Wide-area multi-wind power plant short-term prediction method based on time-space graph convolutional neural network

Abstract

The invention discloses a wide-area multi-wind-power-plant short-term prediction method based on a time-space graph convolutional neural network. The wide-area multi-wind-power-plant short-term prediction method is specifically implemented according to the following steps that S1, strong correlation meteorological characteristics are determined; s2, determining a time-space graph multi-node feature set; s3, taking historical output correlation among multiple wind power plants as a time-space map edge feature; s4, introducing an information propagation mechanism in the time-space graph neural network, and constructing an information propagation mechanism aggregating historical output strong correlation station data weather forecast and historical output high-dimensional features, and S5, introducing a loss function combining absolute error loss and cross entropy loss in the training process of the time-space graph neural network, and completing multi-wind-power-plant short-term prediction in combination with a gating circulation unit layer. According to the wide-area multi-wind-power-plant short-term prediction method based on the time-space graph convolutional neural network, the problem that in the prior art, a prediction value and an actual value of wind power have a large difference is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of short-term wind power prediction, and in particular relates to a wide-area multi-wind farm short-term prediction method based on a time-space graph convolutional neural network. Background Art

[0002] Existing studies have paid little attention to the spatial correlation between adjacent wind farms, and when only the numerical weather forecast data of a single wind farm is used, the useful information contained in the numerical weather forecast data of neighboring stations may be ignored, thereby reducing the accuracy of wind farm prediction. If there is a large difference between the predicted value and the actual value of wind power, it may cause frequency and voltage imbalance in the power system and have a negative impact on the stability of the power system. Improving the accuracy of wind power prediction can reasonably weaken the adverse effects of its uncertainty on the power system when wind power is connected to the grid, and improve the reliability of system operation. Summary of the invention

[0003] The purpose of the present invention is to provide a wide-area multi-wind farm short-term prediction method based on a time-space graph convolutional neural network to solve the problem in the prior art that there is a large difference between the predicted value and the actual value of wind power.

[0004] The technical solution adopted by the present invention is a wide-area multi-wind farm short-term prediction method based on a time-space graph convolutional neural network, which is specifically implemented according to the following steps:

[0005] S1. Analyze the correlation between wind power output and meteorological characteristics of a certain region based on the maximum information coefficient and determine the strongly correlated meteorological characteristics;

[0006] S2, determine the multi-node feature set of the time-space graph through the meteorological characteristics strongly correlated with the output forecast and the historical output of wind power;

[0007] S3, determine the spatial correlation between wind farms through the adjacency matrix, and use the historical output correlation between multiple wind farms as the time-space graph edge feature;

[0008] S4. By introducing the information propagation mechanism in the spatiotemporal graph neural network, an information propagation mechanism is constructed to aggregate the weather forecast data of the station with strong correlation with the historical output and the high-dimensional features of the historical output;

[0009] S5. In the training process of spatiotemporal graph neural network, a loss function combining absolute error loss and cross entropy loss is introduced, and the gated recurrent unit layer is combined to complete the short-term prediction of multiple wind farms.

[0010] The present invention is also characterized in that:

[0011] The meteorological characteristics in S1 include wind speed, wind direction, temperature, humidity and air pressure characteristics.

[0012] The specific process of S1 is as follows:

[0013] Taking the meteorological characteristics within a certain region and a certain time period as variables, and arbitrarily given two variables (T i , T j ), collecting the weather forecast data set D = (x1, y1),..., (x n , y n ) of a certain region and a certain time period, where x i represents the wind power output of the wind farm at the i-th sampling, y i represents the meteorological characteristics at the i-th sampling, i takes values from 1 to n, and n represents the total number of samplings. The formula for calculating the mutual information in the α-by-β grid is:

[0014]

[0015] where I(D, α, β) is the mutual information value in the α-by-β grid, the integer (α, β) is a feature pair formed by any one of the wind power output and the weather forecast characteristics, p(x, y) represents the joint probability density under the current grid division, p(x) represents the marginal probability density of the wind power output, and p(y) represents the marginal probability density of the weather forecast characteristics;

[0016] Find the maximum mutual information value in different grids and normalize it. MIC(T i , T j ) is obtained by calculating the maximum of the mutual information values:

[0017]

[0018] MIC(T i , T j ) = MIC(D) = max αβ<B(n) {M(D)} (3);

[0019] where log(min{α, β}) is for normalization calculation, and (α, β) must satisfy the prerequisite αβ < B(n), and B(n) can be expressed as B(n) = n0.6. By statistically analyzing the daily mean absolute error of the wind farm under different meteorological scenarios, the meteorological characteristics strongly correlated with the mean absolute error of the wind power output prediction are determined according to MIC, that is, wind speed and wind direction.

