A short-term power prediction method for a regional wind power cluster

Abstract

The present invention relates to a short-term power prediction method for a regional wind power cluster, which includes two aspects: spatio-temporal multiple similarity calculation and improved neural network construction. Considering the dynamic characteristics of spatio-temporal correlation, the spatial correlation degree of wind farms is calculated from long and short time scales. For the sub-clusters with strong spatio-temporal correlation classified, an improved hybrid neural network short-term prediction model based on point cloud input is established to realize the short-term power prediction of the sub-clusters. Among them, a spatio-temporal correlation cluster short-term power prediction model based on an improved I-CNN-BILSTM hybrid neural network is established, and the short-term power prediction results of all spatio-temporal related sub-clusters are accumulated to obtain the short-term power prediction result of the regional wind power for the period to be predicted. Compared with the prior art, the present invention has the advantages of improving the prediction accuracy, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a short-term power prediction method for a regional wind power cluster. Background Art

[0002] With the large-scale development of wind power forming a regional wind power cluster, the grid connection of the wind power cluster has caused more impacts on the safe and stable operation and dispatching control of the power system. On the one hand, due to the large-scale and uncertain access of wind power, the system needs to reserve sufficient spare capacity in advance to cope with the volatility of wind power and the system peak shaving problem; on the other hand, the output volatility of large-scale wind power makes its grid connection likely to cause fluctuations in the grid voltage and frequency, affecting the power quality of the grid. Improving the power prediction accuracy of the regional wind power cluster is an effective means to address the above problems. Accurate prediction of the large-scale wind power in the region is an important method to improve the operation stability of the power system and the wind power accommodation capacity. The spatial distribution range of the large-scale wind power in the region is wide and the number of wind farms is large. Due to the persistence of wind resources and the similarity of the terrain where the wind farms are located, the output of multiple wind farms in the region has spatio-temporal correlation characteristics. The complex spatio-temporal correlation between the wind farms increases the difficulty of regional wind power prediction.

[0003] In the area where wind farms are relatively concentrated, the wind speed values and power values of a specific wind farm not only have a certain autocorrelation in time, but also are affected by other wind farms at different positions, the topological structure of the space where they are located, and the environmental conditions. The traditional prediction model based on the historical power and meteorological data of the wind farm itself ignores the spatio-temporal correlation between multiple regional wind farms and it is difficult to accurately predict the aggregated power of multiple regional wind farms. Therefore, for traditional cluster prediction, it is necessary to further combine spatio-temporal correlation information to enhance the short-term power prediction accuracy.

[0004] The existing methods for wind power prediction considering spatio-temporal correlation mainly include: integrating the measurement information of neighboring locations into the power prediction model of the target site through different machine learning methods, establishing prediction models for all power stations in the region separately and superimposing the results, but this method is not applicable to large-scale wind power regions; establishing an input spatio-temporal feature map of multiple wind farms in the region and mining spatio-temporal relationships through a deep neural network, but the processing ability of the model for large-scale wind farms with complex correlation degrees needs to be improved; adding spatial correlation constraints to the sparse control vector autoregressive model to predict large-scale wind power clusters, improving the overall prediction accuracy, but only considering static correlation and ignoring the dynamic correlation caused by short-term changes in wind speed and direction. To sum up, regional wind power prediction based on spatio-temporal correlation has become a research hotspot for domestic and foreign scholars in recent years, but there is still room for improvement. In particular, extracting the dynamic spatio-temporal correlation between multiple wind farms in the region at different time scales and then establishing a spatio-temporal integrated prediction model for strongly correlated clusters can improve the short-term power prediction accuracy of the region. In terms of prediction methods, the existing technologies mainly adopt the following methods: proposing a hybrid network model based on convolutional neural network (CNN) and gated recurrent unit (GRU) to improve the prediction accuracy, but this model simply concatenates CNN and GRU and fails to effectively retain the time structure of the input data, resulting in poor overall prediction effect; at the same time, convolutional neural network is suitable for regular grid data and has insufficient processing ability for irregularly distributed point cloud data, and the spatio-temporal features of the extracted clusters are less, resulting in low prediction accuracy of the clusters. Summary of the Invention

[0005] The purpose of the present invention is to overcome the above-mentioned defects existing in the prior art and provide a short-term power prediction method for regional wind power clusters.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] A short-term power prediction method for regional wind power clusters, the method comprising the following steps:

[0008] S1: Obtain the historical annual power data of regional wind farms, form a daily power sample set of wind farms in units of days, perform similarity clustering on the daily power sample set, extract the typical forms of daily power curves of wind farms at a long-term scale, and define different power curve morphological features as different daily power patterns.

[0009] S2: Form a daily power pattern sequence according to the annual power data, establish a classification matrix Q of daily power pattern sequences of different wind farms, statistically analyze the probability distribution of joint patterns of different wind farms in the classification matrix to form an estimated probability matrix P, and calculate the entropy value of the estimated probability matrix. Classify multiple wind farms in the region into main clusters with similar long-term output patterns according to the entropy value.

[0010] S3: Extract the historical short-term power data of each wind farm in the main cluster before the prediction moment, divide and identify the types of power fluctuation segments, solve the similarity of fluctuation trends using the similar distance of fluctuation trends, and then cluster the sub-clusters that are strongly correlated in fluctuation trends in the short term.

