A distributed wind turbine power prediction method and related device

Abstract

The embodiment of the application discloses a kind of distributed wind turbine power prediction method and related equipment, the method includes: according to the Bayes theory in influencing factor set selects the target factor that most influences distributed wind power;Determine target feature based on the target factor, establish basic data set including time feature and space feature according to the target feature, and in historical sample set, the target sample that similarity is higher than threshold value with the point to be predicted feature is filtered out, constructs training set;The basic data set and the training set are imported into the preset space-time graph neural network, and the distributed wind power prediction value is obtained.Can mine the hidden time feature of each unit historical sample, also combine the space feature between different units, fully learn the information contained in data, improve the prediction accuracy of distributed wind power.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind turbine power prediction, and particularly relates to a decentralized wind turbine power prediction method and related equipment. BACKGROUND

[0002] At present, the wind power prediction method widely used is an artificial neural network prediction model based on a statistical method, which considers weather factors such as wind speed, wind direction, air pressure, temperature, and air density that affect wind power, increases the dimension of model processing and operation period, greatly reduces the operation speed and accuracy of the model, and the selection of hyperparameters such as the number of hidden layer nodes and weight coefficients of the neural network also has certain influence on the wind power prediction result. Unfit or overfitting may occur due to inappropriate hyperparameters. SUMMARY

[0003] Therefore, the present application provides a decentralized wind turbine power prediction method and related equipment to solve the problems of low prediction efficiency and low accuracy of the artificial neural network prediction model based on the statistical method in the prior art. To achieve one or part or all of the above purposes or other purposes, the present application provides a decentralized wind turbine power prediction method, which comprises: selecting a target factor that has the greatest influence on decentralized wind power from an influence factor set according to Bayesian theory;

[0004] determining a target feature based on the target factor, establishing a basic data set including time features and space features according to the target feature, and screening out target samples with a similarity higher than a threshold value to a to-be-predicted point feature from a historical sample set to construct a training set;

[0005] importing the basic data set and the training set into a preset spatio-temporal graph neural network to obtain a decentralized wind power prediction value.

[0006] Optionally, the step of selecting a target factor that has the greatest influence on decentralized wind power from an influence factor set according to Bayesian theory comprises:

[0007] selecting different factors that affect wind power to form the influence factor set;

[0008] determining the prior probability of the influence factor based on historical power data by a probability and statistics method;

[0009] randomly selecting influence factors in the influence factor set, and the selected influence factors form an input feature vector;

[0010] calculating an effectiveness verification error by using an N-fold cross-validation algorithm to obtain a likelihood function of the input feature vector;

[0011] calculating a posterior probability of the input feature vector based on the likelihood function by using a variational inference method;

[0012] comparing the posterior probabilities of the input feature vectors of at least two groups, and selecting an input feature vector with the maximum posterior probability as a target input feature vector;

[0013] selecting a target factor having the greatest impact on the distributed wind power according to the target input feature vector.

[0014] Optionally, before the step of establishing a basic data set including time features and space features according to the target feature, and screening target samples with a similarity higher than a threshold from a historical sample set to construct a training set, the method further comprises:

[0015] performing outlier rejection on the target feature based on a local outlier factor algorithm to obtain an initial target feature;

[0016] performing missing value reconstruction on the initial target feature according to a missing value filling method of a generative adversarial network to obtain an intermediate target feature;

[0017] performing normalization processing on the intermediate target feature to obtain a final target feature.

[0018] Optionally, the step of establishing a basic data set including time features and space features according to the target feature comprises:

[0019] taking the target feature and wind turbine output historical data as time features of each wind turbine;

[0020] representing relative relationships of spatial positions of each distributed wind turbine by an undirected topological net to obtain space features of each wind turbine, the undirected topological net having an adjacency matrix with weights determined by distances between wind turbines;

[0021] constructing a basic data set based on the time features and the space features.

[0022] Optionally, the step of screening target samples with a similarity higher than a threshold from a historical sample set to construct a training set comprises:

[0023] extracting a target sample set with a similarity higher than a threshold from the historical sample set by using a grey correlation algorithm based on entropy weight method weighting;

[0024] arranging the target sample set according to an original time sequence, sequentially sampling one sample from each target sample set without replacement to form a time sequence sample with a preset step length, and constructing an initial training set according to the time sequence sample;

[0025] The relative relationship of the spatial positions of each distributed wind turbine is represented by an undirected topological net, and the spatial characteristics of each wind turbine are obtained, wherein the undirected topological net has a weight adjacency matrix, and the weight value is determined by the distance between each wind turbine.

[0026] The spatial characteristics of each wind turbine are added to the initial training set to obtain a target training set.