[0020] The specific process of S3 is as follows:

[0021] The neighborhood relationship is defined using the spatial correlation between m and n in geographical location or historical output data, that is, node m is connected to node n. The adjacency matrix A is a symmetric square matrix of size N, where each element A(m, n) reflects the spatial correlation between two wind farms m and n. The adjacency matrix is defined as follows:

[0022] Amn =ReLu(MIC(B m ,B n )) (4);

[0023]

[0024] MIC(B m ,B n ) is used to calculate the correlation between nodes, B m ,B n is the corresponding node vector, and ReLu is the activation function used to avoid negative correlation;

[0025] As shown in formula (6), the correlation between the output mean absolute error of multiple wind farms is:

[0026] Construct the adjacency matrix:

[0027] MAE mn =ReLu(MIC(MAE m ,MAE n )) (6);

[0028] Where MAE m with MAE n is the mean absolute error between wind farm m and wind farm n. Since the mean absolute errors between wind farms are different, they are used as edge features of the time-space graph.

[0029] The specific process of S4 is as follows:

[0030] The information propagation mechanism is used for message passing. In the message passing phase, each node and edge aggregates information from its neighbors through an aggregation function and passes the generated messages to adjacent nodes and edges. During feature updating, each node and edge updates its features according to the aggregated messages received from its neighbors. The information passing phase runs T time steps and depends on the message function M. t And the node update function U t , in the message passing phase, the hidden state of each node According to the news To update, the specific process is as follows:

[0031]

[0032] in, The average aggregation function is used to aggregate wind farm information with strong correlation of mean absolute error. R acts on the set of node states and must keep the arrangement of node states unchanged so that the information propagation mechanism remains unchanged for graph isomorphism. y′ can output all updated hidden states and realize full-domain and full-feature enhancement of the spatiotemporal graph.

[0033] The specific process of introducing the loss function combining absolute error loss and cross entropy loss in the training process of spatiotemporal graph neural network in S5 is as follows:

[0034] Optimize the class weight vector of small samples, that is, pay more attention to small sample data. Let K be the number of samples in each class. m The vector of class weights W is expressed as follows:

[0035]

[0036] Using the categorical cross entropy loss function for M classes, the improved L′ CE The loss function can be expressed as:

[0037]

[0038] Where N is the number of training samples, y is the label of the sample, and z is the output of the final layer;

[0039] Let x i and f i represent the actual wind power and the predicted wind power respectively, and the MAE loss is

[0040] The meaning is as follows:

[0041]

[0042] The final loss function is as follows:

[0043] Loss = λ1L′ CE +λ2L MAE (13);

[0044] Wherein, λ1 and λ2 are weighting parameters, and the weighting parameters are set to λ1=0.5, λ2=0.5.

[0045] The specific process of combining the gated cyclic unit layer in S5 to complete the short-term prediction of multiple wind farms is as follows:

[0046] A GRUL with a gating mechanism is established. The enhanced high-dimensional feature vector output from the MPNN at each moment is used as the input of the GRUL to achieve wide-area multi-wind farm wind power prediction. The GRUL includes two multiplication gates, the update gate u t and reset gate r t GRUL uses MPNN to extract the feature output h t , and the hidden layer output h at time t-1 t-1 As input, complete the prediction, the specific formula is as follows:

[0047] u t =σ(A t [X′t ,h t-1 ]EW z +Eb u ) (14);

[0048] r t =σ(A t [X′ t ,h t-1 ]EW r +Eb r ) (15);

[0049] C t =tanh(A t [X′ t , r*h t-1 ]EW c +Eb c ) (16);

[0050] h t =u t *h t-1 +(1-u t )*c t (17);

[0051] Where: σ represents the sigmoid activation function, tanh is the activation function, E, W u , W r , W c , b u , b r and b c are all learnable weight and bias parameters in the model, X′ t represents the node features output by MPNN, A represents the connection relationship between multiple wind farms generated by the adjacency matrix module, and h t-1 represents the output at time t-1, h t Represents the wind power sequence output at time t.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] The present invention provides a wide-area multi-wind farm short-term prediction method based on a time-space graph convolutional neural network. By constructing a wide-area multi-wind farm time-space graph, the meteorological features strongly correlated with the historical wind power output prediction and the historical wind power output are used to determine the time-space graph multi-node feature set. At the same time, the adjacency matrix is ​​used to determine the spatial correlation, and the time-space graph edge features are constructed with the wind power output correlation between multiple wind farms as the spatial correlation. By introducing the information propagation mechanism in the time-space graph convolutional neural network, the average aggregation function is used to extract the numerical weather forecast information of the station with strong feature correlation, and the information propagation mechanism of the numerical weather forecast of the station with strong error correlation and the high-dimensional features of the historical output is constructed to achieve the node-edge feature update of the station to be predicted, fully improve the error-strong correlation station numerical weather forecast utilization rate of the station to be predicted, achieve the full enhancement of the global node-edge feature, and input the gated recurrent unit layer to mine its time domain features. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of a method for short-term prediction of wide-area multiple wind farms based on a time-space graph convolutional neural network according to the present invention;

[0055] Figure 2 It is a schematic diagram of constructing the node features of the present invention;

[0056] Figure 3 It is a schematic diagram of the principle of the node feature of the present invention;

[0057] Figure 4 It is a schematic diagram of the principle of the information propagation mechanism of the present invention;

[0058] Figure 5 Schematic diagram of the effect of the prediction effect of the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments.

