forecast and the NWP wind speed trend of the set hour before the forecast; The trend triangle generation submodule 022 is used to connect the starting point, maximum point and end point of wind speed and power to form a power trend triangle and an NWP wind speed trend triangle; The trend determination submodule 023 is configured to extract the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle, respectively; and determine whether the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle have the same positive and negative relationship. If the positive and negative relationships of the three sides are all the same, it is determined that the NWP wind speed trend effectively tracks the power trend and the trends are the same. If the positive and negative relationships of the three sides are not all the same, it is determined that the NWP wind speed trend cannot effectively track the power trend and the trends are different.

[0069] In this embodiment, in the inter-station fluctuation correlation calculation module 003, the calculation formula of the inter-station fluctuation correlation is: ; Where, P a and P b are the power sequences of the two wind farms; c is the defined threshold; JUDGE is used to determine whether there is fluctuation correlation between the two sequences; FC is the calculation result of the fluctuation correlation; SHIFT is the characterization of its lag or lead phenomenon.

[0070] In this embodiment, in the relevance adjacency matrix construction module 004, the expression of the relevance adjacency matrix is: ; Where a and b are two different wind farms; FC is the result of the fluctuation correlation calculation; std is the variance between the two series; c is the defined threshold; e is a natural constant; In the geographic location adjacency matrix construction module 006, the expression of the geographic location adjacency matrix is: ; Where dist(i, j) is the geographical distance between two wind farms; ε is the defined threshold.

[0071] In this embodiment, in the dual-channel graph convolutional neural network prediction module 007, the dual-channel graph convolutional neural network includes two GCN network modules, and uses each wind farm as a node to extract the power characteristics and NWP wind speed characteristics of the station respectively; The GCN network module expression for extracting power features is: ; Where, F power is the GCN network module for extracting power features; L is the length of the extracted time series of wind power; ∈R N*N is the adjacency matrix formed by the locations of wind farms; N is the number of wind farms; ∈R N*N for The degree matrix of d*h is the learnable convolution kernel parameter; is the activation function; x is the time node; y is the station node; V is the number of stations; i is the number of edges; j is the feature dimension; is the power sequence.

[0072] The GCN network module expression for extracting NWP wind speed characteristics is: ; Where, F NWP is the GCN network module for extracting NWP wind speed characteristics; K is the length of the extracted time series of NWP wind speed; is the wind speed sequence.

[0073] In this embodiment, in the dual-channel graph convolutional neural network prediction module 007, the prediction accuracy of the dual-channel graph convolutional neural network is evaluated by the normalized root mean square error; the calculation formula of the normalized root mean square error is: ; Where, P Mi is the actual average power in period i; P Pi is the predicted average power in period i; N is the total daily assessment period; CAP is the wind farm station startup capacity.

[0074] It should be noted that the information interaction, execution process, etc. between the modules of the above-mentioned system are based on the same concept as the method embodiment in Example 1 of the present application, and the technical effects they bring are the same as those of the method embodiment of the present application. For specific contents, please refer to the description in the method embodiment shown above in the present application, and no further details will be given here.

[0075] Example 3

[0076] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, in which the program code of the ultra-short-term wind power prediction method based on graph convolutional neural network is stored. The program code includes instructions for executing the ultra-short-term wind power prediction method based on graph convolutional neural network of embodiment 1 or any possible implementation thereof.

[0077] Computer-readable storage media can be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that includes one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0078] Example 4

[0079] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor; The processor and the memory communicate with each other through a bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the ultra-short-term wind power prediction method based on graph convolutional neural network of Example 1 or any possible implementation method thereof.

[0080] Specifically, the processor can be implemented by hardware or by software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor, which is implemented by reading software code stored in a memory. The memory can be integrated into the processor or located outside the processor and exist independently.

[0081] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode.

[0082] Obviously, those skilled in the art will appreciate that the various modules or steps of the present invention described above can be implemented using a general-purpose computing system. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Alternatively, they can be implemented using program code executable by a computing system, and thus, they can be stored in a storage system and executed by the computing system. In some cases, the steps shown or described herein can be performed in a different order than that shown, or they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0083] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. An ultra-short-term wind power forecasting method based on graph convolutional neural network is characterized by: include: By collecting data from wind power clusters, wind power and NWP wind speed are obtained; determining whether the trends of the wind power and the NWP wind speed are the same, and performing subsequent processing based on the determination result; If the trends are the same, the fluctuation correlation analysis and calculation of wind power between stations is performed based on the fluctuation information between stations to obtain the fluctuation correlation between stations; Constructing a correlation adjacency matrix based on the fluctuation correlation between the stations; If the trends are different, the separation distance between the stations is obtained by calculation; constructing a geographical location adjacency matrix based on the separation distances between stations; According to the correlation adjacency matrix or the geographical location adjacency matrix, prediction processing is performed through a dual-channel graph convolutional neural network to output a wind power prediction result.

2. The method for ultra-short-term wind power prediction based on graph convolutional neural network according to claim 1 is characterized in that: The steps of determining whether the trends of the wind power and the NWP wind speed are the same are as follows: Extract the power trend for the set hour before the forecast and the NWP wind speed trend for the set hour before the forecast; Connect the starting point, maximum point and end point of wind speed and power to form power trend triangle and NWP wind speed trend triangle; Extracting the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle respectively; determining whether the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle have the same positive and negative relationship; if the positive and negative relationships of the three sides are all the same, determining that the NWP wind speed trend effectively tracks the power trend and the trends are the same; If the positive and negative relationships of the three edges are not all the same, it is determined that the NWP wind speed trend cannot effectively track the power trend and the trends are different.

