Wind power cluster power prediction method and device, equipment and storage medium

Abstract

The invention provides a wind power cluster power prediction method and device, equipment and a storage medium. Relates to the technical field of wind power prediction. According to the method, an interpretable space-time variable diagram is constructed through wind power cluster power coding, and space-time coupling characteristics between stations are extracted and represented in a Fourier space. In a prediction stage, a sequence decomposition module is constructed to decouple a trend component and a residual component, a dual-channel decomposition prediction architecture fusing a recurrent neural network and graph convolution is proposed to learn a decoupling feature component, and unified prediction is realized through a fusion mechanism. In the training stage, a composite loss function formed by a plurality of angle errors used for capturing amplitude, trend and error mean value is constructed to guide model parameter updating, and the capturing capability of cluster power change is improved. Example verification results show that the ultra-short-term multi-step prediction precision of the wind power can be effectively improved, and a favorable reference is provided for a dispatching center to respond to the power demand in real time to adjust a power generation plan.

Description

Technical Field

[0001] The present application relates to the technical field of wind power prediction, and in particular to a method, apparatus, device and storage medium for predicting wind power cluster power. Background Art

[0002] As wind power grid penetration increases annually, the demand for wind power generation to ensure stable grid operation is becoming increasingly stringent. However, due to the volatility and intermittent nature of wind energy, the wind power integration process is subject to uncertainty, creating risks to the safe operation of the power system. Fluctuations in wind farm power are a major factor jeopardizing grid safety during wind farm integration. To better regulate power allocation and promote the absorption of wind power, more accurate wind farm power forecasts are required to provide dispatching departments with pre-emptive power generation planning.

[0003] Wind power generation relies on wind energy resources, which means that the power of wind farm cluster sites has the same spatiotemporal correlation as wind energy. Therefore, some researchers have constructed a two-dimensional matrix based on the site's geographic location and used a convolutional network to extract the spatial characteristics of wind farms. This is then input into a long-short-term memory network to learn temporal features for wind power prediction. Other researchers have proposed that there are differences in spatial distance and wind resources between wind farm sites. They propose using correlation coefficients to represent the spatiotemporal correlation between sites, constructing a weighted feature matrix that is input into a convolutional network to extract spatial features, and then using a long-short-term memory network to learn and predict temporal features. However, the geographic locations of wind farms are not regular and orderly, so constructing a two-dimensional matrix and then using a two-dimensional convolutional network does not effectively represent the spatial characteristics of wind farms. In recent years, graph convolution has attracted increasing attention due to its ability to handle arbitrary topological structures. Existing techniques construct the adjacency matrix of a graph convolutional network based on the wind farm's geographic location, the Pearson coefficient, the Spearman coefficient, and other methods. This allows for spatial modeling and representation of wind farm clusters, improving the accuracy of wind farm cluster power prediction. A major trend in current wind power forecasting research is to use graph neural networks to define and extract the complex spatial correlations of wind farms, combined with time series models to capture temporal dependencies. In terms of model architecture, the mainstream model is a tandem model that considers space first and time later. There are also studies exploring the alternating fusion of temporal convolutional networks between graph neural network layers or combining different time series models. Attention mechanisms and multi-task learning are commonly used optimization methods to improve the model's focus on key information and tasks. Although these methods use the spatiotemporal correlations of wind power clusters for analysis and prediction, they all perform characterization and modeling in the spatial and temporal dimensions separately before fusion. The failure to achieve unified modeling and representation in both dimensions and the inability to effectively utilize the power generation size of field stations in wind power clusters is a result of the simultaneous effects of both time and space.

[0004] Existing technologies have made significant contributions to improving the accuracy of wind power cluster power forecasting by focusing on the spatiotemporal correlations of wind power clusters. However, they lack consideration of wind power fluctuation analysis. Based on the evolution of wind power, wind power can be divided into different fluctuation patterns and prediction models can be established for each. Related research focuses on using data similarity matching techniques (such as shape-position similarity, multi-index combination, and dynamic causal relationships) to improve the performance of wind power forecasting or unequal-length data clustering by identifying and matching fluctuation patterns or key information in historical data. However, most of these methods analyze power fluctuations from a single perspective, failing to analyze the fluctuation characteristics of the wind power cluster as a whole. During the wind power grid connection process, when the power of individual stations is aggregated into the power of the wind power cluster, its power fluctuation characteristics are smoothed through cumulative effects, and its variation patterns are relatively stable. Therefore, increasing the predictability of cluster power and incorporating cluster power fluctuation learning and modeling into the forecasting process are essential to improving forecast accuracy.

[0005] In summary, there are two current issues with wind farm power forecasting. 1) While graph convolution-based modeling methods improve forecasting performance, they fail to provide unified modeling across both temporal and spatial dimensions, resulting in insufficient spatiotemporal information mining for wind farm clusters. 2) Traditional power forecasting volatility research relies on analyzing single or localized sites, lacking predictability of overall wind farm cluster trend information.

[0006] Therefore, how to improve the accurate prediction of wind power has become an urgent problem to be solved. Summary of the Invention

[0007] The present application provides a method, apparatus, device, and storage medium for predicting wind power cluster power. First, the site feature sequence is encoded based on the addition of prior knowledge of the wind power cluster and reconstructed into a spatiotemporal variable graph. Secondly, a pure graph structure is introduced to achieve a unified spatiotemporal representation of the wind power cluster to obtain the total power feature sequence of the cluster. Then, the total power feature sequence is decomposed to obtain a long-term stable sequence and a short-term fluctuation sequence, and then different fluctuation characteristic prediction models are used to achieve fusion prediction. Finally, a trend-oriented loss function is used to guide the model to update parameters. This improves the accuracy of wind power prediction.

[0008] In a first aspect, the present application provides a wind power cluster power prediction method, comprising:

[0009] Constructing a prediction model based on spatiotemporal simultaneity; wherein the prediction model based on spatiotemporal simultaneity includes a spatiotemporal coding module, a unified spatiotemporal representation module, and a fluctuation fusion prediction module;

[0010] The spatiotemporal encoding module responds to the input power data of each wind farm, calculates the correlation coefficient between each wind farm to construct an adjacency matrix, and uses a graph convolutional network to encode the feature sequence of each station in the wind farm cluster to obtain a fully connected hypervariable graph;

[0011] The unified spatiotemporal representation module performs unified spatiotemporal representation on the input fully connected hypervariable graph in response to the input, to obtain spatiotemporal features;

[0012] The fluctuation fusion prediction module responds to the input spatiotemporal features, decomposes the spatiotemporal features into trend components and residual components, uses the trend components to perform long-term dependency prediction to obtain first predicted power data, uses the residual components to perform short-term fluctuation prediction to obtain second predicted power data, and fuses the first predicted power data and the second predicted power data to obtain predicted power data for the wind power cluster within a future time range;

[0013] According to the set loss function, the calculation and spatiotemporal simultaneity prediction model is trained, and the wind power cluster power prediction is realized according to the trained calculation and spatiotemporal simultaneity prediction model; wherein, the loss function includes trend error, amplitude error and error mean.