[0060] Example 1

[0061] The present invention provides a wide-area multi-wind farm short-term prediction method based on a time-space graph convolutional neural network, such as Figure 1 As shown, the specific steps are as follows:

[0062] S1. Analyze the correlation between wind power output and meteorological characteristics of a certain region based on the maximum information coefficient and determine the strongly correlated meteorological characteristics.

[0063] S1 is specifically as follows: By comparing various analysis methods, the Maximal Information Coefficient (MIC) is applicable to linear and non-linear data, and is not limited to specific function types, with higher generalization ability. The MIC is an information-theoretic method based on grid partitioning for finding the correlation between two variables. Taking meteorological characteristics in a certain area and a certain time period as variables, given any two variables (T i , T j ), collect the weather forecast data set D = (x1, y1),...,(x n , y n ) in a certain area and a certain time period. MIC(T i , T j ) can be calculated based on the mutual information (Maximal Information, MI). Among them, x i represents the wind power output of the wind farm at the i-th sampling, y i represents the meteorological characteristics at the i-th sampling, i takes values from 1 to n, and n represents the total number of samplings. Calculate the mutual information in the α-by-β grid, which can be described as:

[0064]

[0065] where I(D, α, β) is the size of the mutual information in the α-by-β grid, the integers (α, β) can be any pair, p(x, y) is the joint probability density under the current grid partitioning, and p(x) and p(y) are the marginal probability densities. Find the maximum MI in different grids and normalize it. MIC(T i , T j ) is obtained by calculating the maximum value of MI:

[0066]

[0067] MIC(T i , T j ) = MIC(D) = max αβ＜B(n) {M(D)} (3);

[0068] where log(min{α, β}) is for normalization calculation, and (α, β) must satisfy the prerequisite αβ < B(n), and B(n) can be expressed as B(n) = n0.6. By statistically analyzing the daily mean absolute error of the wind farm under different meteorological scenarios, determine the meteorological characteristics strongly correlated with the mean absolute error of wind power output prediction, namely wind speed and wind direction, according to MIC.

[0069] S2. Determine the spatio-temporal graph multi-node feature set based on the meteorological characteristics strongly correlated with the output prediction and the historical wind power output.

[0070] S2 is specifically as follows: As Figure 2 and Figure 3 As shown in the figure, the meteorological characteristics include wind power, wind direction, wind speed, temperature, air pressure, humidity and other meteorological characteristics for feature selection, that is, the multi-node feature set of the time-space graph is determined by using the strongly correlated meteorological characteristics and the historical output of wind power. The selected features are used as the multi-node feature set to construct the node features of the time-space graph. The node features of the time-space graph can accurately reflect the attributes of the wind farm and provide a basis for the construction of subsequent edge features and time-space graphs.

[0071] S3. Determine the spatial correlation between wind farms through the adjacency matrix, and use the historical output correlation between multiple wind farms as the time-space graph edge feature.

[0072] S3 is specifically: In order to accurately describe the complex time-space coupling association of multiple wind farms in a wide area, the time-space coupling association of wind farms existing in non-Euclidean space is mapped to the Euclidean time-space graph. The neighborhood relationship is defined by the spatial correlation of m and n in the geographical location or historical output data, that is, node m is connected to node n, and the adjacency matrix A is a symmetric square matrix of size N, in which each element A(m,n) reflects the spatial correlation between two wind farms m and n. The adjacency matrix is ​​defined as follows:

[0073] A mn =ReLu(MIC(B m ,B n )) (4);

[0074]

[0075] MIC(B m ,B n ) is used to calculate the correlation between nodes, B m ,B n is the corresponding node vector, and ReLu is the activation function used to avoid negative correlation;

[0076] For wind farms m and j, there are H historical wind power time series values ​​p(t), and t∈[1,H]. Formula (3) is used to take the historical wind power output as the spatial judgment basis. If P mn =ReLu(MIC(p m ,p n ))>τ>0, then (m,n)∈E, otherwise, there is no connection between the two nodes. Similarly, when the geographical distance between wind farms is used as the judgment, that is, D mn =RELu(MIC(d m ,d n ))>τ>0, there is spatial correlation between the two points.

[0077] However, there are some limitations in using the above method to determine spatial correlation. The geographical distance of multiple wind farms does not take into account other factors such as historical wind power output that affect the accuracy of wind power forecasting. The historical wind power output fails to accurately depict the spatial correlation of the average absolute error of output between wind farms, which in turn affects the effect of short-term wind power forecasting. As shown in formula (6).

[0078] The adjacency matrix is ​​constructed based on the correlation of the mean absolute error of output among multiple wind farms.

[0079] MAE mn =ReLu(MIC(MAE m ,MAE n )) (6);

[0080] Where MAE m with MMAE n is the mean absolute error between wind farm m and wind farm n. Since the mean absolute errors between wind farms are different, they are used as edge features of the time-space graph.