[0011] S4: Extract the historical meteorological data and spatial coordinate data of multiple wind farms in the sub-cluster to form an input feature matrix, and establish a short-term power prediction model based on an improved I-CNN-BILSTM hybrid neural network. The historical meteorological data includes wind speed, wind direction, temperature, and historical power sequences, and the spatial coordinate data includes longitude, latitude, and altitude.

[0012] S5: Initialize the weights of the neural network and set the maximum number of iterations.

[0013] S6: Construct a point convolutional neural network based on point cloud input, and input the feature matrix into the point convolutional neural network to extract global features and local features.

[0014] S7: Input the features extracted by the point CNN network into an improved Attention mechanism combined with the SSA algorithm, assign different weights to the transition feature vectors of the neural network model, input the weighted transition feature vectors into the BILSTM layer by time step, output the training results of the improved I-CNN-BILSTM hybrid neural network, read the training loss curve and error curve, and evaluate the convergence performance of the network prediction results according to the vertical distance between the loss curves of the training set and the validation set during the convergence process, as well as the magnitude of the absolute error.

[0015] S8: Determine whether the current number of iterations has reached the set number of iterations. If so, terminate the iteration and output the parameters of the hybrid neural network. Otherwise, increment the number of iterations by one and go to step S4.

[0016] S9: Use the improved I-CNN-BILSTM network trained in the above steps to perform short-term wind power prediction for the corresponding sub-cluster and obtain the predicted power of the sub-cluster.

[0017] S10: Determine whether the power prediction of all sub-clusters has been completed. If so, proceed to S11. Otherwise, go to S4 to predict the remaining sub-clusters.

[0018] S11: Perform superposition calculation on the power prediction values of all sub-clusters to obtain the overall wind power prediction result of the region.

[0019] Furthermore, in S1, the ACFSFDP algorithm is used to perform similarity clustering on the daily power sample set. The specific steps include:

[0020] S101: Form a sample set in days from the normalized historical wind power data, and calculate the local density ρ of the samples i , topological relationship, and distance δ i ;

[0021] S102: Combine the product of the sample local density ρ i and the distance δ i to form a comprehensive evaluation index γ, and select the clustering center according to this comprehensive evaluation index;

[0022] S103: Establish an adaptive index F K for determining the optimal number of clustering clusters, gradually increase the number of clustering centers, and subtract one from the number of clustering centers corresponding to the only maximum peak of the adaptive index as the optimal number of clustering clusters.

[0023] Further, the specific steps of S2 include:

[0024] S201: Form a daily power pattern sequence from the daily power data of a wind farm in the past year according to the clustering result of S1, and use the multiple groups of daily power pattern sequences of all wind farms in the region to form a classification matrix Q. The row vector of the matrix is the daily power pattern sequence of the corresponding wind farm in N days, and the k-th column vector of the matrix represents the joint pattern of the daily power curves of different wind farms;

[0025] S202: Count the probabilities of the occurrence of each joint pattern in the classification matrix Q to form an estimated probability matrix P. The element p ij in the estimated probability matrix P represents the probability of the occurrence of (i, j) T in each column of Q. Calculate the entropy value H of the estimated probability matrix: H = -∑p ij ln(p ij ). Classify the wind farms with the entropy value H less than the set threshold H min into the same cluster; and classify the multiple wind farms in the region into main clusters with similar long-term output patterns according to the magnitude of the entropy value.

[0026] Further, in S3, the specific steps of solving the similarity of the fluctuation trend by using the similarity distance of the fluctuation trend include:

[0027] S301: Define the characteristic parameters describing the power fluctuation, including the volatility S t reflecting the power change trend, the fluctuation standard deviation B reflecting the severity of the power fluctuation, the fluctuation degree W reflecting the frequency of the power fluctuation, and the high power ratio Q h and the low power ratio Q l reflecting the magnitude of the actual output value;

[0028] S302: Extract the historical short-term power series of the wind farm in the three days before the prediction moment. According to the fluctuation parameters, divide the power time series into multiple different types of fluctuation segments, including stable power fluctuation, ramp power fluctuation, oscillating power fluctuation, and peak-valley power fluctuation, which are defined as type-a, type-b, type-c, and type-d fluctuations respectively, and convert the power time series into a fluctuation type sequence Z P ={z1,…,z i ,…}, where z i ∈{a,b,c,d};

[0029] S303: The fluctuation type sequences of all wind farms in the main cluster form a fluctuation type matrix Z. For the power time series P1 and P2, their fluctuation type sequences are and where z i ,z j ∈{a,b,c,d}, calculate the fluctuation type matching distance d(z i ,z j ):

[0030]

[0031] The fluctuation type matching distance forms a distance matrix D k×l and a cumulative matrix R = {r(i,j)} k×l

[0032] D k×l ={d(z i ,z j )} k×l , i = 1,2,…k.j = 1,2,…l

[0033]

[0034] In the formula, r(0,0) = 0, r(i,0) = r(0,j) = ∞;

[0035] The similarity distance of the fluctuation trends of different fluctuation type sequences is Set the minimum distance threshold L min . Sequences with a distance L less than the threshold are clustered into clusters of sequences with similar fluctuation trends, and the corresponding wind farm groups are spatio-temporal correlation sub-clusters that take into account the similarity of long-term output patterns and short-term output fluctuations.