[0014] In one possible design, the spatiotemporal encoding module calculates the correlation coefficients between any two wind farms to construct an adjacency matrix, and uses a graph convolutional network to encode the characteristic sequences of each station in the wind farm cluster. The method for obtaining a fully connected hypervariable graph includes:

[0015] The correlation coefficient between two wind farms is calculated by the following formula:

[0016]

[0017] Where, Pcc(WF i ,WF j ) represents the i-th wind farm WF i and the j-th wind farm WF j The Pearson correlation coefficient of N represents the number of sites in the wind power cluster. and represent the average values ​​of the power data of the i-th wind farm and the j-th wind farm, and Respectively represent the power data of the i-th wind farm and the j-th wind farm at time t;

[0018] Based on a set threshold, edges whose Pearson correlation coefficient exceeds the set threshold are selected to construct an adjacency matrix A;

[0019] According to historical observation data, the multidimensional feature sequence is calculated by the following formula:

[0020] STC=GCNConv({hP 1 ,hP 2 ,…,hP N},A) (2)

[0021] In the formula, STC represents the multidimensional feature sequence, GCNConv represents the graph convolution operation, and hP 1 ,hP 2 ,hP N Respectively represent the observation data of the first, second and N-th wind farms within the historical time range, and the observation data include power data;

[0022] The multidimensional feature sequence is converted into a fully connected hypervariable graph.

[0023] In one possible design, the fully connected hypervariable graph is represented as in, Represents graph node features, represents a randomly initialized adjacency matrix, and the graph node features are determined by the following formula:

[0024]

[0025] Where, Represents the power data of the wind farm, represents the embedding matrix.

[0026] In one possible design, the unified spatiotemporal representation module performs unified spatiotemporal representation on the input fully connected hypervariable graph in response to the input, and obtains spatiotemporal features in a manner including:

[0027] Performing a discrete Fourier transform on the graph node features in the fully connected hypervariable graph to convert the time domain data into the frequency domain to obtain a first feature;

[0028] Based on the first feature, the second feature is calculated using the following formula:

[0029]

[0030] In the formula, Y represents the second characteristic, Y real and Y imag represent the real and imaginary parts of the second characteristic, X real and X imag Represent the real and imaginary parts of the first feature respectively, Relu is the activation function, W1 and W2 are weight matrices, b1 and b2 are biases, @ represents matrix dot product, and stack represents stack concatenation operation;

[0031] Based on the second feature, the third feature is calculated using the following formula:

[0032]

[0033] Where Y' represents the third feature, λ represents the adjustable parameter, SoftShrink represents the nonlinear activation function, and complex represents the complex transformation;

[0034] Based on the first and third characteristics, the fourth characteristic is obtained by the following formula:

[0035]

[0036] Where, Represents the fourth characteristic, represents the inverse discrete Fourier transform;

[0037] A linear projection operation is performed on the fourth feature to obtain a spatiotemporal feature.

[0038] In one possible design, the fluctuation fusion prediction module responds to input spatiotemporal features, decomposes the spatiotemporal features into trend components and residual components, uses the trend components to perform long-term dependency prediction to obtain first predicted power data, uses the residual components to perform short-term fluctuation prediction to obtain second predicted power data, and fuses the first predicted power data and the second predicted power data to obtain predicted power data for the wind power cluster within a future time range. The method includes:

[0039] Perform an average pooling operation on the spatiotemporal features to obtain the trend component, and subtract the trend component from the spatiotemporal features to obtain the residual component;

[0040] Inputting the trend component into the GRU network to obtain the first predicted power data;

[0041] The residual component is input into the Conv1d network, and the second predicted power data is obtained by performing a cross-correlation operation between the convolution kernel and the feature sequence;

[0042] The first predicted power data and the second predicted power data are added point by point to achieve feature fusion to obtain the predicted power data of the wind power cluster within a future time range.

[0043] In one possible design, the temporal and spatial features are average-pooled to obtain the trend component, and the calculation process of subtracting the trend component from the temporal and spatial features to obtain the residual component is expressed as:

[0044] X'=AvgPool(USTR)

[0045] X res =USTR-X' (8)

[0047] Where X' represents the trend component, X res Represents the residual component, USTR represents the spatiotemporal features, and AvgPool represents the average pooling operation;

[0048] In the process of inputting the trend component into the GRU network to obtain the first predicted power data, the parameter update process of the GRU network is:

[0049]

[0050] In the formula, σ represents the sigmoid function, ⊙ represents element-by-element multiplication, and h t represents the hidden state at time t, x t Indicates the input at time t, h t-1 represents the hidden state at time t-1 or the initial hidden state at time point 0, r t ,z t ,n t Represent the reset gate, update gate and new gate respectively, W3 and W4 represent the weight matrix of the reset gate, W5 and W6 represent the weight matrix of the update gate, W7 and W8 represent the weight matrix of the new gate, b3 and b4 represent the bias of the reset gate, b5 and b6 represent the bias of the update gate, b7 and b8 represent the bias of the new gate, and tanh represents the hyperbolic tangent activation function.

[0051] In one possible design, the loss function is expressed as:

[0052]

[0053]

[0054] loss=loss1+loss2+loss3 (15)

[0055] In the formula, loss1, loss2 and loss3 represent trend error, amplitude error and error mean respectively, and loss represents loss value. represents the observation difference, represents the forecast difference, represents the true value of the position at γ, Represents the predicted value at position γ, q, w, ..., o represent the extreme point positions in the data, and the value range of γ is {q, w, ..., o}. represents the true value at time t, Represents the predicted value at time t, t represents time t and its range is in τ, and τ represents the time set of the total length of the data.

[0056] In a second aspect, the present application provides a wind power cluster power prediction device, the device comprising:

[0057] A model building unit is configured to build a prediction model based on spatiotemporal simultaneity; wherein the prediction model based on spatiotemporal simultaneity includes a spatiotemporal coding module, a unified spatiotemporal representation module, and a fluctuation fusion prediction module;

[0058] The spatiotemporal encoding module responds to the input power data of each wind farm, calculates the correlation coefficient between each wind farm to construct an adjacency matrix, and uses a graph convolutional network to encode the feature sequence of each station in the wind farm cluster to obtain a fully connected hypervariable graph;

[0059] The unified spatiotemporal representation module performs unified spatiotemporal representation on the input fully connected hypervariable graph in response to the input, to obtain spatiotemporal features;

[0060] The fluctuation fusion prediction module responds to the input spatiotemporal features, decomposes the spatiotemporal features into trend components and residual components, uses the trend components to perform long-term dependency prediction to obtain first predicted power data, uses the residual components to perform short-term fluctuation prediction to obtain second predicted power data, and fuses the first predicted power data and the second predicted power data to obtain predicted power data for the wind power cluster within a future time range;

[0061] The model training unit is configured to train the calculation and spatiotemporal simultaneity prediction model according to a set loss function, and realize wind power cluster power prediction based on the trained calculation and spatiotemporal simultaneity prediction model; wherein the loss function includes trend error, amplitude error and error mean.

