A large-scale wind power cluster ultra-short-term wind power prediction method

Abstract

The application is a large-scale wind power cluster ultra-short-term wind power prediction method, which has the characteristics of including a wind power cluster division method based on fuzzy clustering, a wind power cluster power prediction method based on a space-time attention neural network, a simulation calculation and error analysis step. Firstly, a statistical feature extraction method is used to construct the fluctuation characteristics of the wind farm, and a fuzzy clustering algorithm is used to construct a wind power cluster division model; then, a space-time attention network is constructed to extract the space-time characteristics of the wind power data, and a large-scale wind power cluster power prediction model is trained; finally, the stability of the wind power prediction performance is analyzed by using a comprehensive index evaluation system. It has the advantages of scientific and reasonable, clear physical meaning, fast and efficient, simple and practical, higher precision.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction, and is a method for predicting the ultra-short-term wind power of large-scale wind power clusters. Background Technology

[0002] In existing technologies, short-term wind power forecasting is based on the output power of a single wind farm. However, from the perspective of grid dispatching, simple wind farm power forecasting can no longer meet the needs of system dispatching. On the one hand, dispatchers pay more attention to the magnitude of uncertain power in the entire system, rather than individual wind farms, when arranging operating modes and spinning reserves. On the other hand, the increase in wind power penetration makes real-time grid operation dispatching and tie-line power control increasingly difficult. Therefore, power forecasting for wind power clusters is more conducive to the formulation of dispatching plans.

[0003] The technical problem that urgently needs to be solved in this field is:

[0004] (1) Current wind power cluster forecasting methods are mostly geared towards small and medium-sized wind power clusters, while forecasting methods for large-scale wind power clusters have not been fully developed. In fact, the power forecasting results of large-scale wind power bases are of greater significance for regional power grid dispatch.

[0005] (2) Unlike the power prediction mode of a single wind farm, the input data for cluster power prediction is a data set composed of data provided by a single wind farm. The divergence of its data makes it difficult to directly characterize the output characteristics of the wind power cluster. It is necessary to further extract the mode that characterizes the output of the wind power cluster from the data set of a single farm.

[0006] (3) Large-scale wind power clusters have a large number of wind farms, and their modeling efficiency decreases as the number of wind farms increases. Therefore, how to improve modeling efficiency and accelerate the convergence of prediction models is a key issue. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a scientific, reasonable, physically meaningful, fast, efficient, simple, practical, and highly accurate method for predicting the ultra-short-term wind power of large-scale wind power clusters, which can reflect the dynamic characteristics of the system and track future power trends.

[0008] The technical solution adopted to achieve the purpose of this invention is: a method for ultra-short-term wind power prediction of large-scale wind power clusters, characterized by the following steps:

[0009] 1) Wind power cluster partitioning method based on fuzzy clustering

[0010] By introducing the concept of membership degree into fuzzy clustering, each sample belongs to different clusters with a certain membership degree. Based on this, the samples are divided into several disjoint subsets, and the observed data matrix is ​​denoted as Equation (1):

[0011] X = [x1, x2, ..., x s (1)

[0012] In the formula, x j =[x i1 ,x i2 ,,x ib ], x ib Let be the b-th observation result of the i-th sample, where i = 1, 2, ..., n, and s is the number of variables;

[0013] Define the objective function as equation (2):

[0014]

[0015] In the formula, V represents the cluster center matrix, where V = [V1, V2, ..., V... c ], c represents the number of cluster centers, U = (u jk ) c×M Let u be the membership matrix. jk For the k-th sample x k The membership degree of belonging to class j, d jk =||x k -v j || Represents sample x k With cluster center v j The Euclidean distance, where M is the number of samples;

[0016] Fuzzy clustering is based on initializing cluster centers. By continuously updating the membership matrix and cluster centers, and with the objective function minimization as the optimization goal, the final classification is obtained. The cluster center and membership update methods are Equations (3)-(4):

[0017]

[0018]

[0019] In the formula, l represents the number of iterations;

[0020] Representative input features are the fundamental condition for ensuring algorithm convergence. By constructing statistical features, the fluctuation characteristics of wind farms are characterized. Using the fluctuation characteristics of each wind farm as input and fuzzy clustering algorithm as classifier, the division of wind power clusters is realized. The wind farm fluctuation feature construction scheme is as follows: Equation (5)-Equation (10):

[0021] ①The first statistical feature is the per-unit value of the maximum output of the wind farm, which represents the maximum output level of the wind farm;

[0022]

[0023] In the formula, n represents the number of fans, f represents the per-unit value, and P represents the power of the fans. This represents the per-unit value indicating the maximum output of a wind farm;

[0024] ②The second statistical feature is the maximum wind speed of the wind farm, which reflects the maximum wind speed of the wind farm;

[0025] V max =max{V1,V2,...,V n}, n>0 (6)

[0026] In the formula, V max V is the maximum wind speed of the wind farm. i Let n represent the wind speed of the i-th fan, and n represent the number of fans.