[0036] Furthermore, the specific content of the short-term power prediction model for the wind power cluster based on the improved I-CNN-BILSTM hybrid neural network is as follows:

[0037] a) Use the point CNN neural network to extract the local and global spatial features of the cluster; specifically:

[0038] a1) Establish the input feature matrix of the wind power cluster, including the historical meteorological data and spatial coordinate data of each wind farm in the cluster;

[0039] a2) Encode the irregular wind farm location information into the weights of the X transformation matrix, and arrange the features of the associated points in the point cloud;

[0040] a3) Construct the point cloud data feature graph G = (α, β), where the vertices α = {1, 2, …, n} represent the information of each data point in the point cloud, and the edges β represent the association information between every two data points in the graph;

[0041] a4) Set the edge function as e ij = h(x i , x j - x i , θ), whose global features are provided by the numerous data points constituting the point cloud, and the local features are provided by the association information between every two points in the point cloud, where hx i , x j - x i , θ) is the edge function for edge feature extraction, θ is the parameter to be learned and optimized in the edge function, and x i and x j are sample data points.

[0042] b) Use the improved Attention mechanism combined with the SSA algorithm to achieve weight allocation between time series; specifically:

[0043] b1) Initialize the weight W of the attention layer;

[0044] b2) Input the attention layer weight W into the Salp Swarm Algorithm for optimization, and transfer the optimized weight set to the BILSTM neural network, where the BILSTM neural network generates corresponding loss values according to the prediction error in the network;

[0045] b3) Select the optimal attention weight set according to the loss situation.

[0046] c) Introduce the Targeted dropout algorithm in the BILSTM network to selectively eliminate neurons; specifically:

[0047] c1) Prune the neurons of the BILSTM network according to the methods of weight pruning and unit pruning, and the calculation formula is:

[0048]

[0049]

[0050] where: ε(W c (θ)) is the network loss function, Wc is the parameter matrix of the neural network model, argmax-k is a function that returns the k largest elements among all elements, and w o is the column vector of the o-th column of the weight matrix W, and W io is the element in the i-th row and o-th column of the weight matrix, and N col and N row respectively represent the number of columns and rows of the parameter matrix;

[0051] c2) Introduce the targeting ratio γ and the deletion probability α. Select the smallest γ|θ| weights as the candidate weights for dropout, and then independently remove the weights in the candidate set with the deletion probability α.

[0052] The short-term power prediction method for regional wind power clusters provided by the present invention has at least the following beneficial effects compared with the prior art:

[0053] First, the present invention considers the spatio-temporal correlation of regional wind farms, divides the spatio-temporal correlation clusters through the spatio-temporal multiple similarity method, clusters the main clusters that are spatially correlated on a long time scale through the joint distribution of daily power patterns, and then identifies the spatio-temporal correlation sub-clusters that take into account both long and short time scales by matching the short-term power fluctuation similarity in the main clusters, having a stronger recognition ability for the dynamic spatio-temporal correlation between wind farms;

[0054] Second, the present invention proposes an improved I-CNN-BILSTM hybrid neural network cluster prediction model. This model establishes a point convolutional neural network based on the point cloud input of the wind power cluster, effectively utilizes the spatial position information of the wind farms, and improves the integrity of the network in extracting spatial features; at the same time, this model proposes an improved Attention mechanism combined with the SSA algorithm, fully extracts the correlation information between feature sequences, and combines the improved Targeted dropout algorithm, which is beneficial to the fitting and optimization of the hybrid model and improves the prediction accuracy of the wind power cluster. Description of the Drawings

[0055] Figure 1 is a schematic diagram of the short-term power prediction process for the regional wind power cluster in the embodiment;

[0056] Figure 2 is the historical long-term output curve of three adjacent wind farms in the region in the embodiment;

[0057] Figure 3 is the short-term output fluctuation curve of three adjacent wind farms in the region in the embodiment;

[0058] Figure 4 is a schematic diagram of the change of the adaptive index and the number of cluster centers proposed by the ACFSFDP algorithm in the embodiment;

[0059] Figure 5Schematic diagram of the Attention mechanism process improved by combining SSA in the embodiment. Detailed implementation manners

[0060] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0061] Embodiment

[0062] The present invention relates to a short-term power prediction method for a regional wind power cluster. The method includes two aspects: spatio-temporal multiple similarity calculation and improved neural network construction. Considering the dynamic characteristics of spatio-temporal correlation, the spatial correlation degree of wind farms is calculated from long and short time scales. For the sub-clusters with strong spatio-temporal correlation classified, an improved hybrid neural network short-term prediction model based on point cloud input is established to realize the short-term power prediction of the sub-clusters, which can further improve the cluster prediction accuracy.

[0063] The main principle of the short-term power prediction of the sub-clusters by the improved hybrid neural network short-term prediction model based on point cloud input established by the present invention is as follows:

[0064] Spatio-temporal multiple similarity clustering aspect: Since the long-term persistence and short-term variability of meteorological environment factors such as wind speed will lead to dynamic spatio-temporal correlation in the output laws of regional wind farms. To improve the power prediction accuracy of regional wind power clusters under large-scale wide-area distribution and spatio-temporal correlation of output, the overall region needs to be decomposed into multiple wind power sub-clusters with strong spatio-temporal coupling correlation, and an effective cluster power prediction model for extracting spatio-temporal features needs to be constructed. Figure 2 The historical long-term output curves of three wind farms are shown. From Figure 2 Analyzing the historical long-term output curves of multiple wind farms, it is not difficult to see that multiple wind farms with adjacent locations and the same terrain have strong similarity in long-term output levels and output patterns. On a daily time scale, the daily power curves of strongly correlated wind farms have high similarity in morphological characteristics. Therefore, a daily power curve similarity matching method can be used to classify wind power clusters that are spatially correlated on a long time scale. From Figure 3 The short-term output sequences with minute-level time resolution shown, it can be seen that during the wind speed mutation stage, the output correlation of the wind power clusters related to the long-term mode changes greatly. To effectively extract the strongly correlated sub-clusters at the prediction moment, the present invention proposes a correlation analysis method based on similarity measurement of fluctuation trends, which identifies and divides the fluctuation segments of the power sequences of different stations, and calculates the matching distance of fluctuation types, classifies the spatio-temporal correlation wind power sub-clusters that take into account both long-term mode and short-term fluctuation similarity, and uses them as the prediction objects of the spatio-temporal prediction model.