[0081] S4. By introducing the information propagation mechanism in the space-time graph neural network, the average aggregation function is used to extract the numerical weather forecast information of the historical output-related stations, and an information propagation mechanism is constructed to aggregate the weather forecast data of the historical output-related stations and the high-dimensional features of the historical output, so as to achieve full-domain and full-feature enhancement.

[0082] S4 is specifically: Figure 4 As shown in Figure 1, the forward propagation process of the information propagation mechanism consists of two phases, namely the message passing phase and the readout phase. In the message passing phase, each node and edge aggregates information from its neighbors through an aggregation function. The generated messages are passed to adjacent nodes and edges. During feature update, each node and edge updates its features based on the aggregated messages received from its neighbors. The message passing phase runs for t time steps and depends on the message function M. t And the node update function U t In the message passing phase, the hidden state of each node is According to the news To update, the specific process is as follows:

[0083]

[0084] in, The average aggregation function is used to aggregate wind farm information with strong correlation of mean absolute error. R acts on the set of node states and must keep the arrangement of node states unchanged so that the information propagation mechanism remains unchanged for graph isomorphism. y′ can output all updated hidden states and realize full domain and full feature enhancement of the spatiotemporal graph.

[0085] S5. In the process of training spatiotemporal graph neural network, a loss function combining absolute error loss and cross entropy loss is introduced to optimize the category weight vectors of different small samples and improve the robustness and accuracy of the prediction model under small samples.

[0086] S5 is specifically: Accurate wind power forecasting is crucial for the safe and stable operation of power systems under low-probability events. Using statistical methods to identify extreme wind speed events, most areas experience 5 to 6 occurrences throughout the year, and the amount of data accounts for about 2%, which is a small sample event. Insufficient sample size is a difficulty in machine learning modeling. Extremely high wind speeds cause additional loads, fatigue, damage, and even sudden failures on wind turbines, which reduces wind power output and increases prediction errors. The present invention proposes a weight adjustment method in the cross entropy loss function to optimize the category weight vector of small samples, that is, to pay more attention to small sample data. Let K be the number of samples in each category K. m The vector of class weights W can be calculated as follows:

[0087]

[0088] Using the categorical cross entropy loss function for M classes, the improved L′ CE The loss function can be expressed as:

[0089]

[0090] Where N is the number of training samples, y is the label of the sample, and z is the output of the final layer.

[0091] In order to accurately and effectively predict wind power in low-probability scenarios, a loss function based on the combination of absolute error and cross entropy is proposed. i and f i represent the actual wind power and the predicted wind power respectively. Then, the definition of MAE loss is as follows:

[0092]

[0093] The final model loss function is as follows:

[0094] Loss = λ1L′ CE +λ2L MAE (13);

[0095] Among them, λ1 and λ2 are weighted parameters, and the weighted parameters are set to λ1 = 0.5, λ2 = 0.5. The loss function combining absolute value loss and cross entropy loss is introduced to handle unbalanced small samples in low-probability scenarios. By solving the category weight vectors of different samples, the robustness and accuracy of small sample prediction are improved.

[0096] S6. Combine the gated recurrent unit layer to complete the short-term prediction of multiple wind farms based on the spatiotemporal graph convolutional neural network.

[0097] S6 is specifically: Figure 5 As shown in the figure, in order to make full use of the time-domain features between the time-space diagrams of multiple wind farms after the full-domain full-feature enhancement, a GRUL with a gating mechanism is established. At each moment, the enhanced historical output high-dimensional feature vector output by MPNN (Message Passing Neural Network) is used as the input of GRUL to realize the wind power prediction of multiple wind farms in a wide area. GRUL includes two multiplication gates, the update gate u t and reset gate r t GRUL uses MPNN

[0098] The extracted feature output h t , and the hidden layer output h at time t-1 t-1 As input, complete the prediction, the specific formula is as follows:

[0099] u t =σ(A t [X′ t ,h t-1 ]EW z +Eb u ) (14);

[0100] r t =σ(A t [X′ t ,h t-1 ]EW r +Eb r ) (15);

[0101] c t =tanh(A t [X′ t ,r*h t-1 ]EW c +Eb c ) (16);

[0102] h t =u t *h t-1 +(1-u t )*c t (17);

[0103] Where: σ represents the sigmoid activation function, tanh is the activation function, E, W u , W r , W c , b u , b r and b c are all learnable weight and bias parameters in the model, X′ t represents the node features output by MPNN, A represents the connection relationship between multiple wind farms generated by the adjacency matrix module, and h t-1 represents the output at time t-1, h t Represents the wind power sequence output at time t.