3. The method for ultra-short-term wind power prediction based on graph convolutional neural network according to claim 2 is characterized in that: The calculation formula for the fluctuation correlation between stations is: ； Where, P a and P b are the power sequences of the two wind farms; c is the defined threshold; JUDGE is used to determine whether there is fluctuation correlation between the two sequences; FC is the calculation result of the fluctuation correlation; SHIFT is the characterization of its lag or lead phenomenon.

4. The method for ultra-short-term wind power prediction based on graph convolutional neural network according to claim 3 is characterized in that: The expression of the association adjacency matrix is: ； Where a and b are two different wind farms; FC is the result of the volatility correlation calculation; std is the variance between the two series; c is the defined threshold; e is the natural constant; The expression of the geographical location adjacency matrix is: ； Where dist(i, j) is the geographical distance between two wind farms; ε is the defined threshold.

5. The method for ultra-short-term wind power prediction based on graph convolutional neural network according to claim 4 is characterized in that: The dual-channel graph convolutional neural network includes two GCN network modules, and uses each wind farm as a node to extract the power characteristics and NWP wind speed characteristics of the station respectively; The GCN network module expression for extracting power features is: ； Where, F power is the GCN network module for extracting power features; L is the length of the extracted time series of wind power; ∈R N *N is the adjacency matrix formed by the locations of wind farms; N is the number of wind farms; ∈R N*N for The degree matrix of d*h is the learnable convolution kernel parameter; is the activation function; x is the time node; y is the station node; V is the number of stations; i is the number of edges; j is the feature dimension; is the power sequence; The GCN network module expression for extracting NWP wind speed characteristics is: ； Where, F NWP is the GCN network module for extracting NWP wind speed characteristics; K is the length of the extracted time series of NWP wind speed; is the wind speed sequence.

6. The method for ultra-short-term wind power prediction based on graph convolutional neural network according to claim 5 is characterized in that: The prediction accuracy of the dual-channel graph convolutional neural network is evaluated by the normalized root mean square error; the calculation formula of the normalized root mean square error is: ； Where, P Mi is the actual average power in period i; P Pi is the predicted average power in period i; N is the total daily assessment period; CAP is the wind farm station startup capacity.

7. A wind power ultra-short-term prediction device based on a graph convolutional neural network, adopting the wind power ultra-short-term prediction method based on a graph convolutional neural network according to any one of claims 1 to 6, characterized in that: include: The data acquisition module is used to obtain wind power and NWP wind speed by collecting data from the wind power cluster; a trend judgment module, configured to judge whether the trends of the wind power and the NWP wind speed are the same, and perform subsequent processing according to the judgment result; Inter-station fluctuation correlation calculation module is used to perform fluctuation correlation analysis and calculation on the wind power between stations based on the fluctuation information between stations if the trends are the same, and obtain the fluctuation correlation between stations; A correlation adjacency matrix construction module, configured to construct a correlation adjacency matrix based on the fluctuation correlation between the stations; A station interval distance calculation module is used to obtain the interval distance between stations by calculation if the trends are different; A geographic location adjacency matrix construction module, configured to construct a geographic location adjacency matrix based on the interval distances between stations; A dual-channel graph convolutional neural network prediction module is used to perform prediction processing through a dual-channel graph convolutional neural network based on the correlation adjacency matrix or the geographic location adjacency matrix, and output a wind power prediction result.

8. The ultra-short-term wind power prediction device based on graph convolutional neural network according to claim 7 is characterized in that: In the trend judgment module, the submodule for judging whether the trends of the wind power and the NWP wind speed are the same includes: The power and wind speed trend extraction submodule is used to extract the power trend of the set hour before the forecast and the NWP wind speed trend of the set hour before the forecast; The trend triangle generation submodule is used to connect the starting point, maximum point and end point of wind speed and power to form the power trend triangle and NWP wind speed trend triangle; The trend judgment submodule is used to respectively extract the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle; and judge whether the slopes of the three sides of the power trend triangle and the slopes of the three sides of the NWP wind speed trend triangle have the same positive and negative relationship; if the positive and negative relationships of the three sides are all the same, it is determined that the NWP wind speed trend effectively tracks the power trend and the trends are the same; if the positive and negative relationships of the three sides are not all the same, it is determined that the NWP wind speed trend cannot effectively track the power trend and the trends are different.

9. The ultra-short-term wind power prediction device based on graph convolutional neural network according to claim 8, characterized in that: In the inter-station fluctuation correlation calculation module, the calculation formula for the inter-station fluctuation correlation is: ； Where, P a and P b are the power sequences of the two wind farms; c is the defined threshold; JUDGE is used to determine whether there is fluctuation correlation between the two sequences; FC is the calculation result of the fluctuation correlation; SHIFT is the characterization of its lag or lead phenomenon.

10. The ultra-short-term wind power prediction device based on graph convolutional neural network according to claim 9, characterized in that: In the relevance adjacency matrix construction module, the expression of the relevance adjacency matrix is: ； Where a and b are two different wind farms; FC is the result of the volatility correlation calculation; std is the variance between the two series; c is the defined threshold; e is the natural constant; In the geographic location adjacency matrix construction module, the expression of the geographic location adjacency matrix is: ； Where dist(i, j) is the geographical distance between two wind farms; ε is the defined threshold.

Images