[0027] ③ The third statistical feature is the power variance of the wind farm, which reflects the power fluctuation of the wind farm;

[0028]

[0029] In the formula, Let P be the power variance of the wind farm. i Let P represent the power of the i-th wind turbine, P represent the average power of the wind farm, and n represent the number of wind turbines.

[0030] ④ The fourth statistical characteristic is the wind speed variance of the wind farm, which reflects the fluctuation of wind speed in the wind farm;

[0031]

[0032] In the formula, V is the variance of wind speed in a wind farm. i Let i represent the wind speed of the i-th fan. The average wind speed of the wind farm is represented by n, and the number of wind turbines is represented by n.

[0033] ⑤ The fifth statistical characteristic is the average wind speed, which reflects the distribution of average wind speed;

[0034]

[0035] In the formula, V represents the average wind speed. i Let n represent the wind speed of the i-th fan, and n represent the number of fans.

[0036] ⑥ The sixth statistical characteristic is the per-unit value of average power, which reflects the distribution of average power;

[0037]

[0038] In the formula, P represents the average power per unit, f represents the calculated per unit value.i Let represent the power of the i-th fan, and n represent the number of fans;

[0039] 2) Wind power cluster power prediction method based on spatiotemporal attention neural network

[0040] Given a multivariate time series X = [X] with N features 1 ,X 2 ,,X N ] T ∈R N×k k represents the length of the input sequence. This represents the input time series of the i-th input variable;

[0041] For each feature Spatial embedding is calculated using a feedforward neural network; for the input data X = [X 1 ,X 2 ,,X N ] T The embedding calculation result of all input variables is D = [D 1 D 2 ,,D N ] T For time step input X = [X1, X2, ..., X...] k The time-step embedding H = [h1, h2, ..., h] is calculated by the coding layer consisting of two LSTM layers. k ];

[0042] A feedforward neural network is used as the alignment model to calculate spatial attention weights. For the output time step j, the i-th spatial attention weight Equations (11)-(12):

[0043]

[0044]

[0045] Where [h′ j-1 ;d i ]∈R p+m , h′ j-1 ∈R p d represents the hidden state of the LSTM decoder. i ∈R m The spatial embedding represents the i-th feature. For attention weights, b e W is the bias value, and W is the weight value. The spatial context vector g is further calculated based on the spatial attention weights. j ;

[0046] For the output time step j, the alignment model is used to calculate the relationship with h. t Corresponding time attention weights It's about h t and the previous hidden layer state h′ of the decoder j-1 The function is given by equation (13):

[0047]

[0048] Among them, b a Where h′ is the bias value, W is the weight value, and [h′] is the weight value. j-1 h t ]∈R p+m Further based on the spatiotemporal context vector s j Calculate attention weights Spatial and temporal attention depend on the hidden layer states of the LSTM decoder, i.e., LSTM G and LSTM S In the computation space context vector g j At that time, the hidden layer state and neuron state of LSTM are defined as h′, respectively. G and c′ G For the output time step j, the input to the spatial attention is the decoder hidden layer state h′. G,j-1 The calculated g j , will r G,j and the previous time step By splicing Input to LSTM G To update the hidden layer state h′ G,j b G W is the bias value. G The weight values ​​are given by equation (14):

[0049]

[0050] For the LSTM time context vector s j The hidden layer and neuron states of the LSTM are h′, respectively. S ∈R p and c′ S ∈R p At time step j, h′ S,j-1 Represents the temporal attention used to compute the temporal context vector, concatenated. and r S,j And input it into LSTMS to update the hidden layer state h′ S,j b S W is the bias value. S The weight values ​​are given by equation (15):

[0051]

[0052] Before prediction, the final step is to connect the hidden state updates of the two LSTMs to [h′]. G,j h′ S,j ];

[0053] Using the STAN model as the predictor, with the current time t and the number of features k, each input feature is taken 16 steps backward, i.e., the features at time t-15t are used as input. Based on spatiotemporal correlation feature mining, the ultra-short-term wind power of a large-scale wind power cluster is predicted, and the prediction is iv:

[0054] i. Construct statistical characteristics based on meteorological and historical power data to characterize the fluctuation characteristics of wind farms, and use the K-Means algorithm to cluster the input features to achieve the division of wind power clusters;