[0104] Example 2

[0105] The present invention provides a wide-area multi-wind farm short-term prediction method based on a time-space graph convolutional neural network, such as Figure 1 As shown, the specific implementation steps are as follows:

[0106] S1. Analyze the correlation between wind power output and meteorological characteristics of a certain region based on the maximum information coefficient, and determine the strongly correlated meteorological characteristics;

[0107] S2, determine the multi-node feature set of the time-space graph through the meteorological characteristics strongly correlated with the output forecast and the historical output of wind power;

[0108] S3, determine the spatial correlation between wind farms through the adjacency matrix, and use the historical output correlation between multiple wind farms as the time-space graph edge feature;

[0109] S4. By introducing the information propagation mechanism in the spatiotemporal graph neural network, an information propagation mechanism is constructed to aggregate the weather forecast data of the station with strong correlation with the historical output and the high-dimensional features of the historical output;

[0110] S5. In the training process of spatiotemporal graph neural network, a loss function combining absolute error loss and cross entropy loss is introduced, and the gated recurrent unit layer is combined to complete the short-term prediction of multiple wind farms.

[0111] Example 3

[0112] Based on the implementation of 2,

[0113] The specific process of S1 and S2 is as follows:

[0114] like Figure 2 and Figure 3 As shown, meteorological characteristics include wind power, wind speed, wind direction, temperature, humidity and air pressure characteristics.

[0115] Taking the meteorological characteristics of a certain region and a certain period of time as variables, any given two variables (T i , Tj ), collect the weather forecast data set D = (x1, y1),...,(x n , y n ) in a certain area and a certain time period, where x i represents the wind power output of the wind farm at the i-th sampling, and y i represents the meteorological characteristics at the i-th sampling. The value of i ranges from 1, 2,..., n, and n represents the total number of samplings. The formula for calculating the mutual information in the α-by-β grid is:

[0116]

[0117] where I(D, α, β) is the mutual information value in the α-by-β grid, the integer (α, β) is a feature pair formed by any one of the wind power output and the weather forecast characteristics, p(x, y) represents the joint probability density under the current grid division, p(x) represents the marginal probability density of the wind power output, and p(y) represents the marginal probability density of the weather forecast characteristics;

[0118] Find the maximum mutual information value in different grids and normalize it. MIC(T i , T j ) is obtained by calculating the maximum value of the mutual information value:

[0119]

[0120] MIC(T i , T h ) = MIC(D) = max αβ＜B(n) {M(D)} (3);

[0121] where log(min{α, β}) is for normalization calculation, and (α, β) must satisfy the prerequisite αβ < B(n), and B(n) can be expressed as B(n) = n0.6. By statistically calculating the daily mean absolute error of the wind farm under different meteorological scenarios, the meteorological characteristics strongly correlated with the mean absolute error of the wind power output prediction are determined according to MIC, that is, wind speed and wind direction.

[0122] Perform feature selection from the meteorological characteristics, that is, use the strongly correlated meteorological characteristics and the historical wind power output to determine the multi-node feature set of the time-space diagram. For the selected features as the multi-node feature set, construct the node features of the time-space diagram. The node features of the time-space diagram can accurately reflect the attributes of the wind farm and provide a basis for the subsequent edge features and the construction of the time-space diagram.

[0123] The specific process of S3 is as follows:

[0124] The neighborhood relationship is defined by the spatial correlation between m and n in the geographical location or historical output data, that is, node m is connected to node n. The adjacency matrix A is a symmetric square matrix of size N, in which each element A(m,n) reflects the spatial correlation between the two wind farms m and n. The adjacency matrix is ​​defined as follows:

[0125] A mn =ReLu(MIC(B m ,B n )) (4);

[0126]

[0127] MIC(B m ,B n ) is used to calculate the correlation between nodes, B m ,B n is the corresponding node vector, and ReLu is the activation function used to avoid negative correlation;

[0128] As shown in formula (6), the adjacency matrix is ​​constructed by the correlation of the mean absolute error of output among multiple wind farms:

[0129] MAE mn =ReLu(MIC(MAE m ,MAE n )) (6);

[0130] Where MAE m with MAE n is the mean absolute error between wind farm m and wind farm n. Since the mean absolute errors between wind farms are different, they are used as edge features of the time-space graph.

[0131] The specific process of S4 is as follows:

[0132] like Figure 4 As shown, the information propagation mechanism is used for message passing. In the message passing phase, each node and edge aggregates information from its neighbors through an aggregation function and passes the generated messages to adjacent nodes and edges. During feature updating, each node and edge updates its features according to the aggregated messages received from its neighbors. The information passing phase runs T time steps and depends on the message function M. t And the node update function U t , in the message passing phase, the hidden state of each node According to the news To update, the specific process is as follows:

[0133]

[0134] in, The average aggregation function is used to aggregate wind farm information with strong correlation of mean absolute error. R acts on the set of node states and must keep the arrangement of node states unchanged so that the information propagation mechanism remains unchanged for graph isomorphism. y′ can output all updated hidden states and realize full-domain and full-feature enhancement of the spatiotemporal graph.