[0062] In a third aspect, an embodiment of the present application provides an electronic device comprising: at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the wind power cluster power prediction method described in the first aspect and various possible designs of the first aspect.

[0063] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores computer-executable instructions. When a processor executes the computer-executable instructions, the wind power cluster power prediction method described in the first aspect and various possible designs of the first aspect is implemented.

[0064] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the wind power cluster power prediction method described in the first aspect and various possible designs of the first aspect.

[0065] The wind power cluster power prediction method, device, equipment, and storage medium provided in this application have at least the following beneficial effects:

[0066] This application not only considers the simultaneity of time and space and proposes a unified time and space representation method based on prior knowledge encoding, so as to improve the model's ability to capture the time and space coupling characteristics of wind power clusters in time and space; it also proposes a dual-channel decomposition prediction module that considers the power trend characteristics of wind power clusters, applies different characteristic networks to different fluctuation characteristic sequences, and improves the accuracy of wind power cluster power trend information prediction; at the same time, it designs a trend-guided loss function to guide the model parameter update, and realizes the effective learning of the model on the power data trend information from three angles of error calculation, thereby improving the accuracy of wind power cluster power prediction. Therefore, this application can reflect the dynamic characteristics of the system, track future power trends, and has the advantages of simple calculation, high prediction performance, clear meaning, effective prediction results, and strong practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0068] Figure 1 A diagram of the wind propagation process in a wind farm group provided in an embodiment of the present application;

[0069] Figure 2 A schematic diagram of the spatiotemporal simultaneity of wind farm power generation in a wind farm cluster provided in an embodiment of the present application;

[0070] Figure 3 A schematic diagram of constructing a hypervariable graph provided in an embodiment of the present application;

[0071] Figure 4 This is a schematic diagram of the network structure of the fusion prediction module provided in an embodiment of the present application;

[0072] Figure 5 A schematic diagram of the trend error between the predicted value and the true value provided in the embodiment of the present application;

[0073] Figure 6 A schematic diagram of the amplitude error between the predicted value and the true value provided in the embodiment of the present application;

[0074] Figure 7 A schematic diagram of the correlation coefficient matrix between adjacent time points of wind farm cluster sites provided in an embodiment of the present application;

[0075] Figure 8 A schematic diagram of the correlation coefficient matrix between adjacent time points of wind farm cluster sites provided in an embodiment of the present application;

[0076] Figure 9Schematic diagram of 1h to 4h predicted power curves of different prediction methods in a possible embodiment of the present application; wherein (a) is 1 hour; (b) is 2 hours; (c) is 3 hours; (d) is 4 hours;

[0077] Figure 10 A schematic diagram of average values ​​of prediction indicators in 16 steps of different prediction methods in a possible embodiment of the present application;

[0078] Figure 11 A schematic diagram of the results of predicting the next 16 steps at different times using different loss functions in a possible embodiment of the present application; wherein (a) is prediction result 1; (b) is prediction result 2; (c) is prediction result 3; (d) is prediction result 4; (e) is prediction result 5; (f) is prediction result 6;

[0079] Figure 12 Schematic diagram of the average values ​​of 16-step prediction indicators of different loss functions in a possible embodiment of the present application; wherein, (a) RMSE; (b) MAE; (c) MAPE;

[0080] Figure 13 This is a structural diagram of a wind power cluster power prediction device provided in an embodiment of the present application.

[0081] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0082] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0083] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of information such as financial data or user data involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0084] It should be noted that in the embodiments of the present application, certain software, components, models and other existing solutions in the industry may be mentioned. They should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0085] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0086] See also Figure 1 In this embodiment, the wind farms in the wind farm cluster are distributed according to the wind belt direction. Wind farms 1 to N are affected by factors such as wind propagation trajectory and wind speed, resulting in different power generation and hysteresis. Figure 2 Assume that wind farm 1, wind farm 2, ..., wind farm N are in the same wind transmission path. Due to the geographical location of the wind farm, the wind arrival time is different, forming a set of non-uniform delay sequences {A0, A1, ..., A n}(Δa=A n -A n-1 ≠const). Within the time window b, that is, each station is in the range of A0~A0+b, A1~A1+b,…,A n ~A n The power generation curves within time range +b show highly similar fluctuation trends. This phenomenon reveals the spatial correlation of wind farm output and the dependence of output time on spatial and geographical distribution. Furthermore, because the power generation of each site is affected by factors such as terrain complexity and atmospheric temperature, the power within time range b exhibits similar trends, but with different amplitudes. This shows that the power generation of each site in a wind farm cluster is the result of the simultaneous effects of spatial and temporal dimensions.

[0087] Based on the above basic principles, an embodiment of the present application provides a wind power cluster power prediction method, which includes two steps, namely steps S10 and S20, which are described in detail as follows.

[0088] S10: Construct a prediction model based on spatiotemporal simultaneity, wherein the prediction model based on spatiotemporal simultaneity includes a spatiotemporal coding module, a unified spatiotemporal representation module, and a fluctuation fusion prediction module.

[0089] The power size of each station in the wind power cluster is the result of complex nonlinear dynamic characteristics formed by the simultaneous action of time and space dimensions. If the two dimensions are modeled separately and then fused and represented, the model will not be able to capture the spatiotemporal coupling characteristics. Therefore, this embodiment constructs a spatiotemporal simultaneity prediction model that combines modeling and representation in the time and space dimensions and fluctuation decomposition and fusion prediction to improve the unified spatiotemporal representation capability of the wind power cluster. It includes three parts: a spatiotemporal coding module, a unified spatiotemporal representation module, and a fluctuation fusion prediction module. The following will describe how the spatiotemporal simultaneity prediction model can achieve wind power cluster power prediction in combination with the specific data processing process of the above three modules.

[0090] The spatiotemporal encoding module responds to the input power data of each wind farm, calculates the correlation coefficient between each wind farm to construct an adjacency matrix, and uses a graph convolutional network to encode the feature sequence of each station in the wind power cluster to obtain a fully connected hypervariable graph.

[0091] Considering geographical factors such as distance and mountainous obstructions, the spatial correlations between wind farm sites in a wind farm cluster can vary in strength. Using known historical power information, information theory is used to calculate the correlations between sites and construct an adjacency matrix. This is then applied to a single-layer graph convolutional network to learn the feature sequences for each site. These feature sequences simultaneously encode historical power trends and cluster coordinated change information, allowing the model to initialize the coupling patterns between wind farm sites. Furthermore, this provides multidimensional feature sequence information with prior knowledge for the subsequent unified spatiotemporal representation module.

