A multi-step wind power forecasting method considering spatiotemporal distribution characteristics

Abstract

The present invention discloses a multi-step prediction method for wind power considering the characteristics of spatiotemporal distribution, including: clustering the wind power data using the acquired wind turbine group data into several groups of wind turbines to extract the spatial distribution characteristics of wind farm power; performing EEMD decomposition on the power sequence of each group of wind turbines to obtain several wind farm power subsequences and extract the time distribution characteristics of the wind power sequence; establishing an encoder-decoder multi-step prediction network model for each subsequence of the group of wind turbines, the network model is composed of two GRU networks, the first network extracts the effective information of the input power sequence and encodes it into a feature vector; the second network decodes the feature vector transmitted by the encoder to obtain a predicted power sequence; reconstructs the predicted power of each group of wind turbines to obtain a prediction curve of the total power of the wind farm. This method ensures the prediction scale and prediction accuracy, and is closer to the actual wind power scheduling task scenario.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a wind power multi-step prediction method considering temporal and spatial distribution characteristics. Background Art

[0002] In recent years, with the increasing depletion of fossil energy and the popularization of low-carbon and environmentally friendly concepts, the role of new energy power generation, especially wind power generation, has become increasingly prominent. As one of the key technologies for wind power grid connection, wind power prediction refers to: combining historical operation data, unit data, geographic information and numerical weather prediction (NWP) data of wind farms, combined with pre-built prediction models, to predict the active power of wind farms in the future. However, wind power generation is greatly affected by meteorological and geographical conditions, and its output is highly random, volatile and intermittent, which will bring great challenges to the accurate prediction of wind power. Therefore, achieving accurate prediction of wind power, thereby assisting in the scientific dispatch and control of wind power, is an important prerequisite and necessary guarantee for promoting large-scale wind power grid connection.

[0003] At present, the main methods for wind power prediction include physical method, statistical method and artificial intelligence method.

[0004] The physical method mainly starts from the physical process of wind power generation, combines the terrain information and NWP data of the wind farm, obtains the wind speed at the hub height, and then converts it into predicted power through the wind speed-power curve. The advantage of the physical method is that it does not require historical operating data, so it is also suitable for newly built wind farms, but it has high requirements for the resolution of terrain information and the accuracy of NWP data.

[0005] Statistical methods are based on the idea of ​​statistics. Starting from historical power data, the inherent statistical laws are studied. Time series analysis is a typical statistical method. Classic time series models include autoregressive models, moving average models, and autoregressive moving average models. These time series analysis models are simple and effective, but they often require stability of the series and require a large amount of historical operating data as data support, which is not suitable for new wind farms.

[0006] The artificial intelligence method uses the adaptability and powerful ability of neural networks to extract nonlinear features, builds a prediction model, extracts the nonlinear functional relationship between wind power and other influencing factors from massive data, and thus realizes power prediction. The earliest application is the back propagation neural network (BPNN). In order to further consider the dynamics of wind power series, some scholars apply recurrent neural network (RNN) to wind power prediction. The long short-term memory network (LSTM) and gated recurrent unit (GRU) network introduce gating mechanism to control the transmission of information, avoiding the "gradient disappearance / explosion" problem of RNN, which is more suitable for processing time series such as wind power. The advantage of artificial intelligence methods is that they can fully explore the nonlinear characteristics between massive data of wind farms, thereby constructing artificial neural networks for accurate prediction. However, since wind power generation is a "black box", the model has poor interpretability or even cannot be explained.

[0007] At present, the above methods for wind power forecasting have their own limitations. The physical method is complex to model, and has high requirements for the resolution of the terrain information of the wind farm and the accuracy of the NWP data, which are difficult to fully meet with existing technologies. The statistical method is simple and has a small amount of calculation, but it has high requirements for the stationarity of the time series and requires a large amount of historical data as a data source for statistical analysis. It is not suitable for newly built wind farms. The theory of artificial intelligence methods is relatively advanced, but the model has poor interpretability, large amount of calculation, and the network training process has certain uncertainties. In addition, the existing wind power forecasting technology rarely considers the impact of the geographical distribution of wind turbines on wind power, thereby ignoring the spatial distribution characteristics of wind farm power; in terms of the forecast time scale, it also focuses on ultra-short-term forecasts, especially single-step forecasts, which cannot meet the needs of actual engineering applications. Summary of the invention

[0008] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a wind power multi-step prediction method considering the temporal and spatial distribution characteristics.