[0135] The specific process of introducing the loss function combining absolute error loss and cross entropy loss in the training process of spatiotemporal graph neural network in S5 is as follows:

[0136] Optimize the class weight vector of small samples, that is, pay more attention to small sample data. Let K be the number of samples in each class. m The vector of class weights W is expressed as follows:

[0137]

[0138] Using the categorical cross entropy loss function for M classes, the improved L′ CE The loss function can be expressed as:

[0139]

[0140] Where N is the number of training samples, y is the label of the sample, and z is the output of the final layer.

[0141] Let x i and f i They represent the actual wind power and the predicted wind power respectively, and the definition of MAE loss is as follows:

[0142]

[0143] The final loss function is as follows:

[0144] Loss = λ1L′ CE +λ2L MAE (13);

[0145] Wherein, λ1 and λ2 are weighting parameters, and the weighting parameters are set to λ1=0.5, λ2=0.5.

[0146] The specific process of combining the gated cyclic unit layer in S5 to complete the short-term prediction of multiple wind farms is as follows:

[0147] like Figure 5 As shown in the figure, a GRUL with a gating mechanism is established. At each moment, the enhanced historical output high-dimensional feature vector output by MPNN is used as the input of GRUL to realize the wind power prediction of multiple wind farms in a wide area. GRUL includes two multiplication gates, the update gate u t and reset gate r t GRUL uses MPNN to extract the feature output ht , and the hidden layer output h at time t-1 t-1 As input, complete the prediction, the specific formula is as follows:

[0148] u t =σ(A t [X′ t ,h t-1 ]EW z +Eb u ) (14);

[0149] r t =σ(A t [X′ t ,h t-1 ]EW r +Eb r ) (15);

[0150] c t =tanh(A t [X′ t ,r*h t-1 ]EW c +Eb c ) (16);

[0151] h t =u t *h t-1 +(1-u t )*c t (17);

[0152] Where: σ represents the sigmoid activation function, tanh is the activation function, E, W u , W r , W c 、b u 、b r and b c are all learnable weight and bias parameters in the model, X′ t represents the node features output by MPNN, A represents the connection relationship between multiple wind farms generated by the adjacency matrix module, and h t-1 represents the output at time t-1, h t Represents the wind power sequence output at time t.

[0153] Comparative Example 1

[0154] The full-year data is used to analyze the impact of information propagation mechanism and average aggregation function on the prediction effect. As shown in Table 1, the range of mean absolute errors of the spatiotemporal graph convolutional neural network (ST-GCN), no information propagation mechanism (noMPNN) and no average function (noMEAN) in the test set are statistically analyzed.

[0155] Table 1 Evaluation indicators with full domain and full feature enhancement and without enhancement

[0156]

[0157] Table 1 shows the evaluation indicators under the condition of full domain and full feature enhancement. Taking the indicators without information propagation mechanism (noMPNN) and without average function (noMEAN) as the benchmark, the absolute value of the mean absolute error of the spatiotemporal graph convolutional neural network is 0.03MW to 0.42MW (that is, the corresponding error percentage decreases by 1.09% to 4.49%). The problem of not fully updating the high-dimensional features of the numerical weather forecast and historical output of the error-strongly correlated station without information propagation mechanism (noMPNN) has failed to achieve full domain and full feature enhancement of the spatiotemporal graph. Although the no-average function (noMEAN) mines the spatiotemporal coupling correlation between wind farm output data from the non-Euclidean domain, it fails to construct an information propagation mechanism for aggregating the high-dimensional features of the numerical weather forecast and historical output of the strongly correlated station. The spatiotemporal graph convolutional neural network uses information propagation mechanism and average aggregation function to fully improve the utilization rate of the numerical weather forecast of the strongly correlated station to be predicted, and the node connection relationship is not constrained by the geographical location between each wind farm, so as to achieve full feature enhancement of the station to be predicted.

[0158] Comparative Example 2

[0159] The accuracy of the prediction method based on the space-time graph convolutional neural network (ST-GCN) of the present invention is compared with other methods. Most studies regard wind power prediction as a time series prediction problem, focusing on mining information in the time dimension, while ignoring the influence of the space-time coupling correlation between different wind farms. In order to further study the prediction performance of the proposed method, the proposed space-time feature learning model is compared with shallow and deep feature extraction techniques. Convolutional Neural Networks-Long Short-Term Memory (CNN-LSTM), GRU, are selected for comparative experiments, in which the node features on the wide-area multi-wind farm space-time graph are determined by the maximum correlation information coefficient, and the edge features of the space-time graph of multiple wind farms are constructed by analyzing the spatial correlation of multiple wind farms. The relevant parameters refer to the parameter determination method of the above-mentioned literature. In the comparison method, the input features are consistent with the node features of the space-time graph convolutional neural network space-time graph.

[0160] As shown in Table 2, the prediction index range in typical scenarios is selected from March 21 to June 20, 2014 (92 days) in spring; June 21 to September 22 (94 days) in summer; September 23 to December 21 (90 days) in autumn; and December 22 to March 20 (89 days) in winter for experiments, and the prediction effects of each index are compared and analyzed. The worst RMSE and MAE of the prediction method proposed in the present invention are significantly improved compared with other methods.