[0055] ii. Principal component analysis algorithm is used to reduce the dimensionality of wind speed, temperature and humidity of each wind power cluster, and a spatial feature set is constructed by combining the average values ​​of wind speed, temperature, humidity and pressure;

[0056] iii. Construct a spatiotemporal attention network, divide it into training and test sets, train the spatiotemporal attention network on the training set, and obtain an ultra-short-term power prediction model;

[0057] iv. Overlay the forecast results of each wind power cluster to reconstruct the power forecast results of the wind power base;

[0058] v. Combining a comprehensive index evaluation system, conduct a comprehensive evaluation of the accuracy and stability of the ultra-short-term power prediction results for large-scale wind power clusters;

[0059] 3) Simulation calculation

[0060] Simulation input: Analyze the measured data of the electric field to determine the total installed capacity of the wind farm; Input data: historical power of the wind farm, wind speed, temperature and humidity NWP data of the wind farm; Data sampling interval is 15 min; According to steps 1) to 2), the ultra-short-term wind power prediction results for the prediction period are obtained;

[0061] 4) Error Analysis

[0062] The root mean square error is calculated using equation (16):

[0063]

[0064] In the formula, y i Represents the actual power at time i. The predicted power at time i represents the test set length, and Cap represents the actual wind farm's operating capacity. Since the actual wind farm's operating capacity is difficult to obtain, the installed capacity is set to the actual wind farm's operating capacity.

[0065] The mean absolute error is calculated using equation (17):

[0066]

[0067] In the formula, R mae y represents the mean absolute error. i y represents the actual power at time i. i Represents the predicted power at time i;

[0068] The pass rate is calculated using formula (12):

[0069]

[0070] In the formula, B i B represents the pass rate assessment standard. i =0 means that the prediction for sample i is qualified, B i =1 indicates that the prediction of sample i is qualified. In addition, the maximum positive error and the minimum negative error are used to evaluate extreme error events. The error is calculated by equation (19):

[0071] err(t)=y(t)-y pred (t) (19)

[0072] In the formula, err(t) represents the error value, y(t) represents the actual value, and y pred (t) represents the predicted value;

[0073] The maximum positive error is calculated using equation (20):

[0074] MPE = argmax(err(t)) (20)

[0075] In the formula, err(t) represents the error value, and MPE represents the maximum positive error;

[0076] The minimum negative error is calculated using equation (21):

[0077] MNE=argmin(err(t)) (21)

[0078] In the formula, MNE represents the minimum negative error;

[0079] Input the simulation input in step 3), and calculate the error between the predicted power calculated by the model and the measured power using the error evaluation standard formulas (16) and (21) in step 4) to obtain the root mean square error, mean absolute error, pass rate, maximum positive error and minimum negative error of the prediction.

[0080] This invention presents a method for ultra-short-term wind power prediction in large-scale wind power clusters. It effectively identifies the power output characteristics of wind farms and rationally divides wind farm clusters. Principal component analysis (PCA) extracts spatial features of wind power clusters, reduces the dimensionality of input data, accelerates algorithm training, and promotes algorithm convergence. The STAN model enables rapid modeling and prediction of regional wind power. It reflects the dynamic characteristics of the system and tracks future power trends. This method has the advantages of being scientifically sound, having clear physical meaning, being fast and efficient, simple and practical, and having higher accuracy. The predicted RMSE on the test set in the 4th hour was 0.07631, with an extreme error of approximately 25% of the installed capacity. The prediction pass rate was close to 100%, demonstrating stable prediction performance. Attached Figure Description

[0081] Figure 1 A schematic diagram illustrating the calculation of spatiotemporal attention weights for time step j;

[0082] Figure 2 Spatiotemporal feature matrix diagram;

[0083] Figure 3 A schematic diagram of the wind power cluster prediction framework;

[0084] Figure 4 A schematic diagram showing the contribution rates of principal components with different characteristics;

[0085] Figure 5 For predicting the power curve;

[0086] Figure 6 This is a diagram illustrating extreme error events. Detailed Implementation

[0087] The following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates a method for predicting ultra-short-term wind power in large-scale wind power clusters according to the present invention.

[0088] Combination Figures 1-3 The present invention provides a method for predicting ultra-short-term wind power in large-scale wind power clusters, comprising the following steps:

[0089] 1) Wind power cluster partitioning method based on fuzzy clustering

[0090] By introducing the concept of membership degree into fuzzy clustering, each sample belongs to different clusters with a certain membership degree. Based on this, the samples are divided into several disjoint subsets, and the observed data matrix is ​​denoted as Equation (1):

[0091] X = [x1, x2, ..., x s (5)

[0092] In the formula, x j =[x i1 ,x i2 ,,xib ], x ib Let be the b-th observation result of the i-th sample, where i = 1, 2, ..., n, and s is the number of variables.