[0027] Optionally, before the step of importing the preset spatio-temporal graph neural network with the basic data set and the training set to obtain the distributed wind power prediction value, the method further comprises:

[0028] The hyperparameters of the preset spatio-temporal graph neural network are optimized by adopting a Bayesian optimization method, and the optimal calculation dimension and hyperparameters are selected;

[0029] The preset spatio-temporal graph neural network is constructed based on the calculation dimension and the hyperparameters, and the preset spatio-temporal graph neural network comprises a graph convolution network for extracting spatial characteristics of each wind turbine, an LSTM model for extracting time characteristics of each wind turbine, a residual layer for preventing network degradation, and a prediction convolution layer for outputting a distributed wind power prediction value.

[0030] Optionally, the step of optimizing the hyperparameters of the preset spatio-temporal graph neural network by adopting the Bayesian optimization method comprises:

[0031] An initial sample point is randomly generated within a set hyperparameter range and is input into a Gaussian process to obtain a target function loss value;

[0032] The Gaussian model is corrected based on the target function loss value until the target function loss value meets a preset requirement, and the hyperparameter value when the target function loss value meets the preset requirement is obtained.

[0033] On the other hand, the application provides a distributed wind turbine power prediction device, and the maintenance device comprises:

[0034] A selection module is configured to select a target factor having the greatest influence on the distributed wind power from the influence factor set according to the Bayesian theory;

[0035] A construction module is configured to determine a target feature based on the target factor, establish a basic data set comprising time characteristics and spatial characteristics according to the target feature, and select a target sample having a similarity higher than a threshold to a to-be-predicted point feature from a historical sample set to construct a training set.

[0036] A prediction module is configured to import the basic data set and the training set into a preset spatio-temporal graph neural network to obtain a distributed wind power prediction value.

[0037] In a third aspect, an electronic device is provided, comprising a processor, a memory, and a bus, the memory storing machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, the machine readable instructions are executed by the processor to perform the steps of the distributed wind turbine power prediction method as described above.

[0038] In a fourth aspect, a computer readable storage medium is provided, the computer readable storage medium storing a computer program, when the computer program is run by a processor, the steps of the distributed wind turbine power prediction method as described above are performed.

[0039] The implementation of the embodiments of the present application will have the following beneficial effects:

[0040] The factors affecting wind power are selected in combination with historical data, and the hyperparameters are optimized by adopting the Bayesian optimization method to select the optimal calculation dimension and hyperparameters, thereby greatly reducing the required training data and improving data efficiency, so that the wind power prediction is efficient and accurate. The training set of historical data similarity is constructed based on the weighted grey correlation method to correct the prediction result of the basis data, so that the model prediction accuracy is improved. The wind speed, wind direction, temperature and other meteorological features related to the distributed wind turbine output are very small in a specific regional range, so the spatial and geographical location information of each unit in the region can be used to obtain the output size relationship between each unit to extract the spatial features of the wind turbine, and the distributed wind short-term power prediction is performed based on the space-time graph neural network, which can fully excavate the hidden time features of each unit historical sample, effectively capture the spatial features between different units, fully learn the information contained in the data, and improve the prediction accuracy of the distributed wind power. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0042] Among them:

[0043] Figure 1 is a flowchart of a distributed wind turbine power prediction method provided by an embodiment of the present application;

[0044] Figure 2 is a flowchart of another distributed wind turbine power prediction method provided by an embodiment of the present application;

[0045] Figure 3is a flow chart of another distributed wind turbine power prediction method provided by the embodiment of the present application;

[0046] Figure 4 is a structural schematic diagram of a distributed wind turbine power prediction device provided by the embodiment of the present application;

[0047] Figure 5 is a structural schematic diagram of an electronic device provided by the embodiment of the present application;

[0048] Figure 6 is a structural schematic diagram of a storage medium provided by the embodiment of the present application. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0050] As shown in Figure 1 , the embodiment of the present application provides a distributed wind turbine power prediction method, comprising:

[0051] S101, selecting a target factor having the greatest influence on distributed wind power from an influence factor set according to Bayesian theory;

[0052] In a possible implementation, the step of selecting a target factor having the greatest influence on distributed wind power from an influence factor set according to Bayesian theory comprises:

[0053] Selecting different factors influencing wind power to form the influence factor set;

[0054] Based on historical power data, determining the prior probability of the influence factor by a probability and statistics method;

[0055] Randomly selecting the influence factor in the influence factor set, and the selected influence factor forming an input feature vector;

[0056] Calculating the validity verification error by using an N-fold cross-validation algorithm to obtain a likelihood function of the input feature vector;

[0057] Based on the likelihood function, calculating the posterior probability of the input feature vector by using a variational inference method;

[0058] Comparing the posterior probabilities of at least two groups of input feature vectors, and selecting the input feature vector with the largest posterior probability as a target input feature vector;

[0059] According to the target input feature vector, a target factor having the greatest influence on the distributed wind power is selected.