[0161] Table 2 Range of prediction indicators under typical scenarios

[0162]

[0163] Robustness evaluation under low-probability scenarios, taking into account the potential impact of extreme wind speed fluctuation characteristics on the wind speed prediction model of multiple wind farms, the wind speed threshold is determined according to the actual situation of the selected wind farm area. High MAE in the prediction under low-probability scenarios means that the power system requires more flexibility resources and has higher costs. When the average wind speed on a certain day exceeds the 95th percentile threshold of the annual data, the event is classified as an extreme high wind speed event. According to the above principles, after screening from the data of a certain year, there were 7 days of extreme high wind speed days that exceeded the set threshold. As shown in Table 3, the prediction range of multiple wind farms by different methods under low-probability scenarios:

[0164] Table 3 Prediction range of multiple wind farms by different methods under low probability scenarios

[0165]

[0166] As shown in Table 3, the prediction range of GRU, CNN-LSTM, ST-GCN and ST-GCN without absolute error and cross entropy loss function in the test set is that ST-GCN has a maximum absolute error of RMSE reduced by 0.12MW to 1.14MW (relative percentage reduced by 0.89% to 7.85%) compared to GRU and CNN-LSTM without mixed loss function (noloss). At the same time, ST-GCN has a MAE reduction of 0.33MW compared to the MAE without mixed loss function (noloss), and a MAE reduction of 0.36MW and 0.59MW compared to CNN-LSTM and GRU.

[0167] Most studies regard wind power forecasting as a time series forecasting problem, focusing on mining information in the time dimension, while ignoring the influence of the temporal-spatial coupling correlation between different wind farms. Therefore, a temporal-spatial graph convolutional neural network taking into account the spatial correlation of historical wind power output of wide-area multi-wind farms is constructed to carry out wind power forecasting of wide-area multi-wind farms. The constructed temporal-spatial graph of wide-area multi-wind farms can directly reflect the degree of spatial coupling correlation between multiple wind farms, further improving the spatial correlation of the model. First, the multi-node feature set of the temporal-spatial graph is determined by using strongly correlated meteorological characteristics and historical wind power output. At the same time, this paper crosses the constraints of geographical proximity of multiple sites and constructs the temporal-spatial graph edge features with the historical output correlation between multiple wind farms as the spatial correlation. Then, by introducing the information propagation mechanism in the temporal-spatial graph convolutional neural network, the average aggregation function is used to extract the numerical weather forecast information of strongly correlated stations, and the information propagation mechanism that aggregates the numerical weather forecast and historical output high-dimensional features of strongly correlated stations is constructed to achieve full-domain and full-feature enhancement. Finally, a loss function combining absolute error loss and cross entropy loss is introduced in the training process of the spatiotemporal graph convolutional neural network to optimize the category weight vectors of different small samples, and input the gated unit circulation layer to mine its time domain features. On this basis, the short-term wind power prediction of wide-area multi-wind farms with full-domain and full-feature enhancement in time-space is realized. The results of short-term wind power prediction of wide-area multi-wind farms show that compared with the mean absolute error of other methods, as shown in Table 3, the RMSE and MAE absolute value percentages of the prediction method proposed in the present invention are reduced by 0.89% to 7.85% and 3.56% to 6.19.

Claims

1. A wide-area multi-wind farm short-term prediction method based on spatiotemporal graph convolutional neural network, characterized in that: Follow the steps below to implement it: S1. Analyze the correlation between wind power output and meteorological characteristics of a certain region based on the maximum information coefficient, and determine the strongly correlated meteorological characteristics; S2, determine the multi-node feature set of the time-space graph through the meteorological characteristics strongly correlated with the output forecast and the historical wind power output; S3, determine the spatial correlation between wind farms through the adjacency matrix, and use the historical output correlation between multiple wind farms as the time-space graph edge feature; S4. By introducing the information propagation mechanism in the spatiotemporal graph neural network, an information propagation mechanism is constructed to aggregate the weather forecast data of the station with strong correlation with the historical output and the high-dimensional features of the historical output; S5. In the training process of spatiotemporal graph neural network, a loss function combining absolute error loss and cross entropy loss is introduced, and the gated recurrent unit layer is combined to complete the short-term prediction of multiple wind farms.

2. The wide-area multi-wind farm short-term prediction method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: The meteorological characteristics described in S1 include wind speed, wind direction, temperature, humidity and air pressure characteristics.