[0093] Define the objective function as equation (2):

[0094]

[0095] In the formula, V represents the cluster center matrix, where V = [V1, V2, ..., V... c ], c represents the number of cluster centers, U = (u jk ) c×M Let u be the membership matrix. jk For the k-th sample x k The membership degree of belonging to class j, d jk =||x k -v j || Represents sample x k With cluster center v j The Euclidean distance, where M is the number of samples.

[0096] Fuzzy clustering is based on initializing cluster centers. By continuously updating the membership matrix and cluster centers, and with the objective function minimization as the optimization goal, the final classification is obtained. The cluster center and membership update methods are Equations (3)-(4):

[0097]

[0098]

[0099] In the formula, l represents the number of iterations.

[0100] Representative input features are the fundamental condition for ensuring algorithm convergence. By constructing statistical features, the fluctuation characteristics of wind farms are characterized. Using the fluctuation characteristics of each wind farm as input and fuzzy clustering algorithm as classifier, the division of wind power clusters is realized. The wind farm fluctuation feature construction scheme is as follows: Equation (5)-Equation (10):

[0101] ①The first statistical feature is the per-unit value of the maximum output of the wind farm, which represents the maximum output level of the wind farm;

[0102]

[0103] In the formula, n represents the number of fans, f represents the per-unit value, and P represents the power of the fans. This represents the per-unit value indicating the maximum output of a wind farm;

[0104] ②The second statistical feature is the maximum wind speed of the wind farm, which reflects the maximum wind speed of the wind farm;

[0105] V max =max{V1,V2,...,V n}, n>0 (6)

[0106] In the formula, V max V represents the maximum wind speed of the wind farm, V represents the wind speed of the wind turbines, and n represents the number of wind turbines.

[0107] ③ The third statistical feature is the power variance of the wind farm, which reflects the power fluctuation of the wind farm;

[0108]

[0109] In the formula, Let P be the power variance of the wind farm. i Let represent the power of the i-th wind turbine. This represents the average power of the wind farm, and n represents the number of wind turbines.

[0110] ④ The fourth statistical characteristic is the wind speed variance of the wind farm, which reflects the fluctuation of wind speed in the wind farm;

[0111]

[0112] In the formula, V is the variance of wind speed in a wind farm. i Let i represent the wind speed of the i-th fan. The average wind speed of the wind farm is represented by n, and the number of wind turbines is represented by n.

[0113] ⑤ The fifth statistical characteristic is the average wind speed, which reflects the distribution of average wind speed;

[0114]

[0115] In the formula, V represents the average wind speed. i Let n represent the wind speed of the i-th fan, and n represent the number of fans.

[0116] ⑥ The sixth statistical characteristic is the per-unit value of average power, which reflects the distribution of average power;

[0117]

[0118] In the formula, P represents the average power per unit, f represents the calculated per unit value. i Let n represent the power of the i-th wind turbine, and n represent the number of wind turbines.

[0119] Using six statistical features of 159 wind farms as input, a fuzzy clustering algorithm was employed to divide them into several wind power clusters. 2) Wind power cluster power prediction method based on spatiotemporal attention neural network.

[0120] Given a multivariate time series X = [X] with N features 1 ,X 2 ,,X N ] T ∈R N×k k represents the length of the input sequence. This represents the input time series of the i-th input variable;

[0121] For each feature Spatial embedding is calculated using a feedforward neural network; for the input data X = [X 1 ,X 2 ,,X N ] T The embedding calculation result of all input variables is D = [D 1 D 2 ,,D N ] T For time step input X = [X1, X2, ..., X...] k The time-step embedding H = [h1, h2, ..., h] is calculated by the coding layer consisting of two LSTM layers. k ].

[0122] A feedforward neural network is used as the alignment model to calculate spatial attention weights. For the output time step j, the i-th spatial attention weight Equations (11)-(12):

[0123]

[0124]

[0125] Where [h′ j-1 ;d i ]∈R p+m , h′ j-1 ∈R p d represents the hidden state of the LSTM decoder. i ∈R m The spatial embedding represents the i-th feature. For attention weights, b e W is the bias value, and W is the weight value. The spatial context vector g is further calculated based on the spatial attention weights. j .