[0065] In terms of prediction algorithms: For the feature matrix composed of cluster feature data by the Point CNN-BILSTM hybrid model, the Point CNN is fully utilized to extract the spatial local features and global features of the data, and the BILSTM can make up for the defect that the Point CNN is difficult to capture the long-term dependence relationship in the sequence data. Since the features for cluster wind power, such as wind speed, wind direction, temperature, spatial position, etc., are relatively independent feature time series, it is difficult to describe the internal relationship between the time series of each feature. Using only the Point CNN or BILSTM alone cannot extract the relevant characteristics between sequences and the long-term law of feature time series at the same time. The traditional CNN-BILSTM simply concatenates the CNN and the BILSTM, which is likely to destroy the correlation between time series. Therefore, it is necessary to make improvements on the basis of the traditional CNN-BILSTM to eliminate the above disadvantages.

[0066] Wind power clusters often have an irregular point cloud distribution. Directly applying the traditional CNN to extract cluster data will lose the inherent spatial position information of the wind farm and cannot extract complete spatio-temporal feature information. To address the above problems, the present invention proposes an improved I-CNN-BILSTM hybrid neural network algorithm. The Point CNN learns the spatial coordinate information of the cluster wind farm to obtain the X transformation matrix, which is encoded into the data feature matrices such as wind speed, wind direction, temperature, and power, so as to perform spatial feature extraction. The SSA algorithm is introduced to improve the Attention mechanism in the convolutional layer and the BILSTM layer, and a dropout algorithm based on a pruning strategy is used to suppress overfitting in model training.

[0067] Based on the above principles and design ideas, the short-term power prediction method for regional wind power clusters of the present invention specifically includes the following steps:

[0068] Step 1: Statistically analyze the power data of all wind farms in the region for the past year, calculate the correlation of the long-term output sequence, and realize the clustering of the daily power curve patterns of the wind farms. The specific operations are as follows:

[0069] Due to the high randomness and volatility of wind power output, the power data sample set has an irregular shape. The Clustering by Fast Search and Find of Density Peaks (CFSFDP) algorithm has a good clustering effect on samples of any shape, but it requires the number of cluster centers to be determined artificially, and subjective factors will bring large errors to the clustering effect. Therefore, this patent improves the traditional CFSFDP algorithm by constructing an adaptive index to automatically determine the optimal number of cluster centers. The proposed Adaptive Clustering by Fast Search and Find of Density Peaks (ACFSFDP) extracts typical patterns from daily power curves. If the data set C to be clustered is n samples to be classified, that is:

[0070] C = {p1, p2, …, p n}

[0071] where p i is the sample point to be classified, and i = 1, …, n.

[0072] Then the basic process of ACFSFDP clustering is as follows:

[0073] A) Calculate the improved local density ρ i of each sample point p i .

[0074]

[0075] where d ij is the Euclidean distance between data points p i and p j ; d c is the cut-off distance.

[0076] B) Calculate the topological relationship and distance δ i between samples. The calculation formula is as follows:

[0077]

[0078] where p j is the sample with a density greater than p i and the closest distance. If p i is the sample with the largest density, then δ i = m j in(d ij ).

[0079] C) Select samples with a relatively large density and a relatively large distance between samples as cluster centers, using the comprehensive evaluation index γi Determine the clustering centers based on the values.

[0080] γ i = ρ i δ i

[0081] D) According to the arrangement order of the γ i values, assign the remaining samples to the cluster where the sample with the nearest neighbor and a density greater than it is located.

[0082] Since the traditional fast search density peak clustering algorithm requires the number of clustering centers to be determined manually, subjective factors will bring relatively large errors. Therefore, the present invention proposes an adaptive index F K to determine the optimal number of clustering clusters.

[0083]

[0084] In the formula, K is the number of clustering clusters, and C K is the total number of samples that have been assigned to the K clustering centers after clustering is completed.

[0085] Figure 4 is the variation of the adaptive index with the number of clustering centers during daily curve pattern clustering. As Figure 4 can be seen, when the number of clustering centers increases, the adaptive index will show a maximum peak. At this time, the corresponding K value K peak just crosses the optimal number of clusters, resulting in the splitting of the optimal clustering cluster and increasing sample noise. Therefore, the optimal number of clustering clusters is determined as K peak - 1.

[0086] Furthermore, based on the above ACFSFDP algorithm process, the specific steps for the present invention to quickly cluster the daily power curve of a wind farm using the ACFSFDP algorithm include:

[0087] 11) Form a sample set with days as the unit from the normalized historical wind power data, and calculate the local density ρ i of the samples, the topological relationship, and the distance δ i . The normalization calculation formula is as follows:

[0088]

[0089] In the formula, X p.u. is the normalized power, X is the original power sequence, and X min and X max are the minimum and maximum values of the power sequence respectively.

[0090] 12) Combine the product of the sample local density and the distance to form a comprehensive evaluation index γ, and select the clustering centers based on this index.