[0009] To achieve the above object, the technical solution of the present invention is:

[0010] A wind power multi-step prediction method considering spatiotemporal distribution characteristics, comprising:

[0011] Using the acquired wind turbine data, the wind turbine data is clustered into several turbine groups to extract the spatial distribution characteristics of wind farm power;

[0012] Perform EEMD decomposition on the power sequence of each unit group to obtain several wind farm power subsequences and extract the time distribution characteristics of the wind power sequence;

[0013] An encoder-decoder multi-step prediction network model is established for each subsequence of the unit group. The network model consists of two GRU networks. The first network extracts the effective information of the input power sequence and encodes it into a feature vector; the second network decodes the feature vector transmitted by the encoder to obtain the predicted power sequence.

[0014] The predicted power of each unit group is reconstructed to obtain the prediction curve of the total power of the wind farm.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] The present invention extracts the spatiotemporal distribution characteristics of wind farm power and then makes predictions, thereby ensuring the prediction scale and prediction accuracy. The multi-step prediction is also closer to the actual wind power scheduling task scenario, while also avoiding the huge workload brought about by modeling a single unit. The prediction form of the combined method also improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flow chart of a wind power multi-step prediction method considering spatiotemporal distribution characteristics provided by an embodiment of the present invention;

[0018] Figure 2 This is the encoder-decoder multi-step prediction model diagram;

[0019] Figure 3 It is a comparison chart of prediction curves of different prediction methods in data set 1;

[0020] Figure 4 It is a comparison chart of prediction curves of different prediction methods in data set 2;

[0021] Figure 5 This is a diagram of the prediction results of the method of the present invention at times T+1, T+6 and T+12. DETAILED DESCRIPTION

[0022] In this application, the relevant abbreviations and key terms are defined as follows:

[0023] Numerical weather prediction: Numerical weather prediction, NWP

[0024] Backpropagation neural network: Backpropagation neural network, BPNN

[0025] Recurrent neural network: Recurrent neural network, RNN

[0026] Long short-term memory network: Long short-term memory, LSTM

[0027] Gated recurrent unit: Gated recurrent unit, GRU

[0028] Empirical mode decomposition: Empirical mode decomposition, EMD

[0029] Ensemble empirical mode decomposition, EEMD

[0030] Encoder-Decoder: Encoder-Decoder, ED

[0031] Root mean square error: Root mean square error, RMSE

[0032] Mean absolute error: MAE

[0033] Example:

[0034] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0035] The present invention aims to propose a combined method for realizing multi-step accurate prediction of wind power. The method combines the advantages of statistical methods and artificial intelligence methods, realizes power prediction after extracting the spatiotemporal distribution characteristics of wind farm power, and the prediction results can meet the needs of actual wind power scheduling tasks. First, the wind turbines are clustered using the k-means clustering method to extract the spatial distribution characteristics of wind farm power. Secondly, the ensemble empirical mode decomposition (EEMD) algorithm is used to extract the characteristics of different time scales in the power sequence, and the original power sequence with large fluctuations and irregular changes is decomposed into several regular subsequences to fully extract the time distribution characteristics of the wind power sequence. Then, an encoder-decoder (ED) prediction network is established for each subsequence of each unit group. The network consists of two GRU networks. The first network uses a deep GRU to fully extract the effective information of the input power sequence and encode it into a feature vector; the second network decodes the feature vector transmitted by the encoder to obtain a predicted power sequence. Finally, the predicted power of each unit group is reconstructed to obtain the prediction curve of the total power of the wind farm. Prediction is performed after extracting the spatiotemporal distribution characteristics of wind farm power, which ensures the prediction scale and accuracy. Multi-step prediction is also closer to the actual wind power scheduling task scenario. At the same time, it also avoids the huge workload brought by modeling a single unit. The prediction form of the combined method also improves the prediction accuracy.

[0036] See also Figure 1 As shown, this method mainly includes the following steps:

[0037] Unit clustering steps

[0038] Considering that wind turbines at different locations in a wind farm are affected differently by factors such as terrain and weather, their outputs will be different. In order to better improve the prediction effect and avoid modeling a single unit, a clustering method is used to cluster wind turbines, thereby extracting the spatial distribution characteristics of wind farm power. The specific calculation process of unit clustering is as follows:

[0039] Step 1: Data preprocessing.