[0126] For the output time step j, the alignment model is used to calculate the relationship with h. t Corresponding time attention weights It's about h t and the previous hidden layer state h′ of the decoder j-1The function is given by equation (13):

[0127]

[0128] Among them, b a Where h′ is the bias value, W is the weight value, and [h′] is the weight value. j-1 h t ]∈R p+m Further based on the spatiotemporal context vector s j Calculate attention weights Spatial and temporal attention depend on the hidden layer states of the LSTM decoder, i.e., LSTM G and LSTM S In the computation space context vector g j At that time, the hidden layer state and neuron state of LSTM are defined as h′, respectively. G and c′ G For the output time step j, the input to the spatial attention is the decoder hidden layer state h′. G,j-1 The calculated g j , will r G,j and the previous time step By splicing Input to LSTM G To update the hidden layer state h′ G,j b G W is the bias value. G The weight values ​​are given by equation (14):

[0129]

[0130] For the LSTM time context vector s j The hidden layer and neuron states of the LSTM are h′, respectively. S ∈R p and c′ S ∈R p At time step j, h′ S,j-1 Represents the temporal attention used to compute the temporal context vector, concatenated. and r S,j And input into LSTM S Update hidden layer state h′ S,j b S W is the bias value. S The weight values ​​are given by equation (15):

[0131]

[0132] Before prediction, the final step is to connect the hidden state updates of the two LSTMs to [h′]. G,j h′ S,j].

[0133] Using the STAN model as the predictor, let the current time be t and the number of features be k. Each input feature is taken 16 steps backward, that is, the features from time t-15 to time t are used as input.

[0134] like Figure 3 As shown, the ultra-short-term wind power of a large-scale wind power cluster is predicted based on spatiotemporal correlation feature mining, and the prediction is iv:

[0135] i. Construct statistical characteristics based on meteorological and historical power data to characterize the fluctuation characteristics of wind farms, and use the K-Means algorithm to cluster the input features to achieve the division of wind power clusters;

[0136] ii. Principal component analysis algorithm is used to reduce the dimensionality of wind speed, temperature and humidity of each wind power cluster, and a spatial feature set is constructed by combining the average values ​​of wind speed, temperature, humidity and pressure;

[0137] iii. Construct a spatiotemporal attention network, divide it into training and test sets, train the spatiotemporal attention network on the training set, and obtain an ultra-short-term power prediction model;

[0138] iv. Overlay the forecast results of each wind power cluster to reconstruct the power forecast results of the wind power base;

[0139] v. Combining a comprehensive index evaluation system, conduct a comprehensive evaluation of the accuracy and stability of the ultra-short-term power prediction results for large-scale wind power clusters.

[0140] 3) Simulation calculation

[0141] Simulation input: Analyze the measured data of the electric field to determine the total installed capacity of the wind farm; Input data: historical power of the wind farm, wind speed, temperature and humidity NWP data of the wind farm; Data sampling interval is 15 min; According to steps 1) to 2), the ultra-short-term wind power prediction results for the prediction period are obtained.

[0142] 4) Error Analysis

[0143] The root mean square error is calculated using equation (16):

[0144]

[0145] In the formula, y i Represents the actual power at time i. represents the predicted power at time i, n represents the test set length, and Cap represents the actual wind farm operating capacity. Since the actual wind farm operating capacity is difficult to obtain, the installed capacity is set as the actual wind farm operating capacity.

[0146] The mean absolute error is calculated using equation (17):

[0147]

[0148] In the formula, R mae y represents the mean absolute error. i y represents the actual power at time i. i This represents the predicted power at time i.

[0149] The pass rate is calculated using formula (12):

[0150]

[0151] In the formula, B i B represents the pass rate assessment standard. i =0 means that the prediction for sample i is qualified, B i =1 means that the prediction for sample i is qualified.

[0152] Furthermore, the maximum positive error and the minimum negative error are used to evaluate extreme error events, and the error is calculated using equation (19):

[0153] err(t)=y(t)-y pred (t) (19)

[0154] In the formula, err(t) represents the error value, y(t) represents the actual value, and y pred (t) represents the predicted value.

[0155] The maximum positive error is calculated using equation (20):

[0156] MPE = argmax(err(t)) (20)

[0157] In the formula, err(t) represents the error value, and MPE represents the maximum positive error.

[0158] The minimum negative error is calculated using equation (21):

[0159] MNE=argmin(err(t)) (21)

[0160] In the formula, MNE represents the minimum negative error.

[0161] Input the simulation input in step 3), and calculate the error between the predicted power calculated by the model and the measured power using the error evaluation standard formulas (16) and (21) in step 4) to obtain the root mean square error, mean absolute error, pass rate, maximum positive error and minimum negative error of the prediction.