[0092] Therefore, in a possible embodiment, the Pearson coefficient is used to calculate the correlation coefficient between stations to construct an adjacency matrix, and then the graph convolution network is combined with the correlation to realize the encoding of the characteristic sequence of each station in the wind power cluster. i ∈N is a graph node. The correlation strength between stations, i.e., the edge relationship in the graph structure, is determined by the Pearson coefficient. The calculation process is as follows:

[0093]

[0094] Where, Pcc(WF i ,WF j ) represents the i-th wind farm WF i and the j-th wind farm WF j The Pearson correlation coefficient of N represents the number of sites in the wind power cluster. and represent the average values ​​of the power data of the i-th wind farm and the j-th wind farm, and They represent the power data of the i-th wind farm and the j-th wind farm at time t respectively.

[0095] Set a threshold and select edges whose Pearson coefficient exceeds the threshold to construct the adjacency matrix A∈R N×N , R represents a set of real numbers. Given the observation data of the i-th wind farm within the historical time range, Among them, hP i ∈R Z×T represents the i-th wind farm WF i Z feature data within the time range, represents the characteristic data of the i-th wind farm at time t, Z is the number of node characteristics, and T is the length of the historical time series. The calculation process of the spatiotemporal coding module is as shown in formula (2).

[0096] STC=GCNConv({hP 1 ,hP 2 ,…,hP N},A) (2)

[0097] In the formula, STC represents the multidimensional feature sequence, GCNConv represents the graph convolution operation, and hP 1 ,hP 2 ,hP N They represent the observation data of the first, second and N-th wind farms within the historical time range respectively, and the observation data include power data.

[0098] The input of the graph convolution operation GCNConv is the historical observation data {hP 0 ,…,hP i} and adjacency matrix A, the output is STC∈R N×T , that is, the historical power trend in the time dimension and the coupling law of cluster coordinated changes in the spatial dimension are encoded simultaneously to obtain a multi-dimensional feature sequence.

[0099] like Figure 3 , further converting the multidimensional feature sequence STC obtained by spatiotemporal coding into the initial fully connected hypervariable graph By constructing the high-resolution spatiotemporal dependency of any two wind farm station characteristic values ​​at any two time stamps. Indicates that all station data are treated as a graph node feature. represents a randomly initialized adjacency matrix. Subsequently, by embedding the matrix Assign a d-dimensional vector to each station data to achieve Projection The calculation process is as shown in Formula 3.

[0100]

[0101] The final fully connected hypervariable graph is represented as

[0102] The unified spatiotemporal representation module responds to the input fully connected hypervariable graph, performs unified spatiotemporal representation on the fully connected hypervariable graph, and obtains spatiotemporal features.

[0103] The goal of the unified spatiotemporal representation module is to uniformly represent the power generation of wind power clusters in both time and space dimensions.

[0104] In a possible embodiment, in order to simultaneously capture the characteristic information of the wind power cluster's interaction in time and space, the hypervariable graph in Fourier space is To perform unified spatiotemporal representation, the specific steps are as follows:

[0105] First, yes In the discrete space-time dimension, discrete Fourier transform (DFT) is performed to convert the time domain data into the frequency domain, and the complex output is obtained, that is, the first feature X, whose real part and imaginary part are X respectively. real ∈R (NT / 2)×d 、X imag ∈R (NT / 2)×d The calculation process is as shown in formula (4). Represents discrete Fourier transform.

[0106]

[0107] Randomly initialize the weight matrix W1∈R d×d , W2∈R d×d and the deviation b1∈R d ,b2∈R d The complex features converted into frequency domain are calculated by formula (4) to obtain the new real part Y real ∈R (NT / 2)×d and the imaginary part Y real ∈R (NT / 2)×d Value. And Y real and Y imag Use the stack() function to splice and get Y∈R (NT / 2)×d×2 .

[0108]

[0109] In the formula, Y represents the second characteristic, Y real and Y imag represent the real and imaginary parts of the second characteristic, X real and X imag Represent the real and imaginary parts of the first feature respectively, @ represents the matrix dot product, and stack represents the stack splicing operation.

[0110] The second feature Y is compressed using the SoftShrink nonlinear activation function to reduce unimportant features, and then a complex number is converted to a complex number, namely the third feature Y'. The calculation process is as shown in Equation (6), where λ is an adjustable parameter.

[0111]

[0112] Where Y' represents the third feature, λ represents the adjustable parameter, SoftmSHrink represents the nonlinear activation function, and complex represents the complex conversion.

[0113] Then input Y'+X into the inverse discrete Fourier transform (IDFT) to convert the frequency domain data back to the time domain to obtain The calculation process is as shown in formula (7).

[0114]

[0115] Where, Represents the fourth characteristic, represents the inverse discrete Fourier transform.

[0116] Finally, yes The linear projection operation obtains the universal spatio-temporal representation result, that is, the spatio-temporal feature USTR∈R 1×T .

[0117] The fluctuation fusion prediction module responds to the input spatiotemporal characteristics, decomposes the spatiotemporal characteristics into trend components and residual components, uses the trend components to perform long-term dependency prediction to obtain first predicted power data, uses the residual components to perform short-term fluctuation prediction to obtain second predicted power data, and fuses the first predicted power data and the second predicted power data to obtain predicted power data within the future time range of the wind power cluster.

[0118] The fluctuation fusion prediction module is aimed at the fluctuation characteristics of the total power of the wind power cluster. Figure 4 As shown, time series data is first deeply decomposed to obtain trend and residual components. The trend component is a low-frequency component characterized by low volatility and slow changes in the feature sequence. The residual component is a high-frequency component characterized by high volatility and dramatic changes in the feature sequence. Different methods are then used to learn and predict the characteristics of each component. GRU is used to capture the advantages of long-term trends in feature sequences and applied to low-frequency components for long-term dependency prediction. One-dimensional convolution is used to capture local features using convolution kernels and applied to high-frequency components for short-term fluctuation prediction. Finally, the outputs of the two methods are fused to achieve the final wind power cluster power prediction.

[0119] In a possible embodiment, the calculation process of decomposing the feature sequence into trend and residual components is as shown in formula (9), and the USTR is average pooled by AvgPool to obtain the trend component X trend , then use the original data X minus X' to get the residual component X res Input the trend component X' into the GRU network to obtain X' trend , its parameter update process is as shown in formula (9). res Input to Conv1d network to get X' res , through the cross-correlation operation between the convolution kernel and the feature sequence, we can get (X' res ) t Finally, as in formula (10), X' trend and X' res The feature fusion is achieved by point-by-point addition to obtain the predicted power data pP∈R of the wind power cluster within the future τ time range. 1×τ .