[0091] 13) Establish an adaptive index F for determining the optimal number of clustering clusters. K Gradually increase the number of cluster centers. The adaptive index will have a unique maximum peak. Subtract 1 from the number of cluster centers corresponding to the peak, and the result is the optimal number of clustering clusters.

[0092] By extracting the typical forms of the daily power curves of the wind farm on a long-term scale and defining different power curve morphological features as daily power patterns E1, E2, E3, E4, and E5, these patterns reflect the numerical values and change trends of the daily power curves of the wind farm. Then, form a daily power pattern sequence from the daily power data of the wind farm over the past year according to the above clustering results.

[0093] Step 2: Based on the daily power pattern sequence of the wind farm obtained in Step 1, construct a pattern classification matrix Q. The row vectors of the matrix are the daily power pattern sequences of the corresponding wind farm over N days. The k-th column vector of the matrix represents the joint pattern of the daily power curves of different wind farms. Statistically calculate the probabilities of the occurrence of each joint pattern in the pattern classification matrix to form an estimated probability matrix P. Its element p ij represents the probability of the occurrence of (i,j) in each column of Q T Calculate the entropy value H of the estimated probability matrix: H = -∑p ij ln(p ij ). Wind farms with entropy value H less than the set threshold H min belong to the same cluster. Classify the multiple wind farms in the region into main clusters with similar long-term output patterns according to the entropy values.

[0094] Step 3: Use the fluctuation segment analysis method to refine the extraction of the fluctuation trend of the power sequence. Determine the fluctuation trend of the short-term power sequence of the wind farms in the main cluster three days before the prediction moment, and propose characteristic parameters describing the degree of power fluctuation. The specific parameter calculation methods are as follows:

[0095] A) Volatility S t : The difference between the endpoint powers, reflecting the power change trend;

[0096] S t = S0 - S1

[0097] B) Fluctuation standard deviation B: The standard deviation of the sequence power values, reflecting the severity of power fluctuations;

[0098]

[0099] C) Fluctuation degree W: The ratio of the number of extreme points in the fluctuation segment to the time period, reflecting the frequency of power fluctuations;

[0100] W = N / T

[0101] D) High power ratio Q h: The ratio of the time when the output level is above 75% of the rated capacity to T;

[0102] Q h = T H / T

[0103] E) Low power ratio Q l : The ratio of the time when the output level is below 20% of the rated capacity to T;

[0104] Q l = T L / T

[0105] Wherein, T is the duration of the fluctuation section; S0 and S1 are the powers at the two end points of the fluctuation section respectively; N is the number of extreme points within the fluctuation section; is the mean value within the fluctuation section; T H and T L are the times when the output power is above 75% and below 20% of the installed capacity of the wind farm respectively.

[0106] By inputting the fluctuation indexes of the segmented power sequence extracted into the k-means algorithm for clustering calculation, the overall sequence can be divided into four different fluctuation types, including stable power fluctuation, ramp power fluctuation, oscillating power fluctuation and peak-valley power fluctuation. The specific corresponding actual power change situations are as follows:

[0107] 1) Stable power fluctuation: Corresponding to the power fluctuation of the wind farm under relatively stable high wind speeds and relatively stable low wind speeds; denoted as type a fluctuation.

[0108] 2) Ramp power fluctuation: The wind speed of the wind farm is in an obvious rising or falling trend and there is no accompanying violent fluctuation; denoted as type b fluctuation.

[0109] 3) Oscillating power fluctuation: Corresponding to the random fluctuation of the wind farm output power for a period of time under the influence of a relatively complex meteorological environment, and containing a large number of high-frequency oscillations; denoted as type c fluctuation.

[0110] 4) Peak-valley power fluctuation: Corresponding to the drastic change of meteorological factors such as environmental wind speed and wind direction of the wind farm, the output curve of the fan generates a large amplitude oscillation, and there is an obvious peak-valley difference. Denoted as type d fluctuation.

[0111] Step 4. Extract the fluctuation type sequence of the short-term power data of the main cluster wind farm to form a fluctuation type matrix Z. For the power time series P1, P2, their fluctuation type sequences are and where z i , z j ∈ {a, b, c, d}, calculate the pattern matching distance d(z i , z j ):

[0112]

[0113] The pattern matching distance forms the distance matrix D of two fluctuation sequences k×l and the cumulative matrix R = {r(i, j)} k×l :

[0114] D k×l = {d(z i , z j )} k×l , i = 1, 2, … k. j = 1, 2, … l

[0115]

[0116] where r(0, 0) = 0, r(i, 0) = r(0, j) = ∞.

[0117] The fluctuation pattern matching distance of the fluctuation type sequence Cluster the sequences with similar fluctuation trends by the matching distance L, and the corresponding wind farm group is a spatio-temporal correlation sub-cluster that takes into account the similarities of long-term patterns and short-term fluctuations.

[0118] Step 5: Based on the historical wind speed, wind direction, temperature and other meteorological data, historical output time series, and spatial coordinate data of the station longitude, latitude and altitude of the cluster wind farms, construct a cluster input matrix and establish a cluster power prediction model based on an improved I-CNN-BILSTM hybrid neural network with point cloud input. The cluster power has dual characteristics of time and space and can be expressed by the following formula:

[0119]

[0120] where: P clu (t) is the cluster output power at time t; f1 is the time correlation function of the wind power cluster; θ is the prediction time scale; E clu (t) is the error at time t; f2 is the spatial correlation function of the cluster, and P Wi (t - θ) is the power of the i-th wind farm in the cluster at time t - θ.