[0162] This invention discloses a method for ultra-short-term wind power prediction for large-scale wind power clusters. The analysis is based on measured data from a large-scale wind power cluster of 159 wind farms in a certain area, with a sampling interval of 15 minutes. The installed capacity of this power station is 17286.6 MW, and the cluster division results are shown in Table 1.

[0163] Table 1 Cluster Partition Results

[0164]

[0165] To ensure consistent measurement and improve computational efficiency, the max-min normalization method is used to normalize the input and output data. Taking a specific feature x as an example, the principle of normalization is as follows:

[0166]

[0167] Where x′ represents the normalized feature vector, x min and x max These represent the maximum and minimum values ​​of x, respectively. During the prediction phase, the prediction results are restored to the original power range based on the inverse normalization formula. The inverse normalization principle is shown in formula (23):

[0168] x=x'(x max -x min )+x min (twenty three)

[0169] Taking cluster 1 as an example, principal component analysis was used to reduce the dimensionality of the three features (wind speed, temperature, and humidity) of the 32 wind farms in the space, with 10 principal components selected. The principal component contribution rates of each feature are as follows: Figure 4 As shown, the first and second principal components of wind speed have relatively high contribution rates, exceeding 70% combined; therefore, they are used as modeling inputs. The first principal component of temperature exceeds 90%, so it is also used as a modeling input. Similarly, the first principal component of humidity exceeds 70%, making it the first principal component of temperature the chosen modeling input. The principal component analysis algorithm reduces the 96 spatial features composed of wind speed, temperature, and humidity to just 4, significantly reducing data dimensionality, preserving spatial feature consistency, and promoting algorithm convergence.

[0170] The training parameters for the data-driven model are shown in Table 2. To reduce the risk of overfitting, dropout is set to 0.2.

[0171] Table 2 Training Parameters

[0172]

[0173] For the 16-step power extrapolation modeling, the spatial attention weights do not vary significantly across different prediction step lengths. However, for a given prediction step length, historical power has the highest attention weight, indicating that the autocorrelation of the power series plays a dominant role in ultra-short-term power prediction modeling. Wind speed is the second most important component, followed by the spatially averaged wind speed, suggesting that wind speed data from numerical weather prediction also contributes to accurate modeling. Strictly speaking, there is a causal relationship between wind speed and wind power, and wind speed should have a high attention weight. However, due to errors in numerical weather prediction, this causal relationship is weakened in the externalization of attention weights. Compared to wind speed and historical power, temperature, humidity, and pressure contribute less.

[0174] The 16th step prediction curve for regional wind power clusters and wind power groups is as follows: Figure 5 As shown, the predicted curve can accurately track the actual power curve with a small error. However, it is worth noting that, like other multi-step prediction models, the STAN model also exhibits a time delay in its prediction results, and this delay becomes increasingly pronounced as the prediction step size increases. Tables 3-7 show the prediction performance of clusters 1-5, respectively.

[0175] Table 3 Performance of Cluster 1

[0176]

[0177] Table 4 Performance of Cluster 2

[0178]

[0179] Table 5 Performance of Cluster 3

[0180]

[0181] Table 6 Performance of Cluster 4

[0182]

[0183]

[0184] Table 7 Performance of Cluster 5

[0185]

[0186] The STAN model's pass rate remained almost 100%, with the absolute error of predictions for most samples below 25% of the installed capacity. The correlation coefficient and R² coefficient decreased with increasing prediction timescale, but remained high even after extrapolation to step 16, indicating a high degree of similarity between the predicted and actual power curves, with effective fitting of peaks and troughs. The MAE and RMSE error assessment metrics showed a slight upward trend, but the magnitude was small, indicating low absolute error on the test set. For the prediction results at hour 4, the RMSE remained stable within 10% of the installed capacity, demonstrating the STAN model's stable power prediction performance for large-scale wind power clusters.

[0187] Extreme error events have the greatest impact on the power system. Extreme positive errors indicate that the predicted value is significantly lower than the actual value, resulting in higher actual wind power output and a higher likelihood of wind curtailment. Extreme negative errors indicate that the predicted value is significantly higher than the actual value, resulting in lower actual wind power output and a higher likelihood of load shedding events. The extreme error statistics for the 16-step prediction of the STAN model are as follows: Figure 6 As shown, according to the principle that an absolute error exceeding 40% of the installed capacity is considered an extreme error event, extreme error events occur very rarely when predicting large-scale wind power clusters, compared to predictions based on a single wind farm. As the prediction step size increases, the absolute value of the extreme error generally increases. For the prediction in step 16, the extreme positive error is 25.278% of the installed capacity, and the extreme positive error is 25.719% of the installed capacity, both significantly lower than 40% of the installed capacity, thus demonstrating the stability of the STAN model.