[0120] X'=AvgPool(USTR)

[0121] X res =USTR-X' (8)

[0123]

[0124] Where X' represents the trend component, X res represents the residual component, USTR represents the spatiotemporal feature, AvgPool represents the average pooling operation; σ represents the sigmoid function, ⊙ represents element-by-element multiplication, h t represents the hidden state at time t, x t Indicates the input at time t, h t-1 represents the hidden state at time t-1 or the initial hidden state at time point 0, r t ,z t ,n t Represent the reset gate, update gate and new gate respectively, W3 and W4 represent the weight matrix of the reset gate, W5 and W6 represent the weight matrix of the update gate, W7 and W8 represent the weight matrix of the new gate, b3 and b4 represent the bias of the reset gate, b5 and b6 represent the bias of the update gate, b7 and b8 represent the bias of the new gate, and tanh represents the hyperbolic tangent activation function; Represents point-by-point addition.

[0125] S20. According to the set loss function, the calculation and spatiotemporal simultaneity prediction model is trained, and the wind power cluster power prediction is realized according to the trained calculation and spatiotemporal simultaneity prediction model; wherein the loss function includes trend error, amplitude error and error mean.

[0126] After step S10 is implemented, a unified spatiotemporal representation forecast representing the future power state of the wind power cluster is obtained by taking into account the spatiotemporal simultaneity prediction model. The traditional MSE and MAE loss functions allow the model to accurately capture the long-term trend or seasonality of wind power, but ignore the short-term volatility of wind power. In order to further improve the multi-step prediction model's ability to fit the fluctuation of wind power curves, this paper designs a loss function CFLoss guided by the trend change of wind power. This loss function analyzes and calculates the final error from three perspectives: trend, amplitude, and error mean. Set the true value to The predicted value is

[0127] Regarding trend error, in the ultra-short-term prediction of wind power, there is a situation where the predicted value and the true value fail to effectively follow the trend change in the long-term range. To this end, the goal of the trend error is to make the predicted value of the model tend to be the same as the true value in terms of trend change. Therefore, this embodiment first calculates the trend change of the predicted value and the true value in the same time range, and then calculates the error value of the trend change of the two. The gradient descent guidance model is used to reduce the error, so that the prediction result is close to the true value in terms of trend. The calculation process is shown in formula (11).

[0128] Diff(y m+1 ,y m )=y' m =y m+1 -y m (11)

[0130] Where Diff() is the difference between the continuous sampling points of the time series data. Applying it to the actual value and the predicted value, the observed difference is obtained and the predicted difference And satisfy m = 1, 2, ..., τ-1. Then, the error of the two differences is calculated using loss1 in formula (12), guiding the predicted value change rate of the network model to gradually approach the true value change rate.

[0131]

[0132] Where loss1 is the trend error, and t represents time or moment.

[0133] The goal of the amplitude error is to minimize the upper and lower amplitude errors between the model's predicted value and the true value within the same time range, thereby guiding the model to learn different output states of the wind power cluster. First, calculate all extreme points in the true value. And according to its index {q, w, …, o}, the value of the corresponding position of the predicted value is obtained, and then the error between the maximum and minimum values ​​of the predicted value and the true value is calculated by formula (13), guiding the network model to approach the amplitude of the true value.

[0134]

[0135] Where loss2 represents the amplitude error, represents the true value of the position at γ, Represents the predicted value at position γ, q, w, …, o represent the extreme point locations in the data, and the value range of γ is {q, w, …, o}. Regarding the mean error, the model fails to achieve the problem of homogenization of the prediction error value due to the difference in the error between the model's predicted value and the true value at each moment. To this end, the goal of the mean error is to focus on the absolute error between the model's predicted value and the true value in a numerically averaged state, reducing the average value of the sum of the errors at all moments, namely the mean absolute error (MAE), calculated as shown in Equation (14).

[0136]

[0137] Where loss3 is the error mean, represents the true value, Represents the predicted value.

[0138] Therefore, the calculation process of the loss function CFLoss designed in this application is as shown in formula (15), which adds the three groups of error values ​​analyzed from different angles as the direction of gradient descent for model parameter update, thereby improving the accuracy of the model in predicting the observed power of the wind power cluster.

[0139] loss=loss1+loss2+loss3 (15)

[0140] Where loss represents the loss value.

[0141] In a possible embodiment, the prediction performance of the prediction model taking into account the spatiotemporal simultaneity is evaluated by the root mean square error, mean absolute error, and mean absolute percentage error of the prediction results;

[0142] The expression of the root mean square error is:

[0143]

[0144] Where RMSE is the root mean square error; i To predict wind power; is the actual wind power; cap is the wind farm station startup capacity; n is the number of prediction points;

[0145] The expression of the mean absolute error is:

[0146]

[0147] Where, MAE is the mean absolute error;

[0148] The expression of the mean absolute percentage error is:

[0149]

[0150] Where MAPE is the mean average percentage error.

[0151] In a possible embodiment, when the wind farm cluster power prediction method is implemented, the wind farm cluster data is spatiotemporally encoded using the Pearson correlation coefficient based on the historical station power generation. Figure 3 As shown in , the encoded representation data is reconstructed into a hypervariable graph, and the unified spatiotemporal representation module of the input obtains sequence feature information in the unified representation of time and space dimensions. Then, as Figure 4 The trend and residual components are obtained by deep decomposition and input into different characteristic neural networks for prediction and fusion to obtain the total power of wind power cluster. Figure 5 、 Figure 6 and Figure 7 A power fluctuation-oriented loss function is constructed to guide the update of model parameters, so that the final wind power cluster power prediction value is closer to the actual value fluctuation trend and amplitude change.

[0152] In a possible embodiment, a specific wind power prediction example is provided as follows:

[0153] This example uses multiple prediction models to perform ultra-short-term wind power forecasts on a test set, validated using real data from a regional wind farm fleet. Wind turbine power is sampled at 15-minute intervals at the wind farms, and meteorological data such as temperature, humidity, air pressure, wind speed, and wind direction are also predicted at 15-minute intervals by the meteorological forecast center. The collected dataset covers 2021 and includes 122 wind farms with a total installed capacity of 13,534.81 MW.

[0154] Figure 8This paper demonstrates a unified spatiotemporal representation of a wind farm cluster. The Pearson correlation coefficients of historical power between sites at adjacent moments are calculated. A correlation matrix is ​​shown for the four-hour historical power of the wind farm cluster at eight consecutive moments. Darker colors indicate stronger correlations, while darker colors indicate weaker correlations. The figure shows that each site has a high correlation coefficient with a fixed number of sites. Analysis of the site's geographic distribution reveals that sites with strong correlations are geographically close, indicating robust spatial interactions between wind farms. However, the strength of the correlations between sites at moments t1, t2, …, and t8 varies, indicating that each site is influenced by its own environment at different moments, resulting in autocorrelations in its own power generation. These factors demonstrate that the power generation of sites within a wind farm cluster is a function of both spatial and temporal dimensions and cannot be characterized and modeled solely through a single dimension.