[0040] The original wind speed data is normalized and the calculation formula is as follows:

[0041]

[0042] In the formula, v i，t and v′ i，t are the original wind speed and normalized wind speed of the i-th wind turbine at time t, V i，min = min{v i，1, v i，2 , ..., v i，T} and V i，max =max{v i，1 , v i，2 , ..., v i，T} are the minimum and maximum wind speed values ​​of the i-th wind turbine, and N is the number of wind turbines.

[0043] The same method is used to normalize the original power data.

[0044] Step 2: Determine the characteristic vector of the wind turbine.

[0045] The average wind speed, wind speed standard deviation, average power and power standard deviation observed throughout the year are selected as the characterization of a single wind turbine. Therefore, the characteristic vector x of each wind turbine is i is a 4-dimensional vector, and its expression is as follows:

[0046] x i =[v i，mean , v i，std , p i，mean , p i，std ] T (i=1, 2, ..., N) (2)

[0047] Step 3: k-means clustering.

[0048] 3.1. Determine the distance formula.

[0049] The square of the Euclidean distance is selected as the measure of the distance between the characteristic vectors of wind turbines. The calculation formula is as follows:

[0050] d(x i , x j )=||x i -x j ||2 (3)

[0051] 3.2. The preset number of clusters is k1, k2, ..., k M

[0052] 3.3. Initialization: r = 1

[0053] 3.4. Determine the number of clusters: k = k r

[0054] 3.5. The clustering problem of units is converted into an optimization problem:

[0055]

[0056] In the formula, C r The number of clusters is k r The classification result when is the mean or center of the lth class, and C(i)=l represents the set of all wind turbines belonging to the lth class.

[0057] 3.6. Calculate the silhouette coefficient of the clustering results:

[0058]

[0059] In the formula, a(x i ) represents the feature vector x i The minimum value of the average distance to other types of samples, b(x i ) represents the feature vector x i The average distance between all samples in the same category.

[0060] 3.7. Let r:=r+1, return to 3-4.

[0061] 3.8. Traverse all the preset cluster numbers and select the result with the highest silhouette coefficient as the final cluster result. * A group of units.

[0062]

[0063] Power sequence decomposition steps

[0064] EEMD decomposition is performed on the power sequence of each unit group to obtain the characteristics of different time scales, and the original sequence with large fluctuations is converted into several subsequences with small fluctuations and regularity. The specific calculation process of EEMD decomposition is as follows:

[0065] Step 1: Add different white noise h to the original power sequence x of the g-th unit group i , get N h The sequence y after adding noise i :

[0066] y i =x+h i (i=1,2,...,N h ) (7)

[0067] Step 2: N h The power sequences after adding noise are sequentially decomposed by EMD

[0068] 2-1. Let i = 1

[0069] 2-2. Add noise to the i-th sequence y i Decompose and let s = y i .

[0070] 2-3. Let k = 0,

[0071] 2-4. Find the sequence All maximum and minimum points are fitted into the upper envelope s by interpolation method max and the lower envelope s min

[0072] 2-5. Calculate the mean envelope:

[0073] 2-6. Subtract the original signal from the mean envelope to get the intermediate signal:

[0074] like If the eigenmode function determination condition is met, then is a subsequence component of the original sequence, denoted as imf i，t ; Otherwise, let k:=k+1, and return to 2-4 until the condition is met to obtain the subsequence component. The eigenmode function determination condition is:

[0075]

[0076] in and They are the intermediate signals before and after respectively. When SD is between 0.2 and 0.3, the eigenmode function determination condition is met.

[0077] 2-7. Subtract the component imf from the original sequence i，t , and get the residual signal s r :

[0078] s r =s-imf i，t (9)

[0079] If r When the number of extreme points of is less than the sequence length or is a monotonic function, the EMD decomposition terminates. Otherwise, let s = s r , return to step 2-3.

[0080] Finally get all the M g subsequence components, the decomposition result of the sequence is:

[0081]

[0082] Let i:=i+1, return to 2-2, and continue to perform EMD decomposition on the next sequence after adding noise until i=N h .