[0188] The performance of the STAN algorithm proposed in this invention was compared with that of Temporal Convolutional Network (TCN), Long Short-Term Memory Network (LSTM), Bidirectional Long Short-Term Memory Network (BiLSTM), and Rec-Encoder composed of LSTM. The comparison metric was the average RMSE of the prediction results over the next 4 hours. The results show that the STAN algorithm achieved the lowest error; the TCN algorithm had the lowest performance, with its highest error differing from its lowest RMSE by only 0.00734. However, the total installed capacity of the wind power cluster is 16887MW, which is equivalent to a prediction error difference of 124MW between the two algorithms, equivalent to the installed capacity of a small to medium-sized wind farm.

[0189] The performance of the prediction mode based on the cluster division and sub-cluster prediction results of this invention, the prediction mode based on single-field prediction, and the prediction mode based on the total cluster are compared. The performance comparison is based on the average RMSE of 4-hour predictions. The average RMSE of the prediction mode of this invention is 0.0522, the average RMSE of single-field prediction overlay is 0.0618, and the average RMSE of the prediction mode based on the total cluster is 0.0639. The prediction mode proposed in this invention has the lowest error and makes a greater contribution to ensuring the power supply capacity of the guaranteed area.

[0190] The specific embodiments of this invention are not exhaustive and do not constitute a limitation on the scope of protection of the claims. Those skilled in the art, upon learning from the embodiments of this invention, can conceive of other substantially equivalent alternatives without inventive effort, all of which are within the scope of protection of this invention.