[0155] Figure 9 The prediction curves of the proposed method and other methods are shown. The figure shows that the proposed method performs best at most times during the 1-hour, 2-hour, 3-hour, and 4-hour prediction scales, closely tracking the actual power value. Only the 800-1000 sampling points, which are in a high-output state, perform slightly worse than the GCN network. This is because the high output of a wind farm station is due to increased wind energy, and the spatial correlation at this time has a stronger impact on the station than the temporal correlation. However, over longer time scales, the output of a wind farm is still the result of the combined effects of time and space.

[0156] Table 1 shows the prediction performance indicators of different prediction methods at 1 hour, 2 hours, 3 hours, and 4 hours. Among them, the RMSE value is improved by 2.78%, 1.62%, and 1.15% compared with the suboptimal model at 1 hour, 2 hours, and 3 hours, respectively. Only at the 4-hour time scale is the RMSE value higher than the GRU by 0.04%. The MAE value indicator is the lowest compared with all other methods, and is reduced by 2.25%, 1.27%, 0.97%, and 0.12% compared with the suboptimal model at the 1-hour, 2-hour, 3-hour, and 4-hour time scales. According to the MAPE evaluation indicator, the method of this application outperforms the suboptimal model of the same prediction time scale at the prediction time scale of 1 hour, 2 hours, and 4 hours, with improvements of 2.82%, 1.42%, and 2.08%, respectively. However, at the 2-hour prediction time scale, the performance is 1.03% worse than that of the gated recurrent unit (GRU). As can be seen from the results, the proposed method achieves optimal performance across most metrics at 1-hour, 2-hour, 3-hour, and 4-hour forecast timescales. This demonstrates the effectiveness of unified temporal and spatial modeling of each site in a wind farm. While the 2-hour MAPE and 4-hour RMSE are higher than those of the GRU model, analysis of the forecast curves indicates that while the proposed method can track power trends well, it fails to capture the magnitude of cluster power changes at high output levels.

[0157] Table 1 Model prediction performance at different prediction time scales

[0158]

[0159]

[0160] Figure 10 A comparison chart shows the performance of this application method and other methods in the ultra-short-term 4-hour forecast, accumulating and averaging the performance indicators of each step of the forecast. As can be seen from the chart, this application method achieves the best average results across all evaluation indicators. In terms of RMSE, this application method outperforms the baseline methods of temporal convolutional networks (TCN) and long short-term memory networks (LSTM) by 4.38% and 4.09%, respectively. It also outperforms the suboptimal model by 1.8%. In terms of MAE, this application method outperforms the baseline methods of TCN and LSTM by 3.73% and 3.28%, respectively. It also outperforms the suboptimal model by 1.48%. In terms of MAPE, this application method outperforms the baseline methods of TCN and LSTM by 40.11% and 13.61%, respectively. It also outperforms the suboptimal model by 4.62%.

[0161] Figure 11 A comparison chart of the prediction results of CFLoss and other loss functions MSELoss, MAELoss, SmoothLoss, and SoftDTWLoss for the next 16 moments is shown. It can be seen that compared with other methods, the prediction results of the present application method in the first 15 steps can be closer to the true value. This is because while considering the absolute error between the predicted value and the true value, the present application method adds the calculation of the influencing factors of the time series change and the amplitude change between the maximum and minimum values ​​of the feature sequence. In the 16th step prediction, there are certain moments when the MSELoss prediction is closer to the true value than the CFLoss in this paper. However, this does not affect the performance of the loss function constructed in this paper in predicting the fluctuation trend of power in the entire next 16 steps compared to other loss functions.

[0162] Figure 12 The average values ​​of the three evaluation indicators of the method of this application and other loss methods are shown. It can be seen that the method of this application is the best in all indicators compared with other loss functions. Compared with MSELoss, MAPE is reduced by 5.6%. MAPE represents the ratio of the deviation between the predicted value and the actual value to the actual value. This shows that the loss function constructed in this article can well guide the model to capture the changing trend and fluctuation law of wind power. It also proves the effectiveness of the power fluctuation loss function (CFLoss) proposed in this application. The RMSE and MAE were reduced by 0.6% and 0.5% respectively, which shows that the loss function constructed in this application can not only capture the fluctuation trend of power, but also reduce the error of prediction results.

[0163] Table 2 shows the calculation results of the three evaluation metrics for different combinations of the components designed in this application. Method 1 applies only the unified spatiotemporal representation module in this application, Method 2 combines the agreed spatiotemporal representation module with prior knowledge, Method 3 combines the unified spatiotemporal representation module with a fusion prediction module, and Method 4 combines the unified spatiotemporal representation module with a fusion prediction module based on prior knowledge. +CFLoss replaces the original MSELoss with the composite loss function (CFLoss) designed in this application to guide model training.

[0164] As can be seen, the combined use of all components achieves optimal wind turbine cluster power prediction accuracy. Method 1, which uses only the unified spatiotemporal representation module, achieves a 1.48% lower RMSE than the suboptimal GCN model in Table 1. This is because unified temporal and spatial modeling effectively represents the spatiotemporal interactions of wind turbine cluster sites without violating natural laws by modeling the two separately. The first four prediction performance groups show that Method 2 significantly contributes to improved modeling performance by incorporating prior knowledge of the wind turbine cluster into the model. Method 3, based on the unified spatiotemporal representation, adds decoupling of feature sequences for separate predictions. While the improvement in prediction accuracy is not as significant as that of Method 2, its MAPE value decreases more significantly than that of Method 2, indicating that Method 3's predictions closely follow the true value trend. Combining all three modules achieves optimal prediction performance. Applying the CFLoss proposed in this application shows that all three metrics improve, demonstrating that jointly guiding model parameter updates from the perspectives of amplitude, trend, and mean error is highly effective in improving model prediction performance.

[0165] Table 2 Ablation experiment

[0166] Prediction Methods RMSE MAE MAPE Method 1 0.0672 0.0543 0.2253 Method 2 0.0580 0.0508 0.2113 Method 3 0.0625 0.0529 0.1921 Method 4 0.0540 0.0426 0.2161 Method 1 + CFLoss 0.0562 0.0391 0.1936 Method 2 + CFLoss 0.0515 0.0394 0.1677 Method 3 + CFLoss 0.0595 0.0475 0.1891 Method 4 + CFLoss 0.0487 0.0378 0.1607

[0167] Table 3 shows the prediction time of different models. It can be seen that the method proposed in this application takes less time to train and test than other graph convolution networks using the same MSELoss loss function, and takes more time than recurrent convolution networks and temporal convolution networks. Furthermore, the time required to perform another test set prediction is in seconds, meeting the time requirements of power grid dispatch in ultra-short-term prediction. After applying the CFLoss loss function constructed in this application, the training time of the model increases significantly, but the prediction time of the test set does not change significantly.