3. The wide-area multi-wind farm short-term prediction method based on spatiotemporal graph convolutional neural network according to claim 2 is characterized in that: The specific process of S1 is as follows: Taking the meteorological characteristics of a certain region and a certain period of time as variables, any given two variables (T i , T j ), collect the weather forecast data set D=(x1,y1),...,(x n ,y n ), where x i represents the wind power output of the wind farm sampled at the i-th time, y i represents the meteorological characteristics of the i-th sampling, i takes the value of 1…n, and n represents the total number of sampling times. The formula for calculating mutual information in the α-by-β grid is: Where I(D, α, β) is the mutual information value in the α-by-β grid, the integer (α, β) is the feature pair formed by any one of the wind power output and weather forecast features, p(x, y) represents the joint probability density under this grid division, p(x) represents the marginal probability density of wind power output, and p(y) represents the marginal probability density of weather forecast features; Find the maximum mutual information value in different grids and normalize it, MIC(T i , T j ) is obtained by calculating the maximum value of the mutual information value: MIC(T i ,T j )=MIC(D)=max αβ＜B(n) {M(D)} (3); Among them, log(min{α, β}) is a normalized calculation, and (α, β) must satisfy the prerequisite αβ<B(n), B(n) can be expressed as B(n)=n0.

6. By statistically analyzing the daily average absolute error of wind farms under different meteorological scenarios, the meteorological characteristics with strong correlation with the average absolute error of wind power output prediction, namely wind speed and wind direction, are determined according to MIC.

4. The wide-area multi-wind farm short-term prediction method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: The specific process of S3 is as follows: The neighborhood relationship is defined by the spatial correlation between m and n in the geographical location or historical output data, that is, node m is connected to node n. The adjacency matrix A is a symmetric square matrix of size N, in which each element A(m, n) reflects the spatial correlation between two wind farms m and n. The adjacency matrix is ​​defined as follows: A mn =ReLu(MIC(B m ,B n )) (4); MIC(B m , B n ) is used to calculate the correlation between nodes, B m , B n is the corresponding node vector, and ReLu is the activation function used to avoid negative correlation; As shown in formula (6), the adjacency matrix is ​​constructed by the correlation of the mean absolute error of output among multiple wind farms: THERE IS mn =ReLu(MIC(MAE m , THERE IS n )) (6); Where MAE m with MAE n is the mean absolute error between wind farm m and wind farm n. Since the mean absolute errors between wind farms are different, they are used as edge features of the time-space graph.

5. The wide-area multi-wind farm short-term prediction method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: The specific process of S4 is as follows: The information propagation mechanism is used for message passing. In the message passing phase, each node and edge aggregates information from its neighbors through an aggregation function and passes the generated messages to adjacent nodes and edges. During feature updating, each node and edge updates its features according to the aggregated messages received from its neighbors. The information passing phase runs T time steps and depends on the message function M. t And the node update function U t , in the message passing phase, the hidden state of each node According to the news To update, the specific process is as follows: in, The average aggregation function is used to aggregate wind farm information with strong correlation of mean absolute error. R acts on the set of node states and must keep the arrangement of node states unchanged so that the information propagation mechanism remains unchanged for graph isomorphism. y′ can output all updated hidden states and realize full-domain and full-feature enhancement of the spatiotemporal graph.

6. The wide-area multi-wind farm short-term prediction method based on spatiotemporal graph convolutional neural network according to claim 1, characterized in that: The specific process of introducing the loss function combining absolute error loss and cross entropy loss in the training process of spatiotemporal graph neural network described in S5 is as follows: Optimize the class weight vector of small samples, that is, pay more attention to small sample data. Let K be the number of samples in each class. m The vector of class weights W is expressed as follows: Using the categorical cross entropy loss function for M classes, the improved L′ CE The loss function can be expressed as: Where N is the number of training samples, y is the label of the sample, and z is the output of the final layer; Let x i and f i They represent the actual wind power and the predicted wind power respectively, and the definition of MAE loss is as follows: The final loss function is as follows: Loss=λ1L′ CE +λ2L MAE (13); Wherein, λ1 and λ2 are weighting parameters, and the weighting parameters are set to λ1=0.5, λ2=0.

5.

7. The wide-area multi-wind farm short-term prediction method based on spatiotemporal graph convolutional neural network according to claim 2 is characterized in that: The specific process of completing short-term prediction of multiple wind farms by combining the gated cyclic unit layer described in S5 is as follows: A GRUL with a gating mechanism is established. The enhanced historical output high-dimensional feature vector output by MPNN is used as the input of GRUL at each moment to realize the wind power prediction of multiple wind farms in a wide area. GRUL includes two multiplication gates, the update gate u t and reset gate r t GRUL uses MPNN to extract the feature output h t , and the hidden layer output h at time t-1 t-1 As input, the specific formula is as follows: u t =σ(A t [X′ t ,h t-1 ]EW z +Eb u ) (14)? r t =σ(A t [X′ t , h t-1 ] EW r +Eb r ) (15); c t =tanh(A t [X′ t ,r*h t-1 ]EW c +Eb c ) (16); h t =u t *h t-1 +(1-u t )*c t (17); Where: σ represents the sigmoid activation function, tanh is the activation function, E, W u , W r , W c 、b u 、b r and b c are all learnable weight and bias parameters in the model, X′ t represents the node features output by MPNN, A represents the connection relationship between multiple wind farms generated by the adjacency matrix module, and h t-1 represents the output at time t-1, h t Represents the wind power sequence output at time t.

Images