[0121] Step 6: Initialize the weights of the neural network, set the maximum number of iterations K = 40, and the current number of iterations k0 = 1.

[0122] Specifically:[[]]

[0123] A) Initialize the convolution layer with Gaussian distribution, sampling from a Gaussian distribution with a mean of 0 and a variance of 1 as the initial value;

[0124] B) Initialize the Scala factor of the BN layer to 1 and the shift factor to 0;

[0125] C) Call the zero_state function (the existing initialization function in TensorFlow) to implement the initialization of the BILSTM network.

[0126] Step 7: Construct a point convolutional neural network, and input the cluster feature matrix into the point CNN for spatial feature extraction.

[0127] Step 8: Input the feature vectors extracted by the point CNN network into the Attention layer improved by combining the SSA algorithm, assign different weights to the transition feature vectors of the hybrid neural network, and then input the weighted feature vectors into the BILSTM layer to output the training results of the I-CNN-BILSTM neural network. Read the training loss curve and error curve, observe the vertical distance between the training set and validation set loss curves during convergence, and combine the absolute error conditions of the training set and validation set to visually evaluate the convergence performance of the network prediction results.

[0128] The following are the convergence situations represented by three common fitting states:

[0129] 1) When the training set loss curve does not decrease, this is the underfitting state and it is not convergent.

[0130] 2) When the training set loss curve continues to decrease and the validation set loss curve no longer decreases at a certain moment, this is the overfitting state, which is a convergent state but not a perfect convergence.

[0131] 3) When there is no obvious distance between the training set and validation set loss curves, it is the perfect fitting state and it is perfectly convergent.

[0132] The present invention improves and optimizes the hybrid network, and the specific content is as follows:

[0133] Since there are a large number of cluster features input into the model, in order to highlight the important spatio-temporal features of the clusters, the present invention uses an improved Attention mechanism to assign different weights to the over-vectors. In the traditional Attention mechanism, the information input into the network first will be covered by the information input later, and the vector cannot fully represent the information of the entire sequence. In view of this deficiency, the present invention proposes an improved Attention mechanism combined with the SSA algorithm, which makes up for the deficiency of the network's attention to important relevant features and improves the network's attention degree to spatio-temporal relevant features. As Figure 5 shown.

[0134] The SSA algorithm is used to generate the optimal parameter combination in the attention layer. By simulating the chain group behavior of leaders and followers in the salp swarm, it searches for the parameter combination with the optimal loss and returns the result. The specific steps are as follows:

[0135] 1) Initialize the weight parameters and calculate the fitness of the current parameters;

[0136] 2) Set the parameter position with the optimal fitness as the target position. The first half of the parameter combinations sorted according to fitness are the leaders, and the remaining parameter combinations are the followers.

[0137] 3) Set the update method for the leader position:

[0138]

[0139] where is the position of the leader in the j-th dimensional space; F j is the position of the food source in the j-th dimensional space; ub j , lb j are the upper and lower limits corresponding to the j-th dimensional position respectively; c1, c2, c3 are control parameters. Among them, c1 is the convergence factor, and its expression is:

[0140]

[0141] where l is the current iteration number and L is the maximum iteration number.

[0142] 4) Set the update method for the follower position as:

[0143]

[0144] where, X j i is the position of the i-th follower in the j-th dimension, t is the time, a is the acceleration, and v0 is the initial velocity. Considering that t is iterative, if t = 1 and v0 = 0 during the iteration process, the follower position update method can be expressed as:

[0145]

[0146] where, is the position of the (i - 1)-th follower in the j-th dimension.

[0147] 5) Repeat the above process until the set maximum iteration number is reached. Calculate the prediction error from the updated parameters and output the optimal attention parameters.

[0148] In a neural network model, if there are too many model parameters and too few training samples, the trained model is very likely to overfit. The latest research shows that traditional dropout, which randomly selects neurons, significantly increases the computational amount and time, affecting the speed of the prediction algorithm. Targeted dropout is an improved dropout algorithm that is selective for neurons. It ranks the weights or neurons according to a measure that rapidly approximates the importance of the weights and applies dropout to the elements with lower importance. The specific implementation method is as follows:

[0149] 1) Pruning operation: For a parameterized neural network, through unit pruning and weight pruning, find the optimal parameter θ * , such that the loss function ε(W(θ * )) is as small as possible while retaining the highest order of magnitude of k weights in the neural network. The following are the operation formulas for unit pruning and weight pruning respectively:

[0150]

[0151]

[0152] In the formula: ε(W c (θ)) is the network loss function, W c is the neural network model parameter matrix, argmax-k is the function that returns the largest k elements among all elements, w o is the column vector of the o-th column of the weight matrix W, W io is the element in the i-th row and o-th column of the weight matrix, N col and N row represent the number of columns and rows of the parameter matrix respectively.

[0153] 2) Random process: Introduce the targeting ratio γ and the deletion probability α. Among them, the targeting ratio γ means that the smallest γ|θ| weights will be selected as the candidate weights for dropout, and then the weights in the candidate set will be independently removed with the deletion probability α.

[0154] Step Nine: If the iteration reaches the maximum number of iterations (k0>K), the iteration terminates, and the I-CNN-BILSTM network parameters are output. Otherwise, let k0 = k0 + 1 and go to Step Five.