[0060] For example, as shown in the formula (1), M factors having influences on wind power prediction are set, including wind speed, temperature, humidity, historical power, etc. Figure 2 W represents a variable in a random variable space Ω, J = 1, 2, K, M, and there are N samples in the sample space n = 1, 2, K, N.

[0061] Determine the prior probability distribution: the prior probability represents the probability distribution estimation of the variable , which is obtained based on historical power data by a probability statistical method.

[0062] Determine the likelihood function: calculate by using the N-fold cross-validation effectiveness verification algorithm. The core idea of cross-validation is to divide the data set multiple times, and take the average of the results of multiple evaluations. Divide the training sample set into n subsets, and select one subset as the effective set each time, and the remaining n-1 subsets as the training set. After n times of optimization calculation, the average value of the effective set error is obtained, and the specific formula is as follows:

[0063]

[0064]

[0065] In the formula, E i is the calculation error; R valid_j [f train ] is the relative average error of the regression function f train on the effective set obtained from the training set; N valid is the number of samples in the effective set Z valid ; y j is the actual power value; and is the predicted power value.

[0066] Finally, the input vector when the error E i is the smallest and the likelihood function is the largest is selected, and the expression of the likelihood function is as follows:

[0067]

[0068] In the formula, L is the likelihood ratio; and E i is the error. is the percentage, and the unit is %.

[0069] Calculate the posterior probability: the discrete random variable form of the Bayes theorem can be expressed as the following formula:

[0070]

[0071] wherein, is the posterior probability.

[0072] As an approximate inference method, variational inference is used to make the variational distribution q φ ≈q(W|θ) approximate First, the KL divergence D KL (q φ ||P) is calculated to measure the distance between different probability distributions:

[0073]

[0074] Further, the function ELBO is calculated, and the ELBO maximization is realized by the backprop algorithm, so as to indirectly obtain

[0075]

[0076] wherein, P(E) is the data prior probability.

[0077] The input vector From the selected feature set composed of M features affecting power, different feature combinations are randomly selected to form the input vector Finally, the input vector with the maximum posterior probability corresponds to the target factor.

[0078] S102, determine the target feature based on the target factor, establish a basic data set including time characteristics and space characteristics according to the target feature, and select target samples with a similarity higher than a threshold to the to-be-predicted point characteristics from the historical sample set to construct a training set.

[0079] For example, a basic data set including time characteristics and space characteristics of a distributed wind turbine is constructed, so that the prediction model can not only improve the accuracy of power prediction by learning the regularity of historical data, but also use the correlation of spatial position to form certain constraints on the mutual influence between different time characteristics, thereby improving the interpretability of the model.

[0080] S103, import the basic data set and the training set into a preset spatio-temporal graph neural network to obtain a distributed wind power prediction value.

[0081] For example, the basic data set and the training set are imported into a preset spatio-temporal graph neural network, spatial features are extracted by using a graph convolution network, time features are extracted by using an LSTM model, and finally distributed wind power prediction is realized through a residual layer and a prediction convolution layer.

[0082] The factors influencing the wind power are selected according to historical data, and the Bayesian optimization method is adopted to optimize the hyperparameters, so that the optimal calculation dimension and hyperparameters are selected, the required training data is greatly reduced, the data efficiency is improved, and the wind power prediction is efficient and accurate. The training set of historical data similarity is constructed based on the weighted grey correlation method to correct the prediction result of the basis data, so that the prediction accuracy of the model is improved. The wind speed, wind direction, temperature and other meteorological characteristics related to the output of the dispersed wind turbine are very small in a specific regional range, so the spatial and geographical position information of each unit in the region is used to obtain the output size relationship between each unit, the spatial characteristics of the wind turbine are extracted, and the short-term power prediction of the dispersed wind turbine is carried out based on the space-time graph neural network. The hidden time characteristics of each unit historical sample can be fully mined, the spatial characteristics between different units can be effectively captured, the information contained in the data can be fully learned, and the prediction accuracy of the dispersed wind power is improved.

[0083] In a possible implementation, before the step of establishing the basis data set including the time characteristics and the space characteristics according to the target characteristics, and screening the target sample with a similarity higher than a threshold from the historical sample set to be predicted, the step further includes:

[0084] Performing outlier detection on the target characteristics based on a local outlier factor algorithm to obtain an initial target characteristic;

[0085] Performing missing value reconstruction on the initial target characteristic based on a missing value filling method of a generative adversarial network to obtain an intermediate target characteristic;

[0086] Performing normalization processing on the intermediate target characteristic to obtain a final target characteristic.