Claims

1. A method for ultra-short-term wind power prediction in large-scale wind power clusters, characterized in that: It includes the following steps: 1) Wind power cluster partitioning method based on fuzzy clustering By introducing the concept of membership degree into fuzzy clustering, each sample belongs to different clusters with a certain membership degree. Based on this, the samples are divided into several disjoint subsets, and the observed data matrix is ​​denoted as Equation (1): （1） In the formula, , For the first The first sample The results of this observation, M is the number of samples; Define the objective function as equation (2): （2） In the formula, Represents the cluster center matrix, where , Represents the number of cluster centers. The membership matrix, For the first Sample Belongs to the Membership degree of a class Representative sample With cluster center Euclidean distance, The number of samples; Fuzzy clustering is based on initializing cluster centers. By continuously updating the membership matrix and cluster centers, and with the objective function minimization as the optimization goal, the final classification is obtained. The cluster center and membership update methods are Equations (3)-(4): （3） （4） In the formula This represents the number of iterations. Representative input features are the fundamental condition for ensuring algorithm convergence. By constructing statistical features, the fluctuation characteristics of wind farms are characterized. Using the fluctuation characteristics of each wind farm as input and fuzzy clustering algorithm as classifier, the division of wind power clusters is realized. The wind farm fluctuation feature construction scheme is as follows: Equation (5)-Equation (10): The first statistical feature is the per-unit value of the maximum output of the wind farm, which represents the maximum output level of the wind farm; (5) In the formula, n represents the number of fans, f represents the per-unit value, and P represents the power of the fans. This represents the per-unit value indicating the maximum output of a wind farm; The second statistical feature is the maximum wind speed of the wind farm, which reflects the maximum wind speed of the wind farm. (6) In the formula, V max V is the maximum wind speed of the wind farm. i Let n represent the wind speed of the i-th fan, and n represent the number of fans. The third statistical feature is the power variance of the wind farm, which reflects the power fluctuation of the wind farm. (7) In the formula, Let P be the power variance of the wind farm. i Let represent the power of the i-th fan. This represents the average power of the wind farm, and n represents the number of wind turbines. The fourth statistical characteristic is the wind speed variance of the wind farm, which reflects the fluctuation of wind speed in the wind farm; (8) In the formula, V is the variance of wind speed in a wind farm. i Let i represent the wind speed of the i-th fan. The average wind speed of the wind farm is represented by n, and the number of wind turbines is represented by n. The fifth statistical characteristic is the average wind speed, which reflects the distribution of average wind speed. (9) In the formula, V represents the average wind speed. i Let n represent the wind speed of the i-th fan, and n represent the number of fans. The sixth statistical characteristic is the per-unit value of average power, which reflects the distribution of average power. (10) In the formula, P represents the average power per unit, f represents the calculated per unit value. i Let n represent the power of the i-th fan and n represent the number of fans. 2) Wind power cluster power prediction method based on spatiotemporal attention neural network Give with Multivariate time series with 1 feature , Represents the length of the input sequence. , indicating the first Input time series of 1 input variable; For each feature A feedforward neural network is used to calculate the spatial embedding; for the input data The embedding calculation result of all input variables is For time step input The time-step embedding is computed by the coding layer consisting of two LSTM layers. ; A feedforward neural network is used as the alignment model to calculate spatial attention weights. For output time steps , No. Spatial attention weights Equations (11)-(12): (11) (12) in , This represents the previous hidden layer state of the LSTM decoder. Representing the Spatial embedding of features, For spatial attention weights, b e W is the bias value. e The spatial attention weights are used as weight values, and the spatial context vector is further calculated based on the spatial attention weights. ; For output time step Using the alignment model to calculate and Corresponding time attention weights , It is about and the previous hidden layer state of the decoder The function is given by equation (13): (13) Among them, b a W is the bias value. a For weight values, Further based on the time context vector Calculate the attention weights over time Spatial and temporal attention depend on the hidden layer states of the LSTM decoder, i.e., LSTM G and LSTM S Among them, LSTM G Spatial attention depends on the hidden layer states of the LSTM decoder. S For time attention to depend on the hidden layer states of the LSTM decoder, the computation space context vector is... At that time, the hidden layer states and neuron states of LSTM are defined as follows: and According to the decoder hidden layer state Calculate spatial attention weights This leads to the spatial context vector. ,Will and the previous time step By splicing , Input to LSTM G To update the hidden layer state b G W is the bias value. G The weight values ​​are given by equation (14): (14) For the time context vector of LSTM The hidden layers and neuron states of LSTM are respectively and In time step , Represents the temporal attention used to compute the temporal context vector, concatenated. and And input into LSTM S Update hidden layer state b S W is the bias value. S The weight values ​​are given by equation (15): (15) Before prediction, the final step is to connect the hidden state updates of the two LSTMs together. ; Using the STAN model as the predictor, let the current time be... The number of features is Each input feature is taken 16 steps forward, that is... Using time-specific features as input, and based on spatiotemporal correlation feature mining, the ultra-short-term wind power of large-scale wind power clusters is predicted. - : Based on meteorological and historical power data, statistical characteristics are constructed to characterize the fluctuation characteristics of wind farms. The K-Means algorithm is used to cluster the input features to achieve the division of wind power clusters. Principal component analysis algorithm is used to reduce the dimensionality of wind speed, temperature and humidity of each wind power cluster, and a spatial feature set is constructed by combining the average values ​​of wind speed, temperature, humidity and pressure. • Construct a spatiotemporal attention network, divide it into training and test sets, train the spatiotemporal attention network on the training set, and obtain an ultra-short-term power prediction model; • By overlaying the forecast results of each wind power cluster, the power forecast results of the wind power base are reconstructed; • By combining a comprehensive index evaluation system, the accuracy and stability of the ultra-short-term power prediction results of large-scale wind power clusters are comprehensively evaluated. 3) Simulation calculation Simulation input: Analyze the measured data of the electric field to determine the total installed capacity of the wind farm; Input data: historical power of the wind farm, wind speed, temperature and humidity NWP data of the wind farm; Data sampling interval is 15 min; According to steps 1) to 2), the ultra-short-term wind power prediction results for the prediction period are obtained; 4) Error Analysis The root mean square error is calculated using equation (16): (16) In the formula, represent Actual power at time, represent Predicted power at time, Represents the length of the test set. This represents the actual operating capacity of the wind farm. Since the actual operating capacity of the wind farm is difficult to obtain, the installed capacity is set as the actual operating capacity of the wind farm. The mean absolute error is calculated using equation (17): (17) In the formula, R mae Indicates the mean absolute error. This represents the actual power at time i. Represents the predicted power at time i; The pass rate is calculated using formula (12): (18) In the formula, The pass rate is determined by the standard. Representative sample The prediction was unsatisfactory. Representative sample The prediction is satisfactory. In addition, the maximum positive error and the minimum negative error are used to evaluate extreme error events, and the error is calculated by equation (19): (19) In the formula, err(t) represents the error value, and y(t) represents the actual value. Indicates the predicted value; The maximum positive error is calculated using equation (20): (20) In the formula, err(t) represents the error value, and MPE represents the maximum positive error; The minimum negative error is calculated using equation (21): (21) In the formula, MNE represents the minimum negative error; Input the simulation input in step 3), and calculate the error between the predicted power calculated by the model and the measured power using the error evaluation standard formulas (16) and (21) in step 4) to obtain the root mean square error, mean absolute error, pass rate, maximum positive error and minimum negative error of the prediction.

Images