[0083] Step 3: N hThe EMD decomposition results of the power sequences after adding noise are averaged to obtain the EEMD decomposition results of the original power sequence x as follows:

[0084]

[0085] Encoder-Decoder Multi-step Prediction Model Steps

[0086] For each EEMD component of each group, an encoder-decoder multi-step prediction model is constructed. The model consists of two GRU networks with the following structure: Figure 2 As shown. The first network is the encoder, which uses a deep GRU to fully extract the effective information of the input sequence and encode it into a feature vector; the second network is the decoder. After receiving the feature vector transmitted by the encoder, the decoder starts working, decodes the feature vector through the GRU unit, and converts it into the required dimension through the fully connected layer to obtain the first power prediction value, and then transmits the prediction value as input to the next GRU unit to obtain the second power prediction value, and so on, and finally completes the prediction of the entire power sequence. The calculation expression of the encoder-decoder multi-step prediction model is as follows:

[0087] h t =f enc (x t ,h t-1 ) (12)

[0088]

[0089]

[0090] In the formula, x t represents the input historical power, T is the input step length, represents the predicted power, f enc 、f dec and f out Represent the encoder, decoder and fully connected layer functions respectively.

[0091] Among them, the GRU network is an RNN with a gating mechanism, which is mainly composed of a reset gate and an update gate. The expressions of the two gating mechanisms are:

[0092] r t =σ(w xr x t +w hr h t-1 +b r ) (15)

[0093] u t =σ(w xu x t +whu h t-1 +b u ) (16)

[0094] In the formula, σ represents the Sigmoid activation function, x t is the input at the current moment, h t-1 is the hidden state of the previous moment, b r and b u All are biased.

[0095] The hidden state is the recurrent variable of GRU, and its calculation formula is as follows:

[0096]

[0097]

[0098] In the formula, represents the candidate state of the hidden layer, h t represents the hidden layer state, ⊙ represents the dot product of vector elements, and Tanh represents the hyperbolic tangent activation function;

[0099] The specific expressions of Sigmoid activation function and Tanh activation function are as follows:

[0100]

[0101]

[0102] The following is a simulation example to further verify this method:

[0103] Data Source

[0104] The test was conducted using real data from a wind farm, which consists of 23 wind turbines with a rated power of 1.5 MW and a rated installed capacity of 34.5 MW. The dataset contains multi-location wind speed data and power data from different wind turbines, with a time span of 0:00 on February 1, 2019 to 23:50 on November 30, 2019, and a time resolution of 10 minutes.

[0105] Python 3.6 is used as the programming language. The deep learning architecture is based on the Pytorch framework and the Scikit-learn algorithm. The program running environment is the Intel i9-10900 processor and the Nvidia-Quadro-P1000 graphics card. The drawing tool uses the matplotlib drawing module.

[0106] Example Results

[0107] The original data set is divided into two sample subsets. The first sample subset is the data from February to May 2019 (dataset 1), and the second sample subset is the data from August to November 2019 (dataset 2). Each sample subset is divided into training set, validation set and test set in a ratio of 7:1:2. The historical power of the 12 moments before the current moment is used as the input sequence of the model to predict the wind power of the next 12 moments. At the same time, the prediction results are compared with the traditional prediction methods based on BPNN model and GRU model.

[0108] The prediction performance of each model on dataset 1 and dataset 2 are as follows: Figure 3 and Figure 4 The comparison results of the evaluation indicators are shown in Table 1. Figure 3 and Figure 4 It can be seen that the predicted power curve of the proposed model is basically consistent with the true curve, while the predicted power curves based on the BPNN and GRU models have obvious errors. From the perspective of prediction error, on dataset 1, the RMSE of the prediction results of the proposed model is 0.0376, which is 68.18% and 66.48% lower than that of BPNN and GRU respectively; on dataset 2, the RMSE of the prediction results of the proposed model is 0.0287, which is 73.15% and 70.95 lower than that of BPNN and GRU respectively. From the perspective of prediction accuracy, the accuracy of the proposed model on the two datasets is 96.24% and 97.13% respectively, which is 9.14% and 8.76% higher than that of the BPNN model; and 8.40% and 7.78% higher than that of the GRU model. The test results show that the proposed method has greatly improved compared with the traditional prediction model in the multi-step power prediction of 2 hours in advance.