[0155] Step 10: Perform short-term power prediction for the wind power cluster using the I-CNN-BILSTM network trained according to the above steps to obtain the predicted power at the moment to be predicted. Input the cluster feature matrix into the point CNN network, obtain global features through the edge convolutional layer and the max pooling layer, then input them into the point convolutional layer to extract local features, and the convolutional network outputs a one-dimensional transitional feature vector. Use the improved Attention mechanism to assign weights to the transitional feature vector, and finally input it into the three-layer BILSTM neural network to obtain the prediction result through the fully connected layer.

[0156] Step 11: Use the above steps to perform power prediction for all spatio-temporally related sub-clusters. The aggregated power prediction result of the region is the comprehensive superposition result of all spatio-temporally related clusters, and the calculation method is as follows:

[0157]

[0158] In the formula, P region (t) is the predicted power of the region at time t, P clu-k (t) is the predicted power of cluster k at time t, and N is the total number of spatio-temporally related clusters.

[0159] The present invention considers the spatio-temporal correlation of the regional wind power cluster, proposes a cluster classification method based on spatio-temporal multiple similarities, and adopts the joint distribution matching of the daily power curve pattern to match the main wind power cluster related to the long-term output. Secondly, considering the dynamic change of the correlation in the time domain, a similarity measurement method for the fluctuation trend of the short-term power sequence is proposed. By segmentally identifying the power fluctuation type and the trend similarity distance of the type sequence, wind power sub-clusters that take into account both long-term pattern correlation and short-term fluctuation similarity are classified, and the wind farms in the sub-clusters have strong spatio-temporal correlation at the prediction moment. The proposed I-CNN-BILSTM hybrid neural network model considers the spatial position relationship information of the point cloud distribution of the cluster wind farms, synthesizes the characteristic data of each station, fully extracts the spatial features and time-dependent relationships of the cluster, and introduces the Attention mechanism improved by combining the SSA algorithm and the dropout algorithm with a pruning strategy, which is beneficial to the fitting optimization of the hybrid model and improves the accuracy of the model prediction result.

[0160] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any staff familiar with the technical field of the present invention can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A short-term power prediction method for a regional wind power cluster, characterized in that It includes the following steps: 1) Obtain the historical annual power data of the regional wind farm, form a daily power sample set of the wind farm with days as the unit, perform similarity clustering on the daily power sample set, extract the typical forms of the daily power curves of the wind farm on a long-term scale, and define different power curve form features as different daily power patterns; 2) Form a daily power pattern sequence based on the annual power data, establish a classification matrix Q for the daily power pattern sequences of different wind farms, statistically calculate the probability distribution of the joint patterns of different wind farms in the classification matrix to form an estimated probability matrix P and calculate the entropy value of the estimated probability matrix, and classify the regional multi-wind farms into main clusters with similar long-term output patterns according to the entropy value; 3) Extract the historical short-term power data of each wind farm in the main cluster before the prediction moment, divide and identify the types of power fluctuation segments, solve the similarity of the fluctuation trends by using the similar distance of the fluctuation trends, and then cluster into sub-clusters with strongly correlated short-term fluctuation trends; 4) Extract the historical meteorological data and spatial coordinate data of the multi-wind farms in the sub-cluster to form an input feature matrix, and establish a short-term power prediction model based on the improved I-CNN-BILSTM hybrid neural network; 5) Initialize the weights of the neural network and set the maximum number of iterations; 6) Construct a point convolutional neural network based on point cloud input, and input the feature matrix into the point convolutional neural network to extract global features and local features; 7) Input the features extracted by the point CNN network into the improved Attention mechanism combined with the SSA algorithm, assign different weights to the transition feature vectors of the neural network model, input the weighted transition feature vectors into the BILSTM layer according to the time steps, output the training results of the improved I-CNN-BILSTM hybrid neural network, read the training loss curve and error curve, and evaluate the convergence performance of the network prediction results according to the vertical distance between the loss curves of the training set and the validation set during the convergence process and the magnitude of the absolute error; 8) Determine whether the current number of iterations has reached the set number of iterations. If so, terminate the iteration and output the parameters of the hybrid neural network. Otherwise, increment the number of iterations by one and go to step 4); 9) Use the improved I-CNN-BILSTM network trained in the above steps to perform short-term wind power prediction for the corresponding sub-cluster and obtain the predicted power of the sub-cluster; 10) Determine whether the power prediction of all sub-clusters has been completed. If so, perform step 11). Otherwise, go to step 4) to predict the remaining sub-clusters; 11) Perform superposition calculation on the power prediction values of all sub-clusters to obtain the overall wind power prediction result of the region; The specific content of the short-term power prediction model for the wind power cluster based on the improved I-CNN-BILSTM hybrid neural network is as follows: a) Use the point CNN neural network to extract the local and global spatial features of the cluster; b) Use the improved Attention mechanism combined with the SSA algorithm to achieve weight allocation between time series; c) Introduce the Targeted dropout algorithm in the BILSTM network to selectively eliminate neurons; The specific steps for extracting the local and global spatial features of the cluster using the Point CNN neural network are as follows: a1) Establish an input feature matrix for the wind power cluster, including the historical meteorological data and spatial coordinate data of each wind farm in the cluster; a2) Encode the irregular wind farm location information into the weights of the X transformation matrix, and arrange the features of the associated points in the point cloud; a3) Construct a point cloud data feature map G = (α, β), where the vertices α = {1, 2, …, n} represent the information of each data point in the point cloud, and the edges β represent the association information between every two data points in the graph; a4) Set the edge function to e ij = h(x i , x j - x i , θ), whose global features are provided by numerous data points that make up the point cloud, and local features are provided by the correlation information between two points in the point cloud, where h(x i , x j - x i , θ) is the edge function for edge feature extraction, θ is the parameter to be learned and optimized in the edge function, x i and x j are sample data points; The specific steps for implementing the weight allocation between time series using an improved Attention mechanism combined with the SSA algorithm are as follows: b1) Initialize the weight W of the attention layer; b2) Input the attention layer weight W into the Salp Swarm Algorithm, and transmit the optimized weight set to the Bidirectional Long Short-Term Memory (BILSTM) neural network. In the BILSTM neural network, a corresponding loss value is generated according to the prediction error in the network; b3) Select the optimal attention weight set according to the loss situation; The specific steps for introducing the Targeted dropout algorithm in the BILSTM network to selectively eliminate neurons are as follows: c1) Perform pruning operations on the BILSTM network neurons according to the methods of weight pruning and unit pruning. The calculation formula is: where: ε(W c (θ)) is the network loss function, W c is the neural network model parameter matrix, argmax-k is the function that returns the k largest elements among all elements, w o is the column vector of the o-th column of the weight matrix W, W io is the element in the i-th row and o-th column of the weight matrix, N col and N row represent the number of columns and rows of the parameter matrix respectively; c2) Introduce the targeted ratio γ and the deletion probability α. Select the smallest γ|θ| weights as the candidate weights for dropout, and then independently remove the weights in the candidate set with the deletion probability α.