[0168] Table 3 Time consumption of prediction model

[0169] Training duration (s) Test duration (s) LSTM+MSELoss 104.144 0.04074 GRU+MSELoss 94.713 0.02145 TCN+MSELoss 121.055 0.04203 GCN+MSELoss 743.478 0.64739 GCNLSTM+MSELoss 759.5376 0.64103 STGCN+MSELoss 1118.469 0.86556 Our+MSELoss 443.700 0.26645 Our+CFLoss 3207.503 0.26652

[0170] In summary, this application proposes a unified spatiotemporal representation module based on prior knowledge encoding for ultra-short-term multi-step wind power cluster forecasting, constructs a dual-channel decomposition forecasting module to achieve forecasting of sequences with different characteristics, and constructs a composite loss function CFLoss to guide model parameter updates. Its substantial features and progress are reflected in the following aspects:

[0171] 1) According to the spatiotemporal simultaneity of wind power clusters, a unified spatiotemporal representation based on prior knowledge encoding is constructed to mine the characteristic information of the joint action of sites in the spatiotemporal dimensions, so that the model can fully represent the collaborative evolution information of wind power clusters.

[0172] 2) Considering the trend characteristics of wind power cluster power, the characterization results are decomposed into trend components and residual components, and different characteristic networks are applied for feature learning to achieve prediction, so that the final prediction can follow the power trend.

[0173] 3) Construct a composite loss function CFLoss to calculate the error from three perspectives: power amplitude, trend change and error mean, effectively guiding the update of model parameters to capture sequence change trends and fluctuation information.

[0174] 4) Experimental verification demonstrates that the proposed wind power cluster power prediction method reduces RMSE, MAE, and MAPE by 2.33%, 1.97%, and 10.17%, respectively. This research provides a new approach for renewable energy cluster power prediction and has significant engineering application value. This application provides a practical solution for wind power prediction, and its technical approach is of reference value for feature extraction and multi-task collaborative optimization in time series prediction tasks.

[0175] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied in a distributed scenario, where multiple devices cooperate with each other to complete the method. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiment of the present disclosure, and the multiple devices will interact with each other to complete the method.

[0176] The present application also provides a wind power cluster power prediction device, such as Figure 13 As shown, the wind power cluster power prediction device includes:

[0177] The model construction unit 1301 is configured to construct a spatiotemporal simultaneity prediction model; wherein the spatiotemporal simultaneity prediction model includes a spatiotemporal coding module, a unified spatiotemporal representation module, and a fluctuation fusion prediction module;

[0178] The spatiotemporal encoding module responds to the input power data of each wind farm, calculates the correlation coefficient between each wind farm to construct an adjacency matrix, and uses a graph convolutional network to encode the feature sequence of each station in the wind farm cluster to obtain a fully connected hypervariable graph;

[0179] The unified spatiotemporal representation module performs unified spatiotemporal representation on the input fully connected hypervariable graph in response to the input, to obtain spatiotemporal features;

[0180] The fluctuation fusion prediction module responds to the input spatiotemporal features, decomposes the spatiotemporal features into trend components and residual components, uses the trend components to perform long-term dependency prediction to obtain first predicted power data, uses the residual components to perform short-term fluctuation prediction to obtain second predicted power data, and fuses the first predicted power data and the second predicted power data to obtain predicted power data for the wind power cluster within a future time range;

[0181] The model training unit 1302 is configured to train the temporal and spatial simultaneity prediction model according to a set loss function, and realize wind power cluster power prediction based on the trained temporal and spatial simultaneity prediction model; wherein the loss function includes trend error, amplitude error and error mean.

[0182] An embodiment of the present application provides an electronic device, which may include a processor and a memory, wherein the processor and the memory can communicate with each other; illustratively, the processor and the memory communicate with each other via a communication bus.

[0183] The processor executes the computer-executable instructions stored in the memory, so that the processor implements the solutions in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0184] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. System buses can be categorized as address buses, data buses, and control buses. Transceivers enable communication between the database access device and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) or non-volatile memory.

[0185] The electronic device provided in the embodiment of the present application may be the terminal device of the above embodiment.

[0186] An embodiment of the present application further provides a computer-readable storage medium, in which computer instructions are stored. When the computer instructions are executed on a computer, the computer executes the technical solution of the wind power cluster power prediction method of the above embodiment.

[0187] An embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When at least one processor executes the computer program, it can implement the technical solution of the wind power cluster power prediction method in the above embodiment.

[0188] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or module, which can be electrical, mechanical or other forms.

[0189] Modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network elements. Some or all of these modules may be selected to implement the solution of this embodiment based on actual needs.

[0190] In addition, the functional modules in the various embodiments of the present application may be integrated into a single processing unit, or each module may exist physically separately, or two or more modules may be integrated into a single unit. The above-mentioned modules may be implemented in the form of hardware or hardware plus software functional units.

[0191] The above-mentioned integrated module implemented in the form of a software functional module can be stored in a computer-readable storage medium. The above-mentioned software functional module is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to perform some steps of the methods of various embodiments of the present application.

[0192] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), or application-specific integrated circuits (ASICs). A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the present invention may be directly implemented by a hardware processor or implemented by a combination of hardware and software modules in the processor.

[0193] The memory may include a high-speed RAM memory, and may also include non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk.

[0194] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be divided into address buses, data buses, and control buses.

[0195] The storage medium may be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium may be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0196] An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist as discrete components in an electronic control unit or a main control device.

[0197] Those skilled in the art will appreciate that all or part of the steps in the above-described method embodiments can be implemented using hardware associated with program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A wind power cluster power prediction method, characterized in that: The method comprises: Constructing a prediction model based on spatiotemporal simultaneity; wherein the prediction model based on spatiotemporal simultaneity includes a spatiotemporal coding module, a unified spatiotemporal representation module, and a fluctuation fusion prediction module; The spatiotemporal encoding module responds to the input power data of each wind farm, calculates the correlation coefficient between each wind farm to construct an adjacency matrix, and uses a graph convolutional network to encode the feature sequence of each station in the wind farm cluster to obtain a fully connected hypervariable graph; The unified spatiotemporal representation module performs unified spatiotemporal representation on the input fully connected hypervariable graph in response to the input, to obtain spatiotemporal features; The fluctuation fusion prediction module responds to the input spatiotemporal features, decomposes the spatiotemporal features into trend components and residual components, uses the trend components to perform long-term dependency prediction to obtain first predicted power data, uses the residual components to perform short-term fluctuation prediction to obtain second predicted power data, and fuses the first predicted power data and the second predicted power data to obtain predicted power data for the wind power cluster within a future time range; According to the set loss function, the calculation and spatiotemporal simultaneity prediction model is trained, and the wind power cluster power prediction is realized according to the trained calculation and spatiotemporal simultaneity prediction model; wherein, the loss function includes trend error, amplitude error and error mean.