[0109] Table 1 Comparison of prediction errors of different prediction methods

[0110]

[0111] Table 2 shows the prediction errors of each prediction model at each prediction step. The prediction power curves of the two data sets at T+1 (one step ahead), T+6 (first hour), and T+12 (second hour) are shown in Figure 5As shown in Table 2, as the number of prediction steps increases, the prediction error also gradually increases. The RMSE on dataset 1 is 0.0221-0.0540, the MAE is 0.0162-0.0384, and the correlation coefficient R is 0.956-0.993; the RMSE on dataset 2 is 0.0123-0.0403, the MAE is 0.0085-0.0309, and the correlation coefficient R is 0.987-0.999. The prediction accuracy of the proposed model in the first hour (T+6) and the second hour (T+12) exceeded 94%, and the correlation coefficient R was above 0.95.

[0112] Table 2 Error statistics of the proposed method at each prediction step

[0113]

[0114]

[0115] Prediction results under different prediction scenarios

[0116] In order to further verify the prediction performance of the proposed method in different application scenarios, the prediction performance of various methods in four scenarios of 1h ahead, 2h ahead, 3h ahead and 4h ahead is verified, and the prediction results are given in Table 3. It can be seen from Table 3 that with the increase of prediction time, the prediction error of each model has increased, but the prediction performance of the proposed model is better than other models in each scenario, with RMSE of 0.0193~0.0435, correlation coefficient R is also above 0.98, accuracy is 95.65%~98.07%, and the prediction accuracy is 7.36%~12.18% higher than BPNN and 6.45%~11.88% higher than GRU. This shows that the proposed method has good performance in different prediction scenarios and can be applied to ultra-short-term prediction of wind power.

[0117] Table 3 Prediction errors of various methods in different prediction scenarios

[0118]

[0119] In summary, compared with the prior art, the present invention has the following advantages:

[0120] 1) The proposed method uses historical operation data for prediction, without the need to collect additional NWP data and terrain data, thus reducing the workload of data collection.

[0121] 2) The proposed method belongs to the category of multi-step prediction. Compared with single-step prediction, it can provide more information for wind power scheduling and is closer to the actual wind power scheduling task scenario.

[0122] 3) The proposed method has high prediction accuracy and strong stability. The accuracy of the prediction task of 1 to 4 hours in advance on both datasets is above 95%, and the prediction error is much lower than that of BPNN and GRU.

[0123] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-step wind power prediction method considering spatiotemporal distribution characteristics, characterized in that: include: Using the acquired wind turbine data, the wind turbine data is clustered into several turbine groups to extract the spatial distribution characteristics of wind farm power; Perform EEMD decomposition on the power sequence of each unit group to obtain several wind farm power subsequences and extract the time distribution characteristics of the wind power sequence; An encoder-decoder multi-step prediction network model is established for each subsequence of the unit group. The network model consists of two GRU networks. The first network is used to extract the effective information of the input power sequence and encode it into a feature vector; the second network decodes the feature vector transmitted by the encoder to obtain the predicted power sequence. Reconstruct the predicted power of each unit group to obtain the predicted curve of the total power of the wind farm; The wind turbine data includes wind turbine power and wind speed data; The method of clustering the wind turbine group data using the acquired wind turbine group data into a plurality of wind turbine group groups includes: Step 1: Normalize the wind turbine power and wind speed data respectively; Step 2: Determine the characteristic vector of the wind turbine The average wind speed, wind speed standard deviation, average power and power standard deviation observed throughout the year are selected as the characterization of a single wind turbine. Therefore, the characteristic vector xi of each wind turbine is a 4-dimensional vector, and its expression is as follows: x i =[v i，mean ,v i，std ,P i，mean ,p i，std ] T i=1,2,...,N Step 3: k-means clustering: 3-1. Determine the distance formula The square of the Euclidean distance is selected as the measure of the distance between the characteristic vectors of wind turbines. The calculation formula is as follows: 3-2. The preset number of clusters is k1, k2, …, k M ; 3-3. Initialization: r = 1; 3-4. Determine the number of clusters: k = k r ; 3-5. Convert the unit clustering problem into an optimization problem: In the formula, C r The number of clusters is k r The classification result when is the mean or center of the lth class, C l represents the set of all wind turbines belonging to the lth class; 3-6. Calculate the silhouette coefficient of the clustering results: In the formula, a(x i ) represents the feature vector x i The minimum value of the average distance to other types of samples, b(x i ) represents the feature vector x i The average distance between all samples in the same category; 3-7. Let r:=r+1, and return to step 3-4; 3-8. Traverse all the preset cluster numbers and select the result with the highest silhouette coefficient as the final cluster result; the original N wind turbines are clustered into k * A group of units; The EEMD decomposition of the power sequence of each unit group includes: Step 1: Add different white noise h to the original power sequence x of the g-th unit group i , get N h The sequence y after adding noise i : y i =x+h i i=1,2,...,N h Step 2: N h The power sequences after adding noise are subjected to EMD decomposition in turn; Step 3: N h The EMD decomposition results of the power sequences after adding noise are averaged to obtain the EEMD decomposition result of the original power sequence x.