2. The short-term power prediction method for the regional wind power cluster according to claim 1, characterized in that In step 1), the ACFSFDP algorithm is used to perform similarity clustering on the daily power sample set.

3. The short-term power prediction method for the regional wind power cluster according to claim 2, wherein The specific steps for using the ACFSFDP algorithm to perform similarity clustering on the daily power sample set are as follows: 11) Form a sample set in days from the normalized historical wind power data, and calculate the local density ρ of the samples i , topological relationship, and distance δ i ; 12) Combine the product of the local density ρ of the sample i and the distance δ i to form a comprehensive evaluation index γ, and select the clustering center based on this comprehensive evaluation index; 13) Establish an adaptive index F for determining the optimal number of clustering clusters K , gradually increase the number of cluster centers, and subtract one from the number of cluster centers corresponding to the only maximum peak of the adaptive index as the optimal number of clustering clusters.

4. The short-term power prediction method for the regional wind power cluster according to claim 1, characterized in that The specific steps for step 2) are as follows: 21) Form a daily power pattern sequence for the historical one-year daily power data of the wind farm according to the clustering results of step 1). Use the multiple groups of daily power pattern sequences of all wind farms in the region to form a classification matrix Q. The row vector of the matrix is the daily power pattern sequence of the corresponding wind farm in N days, and the k-th column vector of the matrix represents the joint pattern of the daily power curves of different wind farms; 22) Statistically classify the probabilities of the occurrence of each joint pattern in matrix Q to form an estimated probability matrix P. The element p of the estimated probability matrix P ij represents the probability of the occurrence of (i,j) in each column of Q T . Calculate the entropy value H of the estimated probability matrix: H = -∑p ij ln(p ij ). Wind farms with entropy value H less than the set threshold H min are classified into the same cluster; and classify multiple wind farms in the region into main clusters with similar long-term output patterns according to the magnitude of the entropy value.

5. The short-term power prediction method for the regional wind power cluster according to claim 1, wherein The specific steps for using the fluctuation trend similarity distance to solve the fluctuation trend similarity in step 3) are as follows: 31) Define characteristic parameters for describing power fluctuations, including the volatility S reflecting the power change trend t , the standard deviation of fluctuations B reflecting the severity of power fluctuations, the fluctuation degree W reflecting the frequency of power fluctuations, and the high-power ratio Q reflecting the magnitude of the actual output value h , the low-power ratio Q l ; 32) Extract the historical short-term power sequence of the wind farm in the three days before the prediction moment. According to the fluctuation parameters, divide the power time series into multiple different types of fluctuation segments, including stable power fluctuation, ramp power fluctuation, oscillating power fluctuation, and peak-valley power fluctuation, which are defined as type-a, type-b, type-c, and type-d fluctuations respectively, and convert the power time series into a fluctuation type sequence Z P ={z1,…,z i ,…}, where z i ∈{a,b,c,d}; 33) The fluctuation type matrix Z is composed of the fluctuation type sequences of all wind farms in the main cluster. For the power time series P1 and P2, their fluctuation type sequences are and where z i , z j ∈{a, b, c, d}, calculate the fluctuation type matching distance d(z i , z j ): The distance matrix D of two fluctuation sequences is formed by the fluctuation type matching distance k×l and the cumulative matrix R = {r(i, j)} k×l D k×l = {d(z i , z j )} k×l , i = 1, 2, … k. j = 1, 2, … l where r(0, 0) = 0, r(i, 0) = r(0, j) = ∞; The similarity distance of the fluctuation trends of different fluctuation type sequences is Set the minimum distance threshold L min , and the sequences with a distance L less than the threshold are clustered into a sequence cluster with similar fluctuation trends, and the corresponding wind farm groups are spatio-temporal correlation sub-clusters that take into account the long-term output pattern and the similarity of short-term output fluctuations.

6. The short-term power prediction method for the regional wind power cluster according to claim 1, characterized in that, The historical meteorological data includes wind speed, wind direction, temperature, and historical power sequences, and the spatial coordinate data includes longitude, latitude, and altitude.

Images