2. The wind power cluster power prediction method according to claim 1, characterized in that: The spatiotemporal coding module calculates the correlation coefficients between each wind farm to construct an adjacency matrix, and uses a graph convolutional network to encode the characteristic sequence of each station in the wind farm cluster. The method of obtaining a fully connected hypervariable graph includes: The correlation coefficient between two wind farms is calculated by the following formula: Where, Pcc(WF i ,WF j ) represents the i-th wind farm WF i and the j-th wind farm WF j The Pearson correlation coefficient of N represents the number of sites in the wind power cluster. and represent the average values ​​of the power data of the i-th wind farm and the j-th wind farm, and Respectively represent the power data of the i-th wind farm and the j-th wind farm at time t; Based on a set threshold, edges whose Pearson correlation coefficient exceeds the set threshold are selected to construct an adjacency matrix A; According to historical observation data, the multidimensional feature sequence is calculated by the following formula: STC=GCNConv({hP 1 ,hP 2 ,…,hP N },A) (2) In the formula, STC represents the multidimensional feature sequence, GCNConv represents the graph convolution operation, and hP 1 ,hP 2 ,hP N Respectively represent the observation data of the first, second and N-th wind farms within the historical time range, and the observation data include power data; The multidimensional feature sequence is converted into a fully connected hypervariable graph.

3. The wind power cluster power prediction method according to claim 1 or 2, characterized in that: The fully connected hypervariable graph is represented as in, Represents graph node features, represents a randomly initialized adjacency matrix, and the graph node features are determined by the following formula: Where, Represents the power data of the wind farm, represents the embedding matrix.

4. The wind power cluster power prediction method according to claim 3, characterized in that: The unified spatiotemporal representation module performs unified spatiotemporal representation on the input fully connected hypervariable graph in response to the input, and obtains spatiotemporal features in the following manner: Performing a discrete Fourier transform on the graph node features in the fully connected hypervariable graph to convert the time domain data into the frequency domain to obtain a first feature; Based on the first feature, the second feature is calculated using the following formula: In the formula, Y represents the second characteristic, Y real and Y imag represent the real and imaginary parts of the second characteristic, X real and X imag Represent the real and imaginary parts of the first feature respectively, Relu is the activation function, W1 and W2 are weight matrices, b1 and b2 are biases, @ represents matrix dot product, and stack represents stack concatenation operation; Based on the second feature, the third feature is calculated using the following formula: Where Y' represents the third feature, λ represents the adjustable parameter, SoftShrink represents the nonlinear activation function, and complex represents the complex transformation; Based on the first and third characteristics, the fourth characteristic is obtained by the following formula: Where, Represents the fourth characteristic, represents the inverse discrete Fourier transform; A linear projection operation is performed on the fourth feature to obtain a spatiotemporal feature.

5. The wind power cluster power prediction method according to claim 1, characterized in that: The fluctuation fusion prediction module responds to the input spatiotemporal features, decomposes the spatiotemporal features into trend components and residual components, uses the trend components to perform long-term dependency prediction to obtain first predicted power data, uses the residual components to perform short-term fluctuation prediction to obtain second predicted power data, and fuses the first predicted power data and the second predicted power data to obtain predicted power data for the wind power cluster within a future time range. The method includes: Perform an average pooling operation on the spatiotemporal features to obtain the trend component, and subtract the trend component from the spatiotemporal features to obtain the residual component; Inputting the trend component into the GRU network to obtain the first predicted power data; The residual component is input into the Conv1d network, and the second predicted power data is obtained by performing a cross-correlation operation between the convolution kernel and the feature sequence; The first predicted power data and the second predicted power data are added point by point to achieve feature fusion to obtain the predicted power data of the wind power cluster within a future time range.

6. The wind power cluster power prediction method according to claim 5, characterized in that: The average pooling operation of spatiotemporal features is performed to obtain the trend component, and the calculation process of subtracting the trend component from the spatiotemporal features to obtain the residual component is expressed as: X'=AvgPool(USTR) X res =USTR-X’ (8) Where X' represents the trend component, X res Represents the residual component, USTR represents the spatiotemporal features, and AvgPool represents the average pooling operation; In the process of inputting the trend component into the GRU network to obtain the first predicted power data, the parameter update process of the GRU network is: In the formula, σ represents the sigmoid function, ⊙ represents element-by-element multiplication, and h t represents the hidden state at time t, x t Indicates the input at time t, h t-1 represents the hidden state at time t-1 or the initial hidden state at time point 0, r t ,z t ,n t Represent the reset gate, update gate and new gate respectively, W3 and W4 represent the weight matrix of the reset gate, W5 and W6 represent the weight matrix of the update gate, W7 and W8 represent the weight matrix of the new gate, b3 and b4 represent the bias of the reset gate, b5 and b6 represent the bias of the update gate, b7 and b8 represent the bias of the new gate, and tanh represents the hyperbolic tangent activation function.

7. The wind power cluster power prediction method according to claim 1, characterized in that: The loss function is expressed as: loss=loss1+loss2+loss3 (15) In the formula, loss1, loss2 and loss3 represent trend error, amplitude error and error mean respectively, and loss represents loss value. represents the observation difference, represents the forecast difference, represents the true value of the position at γ, Represents the predicted value at position γ, q, w, ..., o represent the extreme point positions in the data, and the value range of γ is {q, w, ..., o}. represents the true value at time t, Represents the predicted value at time t, t represents time t and its range is in τ, and τ represents the time set of the total length of the data.

8. A wind power cluster power prediction device, characterized in that: The device comprises: A model building unit is configured to build a prediction model based on spatiotemporal simultaneity; wherein the prediction model based on spatiotemporal simultaneity includes a spatiotemporal coding module, a unified spatiotemporal representation module, and a fluctuation fusion prediction module; The spatiotemporal encoding module responds to the input power data of each wind farm, calculates the correlation coefficient between each wind farm to construct an adjacency matrix, and uses a graph convolutional network to encode the feature sequence of each station in the wind farm cluster to obtain a fully connected hypervariable graph; The unified spatiotemporal representation module performs unified spatiotemporal representation on the input fully connected hypervariable graph in response to the input, to obtain spatiotemporal features; The fluctuation fusion prediction module responds to the input spatiotemporal features, decomposes the spatiotemporal features into trend components and residual components, uses the trend components to perform long-term dependency prediction to obtain first predicted power data, uses the residual components to perform short-term fluctuation prediction to obtain second predicted power data, and fuses the first predicted power data and the second predicted power data to obtain predicted power data for the wind power cluster within a future time range; The model training unit is configured to train the calculation and spatiotemporal simultaneity prediction model according to a set loss function, and realize wind power cluster power prediction based on the trained calculation and spatiotemporal simultaneity prediction model; wherein the loss function includes trend error, amplitude error and error mean.

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

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

Images