2. The wind power multi-step prediction method considering the temporal and spatial distribution characteristics according to claim 1, characterized in that: The wind speed data is normalized and the calculation formula is as follows: In the formula, v i,t and v' i,t are the original wind speed and normalized wind speed of the i-th wind turbine at time t, V i,min =min{v i,1 ,v i,2 ,…,v i,T } and V i,max =max{v i,1 ,v i,2 ,…,v i,T } are the minimum and maximum wind speed values ​​of the i-th wind turbine, and N is the number of wind turbines.

3. The wind power multi-step prediction method considering the temporal and spatial distribution characteristics according to claim 1, characterized in that: In the second step, N h The EMD decomposition of the power sequence after adding noise includes the following steps: 2-1. Let i = 1; 2-2. Add noise to the i-th sequence y i Decompose and let s = y i ; 2-3. Let k = 0, 2-4. Find the sequence All maximum and minimum points are fitted into the upper envelope s by interpolation method max and the lower envelope s min ; 2-5. Calculate the mean envelope: 2-6. Subtract the original signal from the mean envelope to get the intermediate signal: like If the eigenmode function determination condition is met, then is a subsequence component of the original sequence, denoted as imf i,t ; Otherwise, let k:=k+1, and return to step 2-4 until the condition is met to obtain the subsequence component; wherein the eigenmode function determination condition is: in and They are the intermediate signals of the previous and the next two times, respectively. When SD is between 0.2 and 0.3, the eigenmode function judgment condition is met; 2-7. Subtract the component imf from the original sequence i,t , and get the residual signal s r : s r =s-imf i,t If r When the number of extreme points of is less than the sequence length or is a monotonic function, the EMD decomposition terminates; otherwise, let s = s r , return to step 2-3; Finally get all the M g subsequence components, the decomposition result of the sequence is: Let i:=i+1, return to step 2-2, and continue to perform EMD decomposition on the next sequence after adding noise until i=N h .

4. The wind power multi-step prediction method considering the temporal and spatial distribution characteristics according to claim 3 is characterized in that: The EEMD decomposition result of the original power sequence x is:

5. The wind power multi-step prediction method considering spatiotemporal distribution characteristics according to claim 1, characterized in that: The calculation expression of the encoder-decoder multi-step prediction network model is as follows: h t =f enc (x t ,h t-1 ) In the formula, x t represents the input historical power, T is the input step length, represents the predicted power, f enc 、f dec and f out Represent the encoder, decoder and fully connected layer functions respectively.

6. The wind power multi-step prediction method considering the temporal and spatial distribution characteristics according to claim 5, characterized in that: The encoder and decoder are both GRU networks. The GRU network is an RNN with a gating mechanism, which is mainly composed of a reset gate and an update gate. The expressions of the two gating mechanisms are: r t =σ(w xr x t +w hr h t-1 +b r ) u t =σ(w xu x t +w hu h t-1 +b u ) In the formula, σ represents the Sigmoid activation function, x t is the input at the current moment, h t-1 is the hidden state of the previous moment, b r and b u All are biased.

7. The wind power multi-step prediction method considering the temporal and spatial distribution characteristics according to claim 6, characterized in that: The hidden state is the recurrent variable of GRU, and its calculation formula is as follows: In the formula, represents the candidate state of the hidden layer, h t represents the hidden layer state, ⊙ represents the dot product of vector elements, and Tanh represents the hyperbolic tangent activation function; The specific expressions of Sigmoid activation function and Tanh activation function are as follows:

Images