[0087] For example, the local outlier factor algorithm is used to detect outliers of wind power raw data. A feature is selected from the feature set selected based on the Bayesian theory in step one, and a value under the feature is selected as a sample p. First, the k-distance of the sample p is calculated, that is, the distance of the point farthest from p, excluding p. The object set with a distance less than or equal to the k-distance of the object p is defined as the k-distance neighborhood N k (p) of the object p. The reach-distance of the object p to the object o is calculated, the object o and the object p belong to the same sample space, and at least k objects q satisfy that the distance between the object p and the object o is less than or equal to the distance between the object p and the object o in the sample space. At most, there are k-1 objects q with a distance less than the distance between the object p and the object o. The reach-distance of the object p to the object o is the larger one of the actual distance and the k-distance of the object o, and the formula is as follows:

[0088] reach-distance (p, o) = max{k-distance (o), ||p - o||}

[0089] The local reachability density of object p, denoted as local-reachability-density (p), is then calculated as:

[0090]

[0091] where lrd k (p) denotes the local reachability density of p, |N k (p) | denotes the number of samples in the k-distance neighborhood of p.

[0092] The local outlier factor of sample p, denoted as local-outlier-factor (p), is finally calculated as:

[0093]

[0094] where LOF k (p) denotes the local outlier factor of p, lrd k (o) denotes the local reachability density of o. The closer the local outlier factor is to 1, the more likely the sample p belongs to the same cluster as its neighbors; the smaller the value is than 1, the more likely p is a dense point; the larger the value is than 1, the more likely p is an outlier.

[0095] The process of missing value reconstruction is as follows: the missing value reconstruction is performed by using a missing value filling method based on a generative adversarial network. The generative adversarial network is composed of a generator G and a discriminator D. Firstly, the parameters θ d of the discriminator D and the parameters θ g of the generator G are initialized; secondly, m samples {x 1 , x 2 ,…, x m} are sampled from the feature samples after removing outliers, m noise samples {z 1 , z 2 ,…, z m} are sampled from a prior distribution noise, and m generated samples The generator G is fixed, the discriminator D is trained to distinguish the real samples and the generated samples as accurately as possible, and the correct samples and the generated samples are distinguished as large as possible; subsequently, the discriminator is updated for k times, and then the parameters of the generator are updated once using a smaller learning rate, the generator is trained to make it as possible to reduce the difference between the generated samples and the real samples, and to make the discriminator distinguish errors as much as possible; finally, the discriminator parameters are updated multiple times to make the discriminator distinguish whether the samples come from the generator output or the real sample set. The missing values are filled with the output of the generator to obtain a complete and clean feature data set.

[0096] The data normalization process is as follows:

[0097] The features with the removed outliers and the reconstructed missing values are normalized in the application, the data range after preprocessing is [0, 1], and the normalization formula is as follows:

[0098]

[0099] In the formula, X is an actual value of a sample feature, X new is a normalized value of the sample feature, X max is a maximum value of the sample feature, X min is a minimum value of the sample feature, and X max -X min is a range.

[0100] In a possible implementation, the step of establishing the basic data set comprising the time feature and the space feature according to the target feature comprises:

[0101] The target feature and the wind turbine output historical data are taken as the time feature of each wind turbine.

[0102] The relative relationship of the spatial positions of each distributed wind turbine is represented by an undirected topological net to obtain the space feature of each wind turbine, and the undirected topological net has an adjacency matrix, and the weight value is determined by the distance between the wind turbines.

[0103] The basic data set is constructed based on the time feature and the space feature.

[0104] For example, the climate feature and the wind power output historical data after data preprocessing are taken as the time feature of the basic sample set, and the relative relationship of the spatial positions of each distributed wind turbine is represented by an undirected topological net G=(V,E,A), wherein |V|=N, N represents the number of nodes in the topological net, that is, the number of distributed wind turbines, E is a set of edges, that is, the connectivity between nodes, and the application adopts a full connection mode, and A is a weight matrix, that is, the distance between the nodes. N×XA weighted adjacency matrix representing graph G, the weight value is determined by the distance between each unit. Each node in the topological network G records the time series of the f-dimensional features, where f∈(1,...,F), x t,i ∈R F represents all feature values of the i-th node at time t, F is the feature dimension of the distributed wind power, so the basic data set of the application is a space-feature-time spatio-temporal tensor,

[0105] In one possible implementation, the step of screening target samples with a similarity higher than a threshold to the to-be-predicted point features from the historical sample set to construct a training set comprises:

[0106] extracting a target sample set with a similarity higher than a threshold to the to-be-predicted point features from the historical sample set using a grey correlation algorithm based on entropy weight method;

[0107] arranging the target sample set according to the original time sequence, and sequentially sampling one sample from each target sample set without replacement to form a time sequence sample with a preset step size, and constructing an initial training set according to the time sequence sample;

[0108] representing the relative relationship of the spatial positions of each distributed wind turbine using an undirected topological network to obtain the spatial features of each wind turbine, the undirected topological network having a weighted adjacency matrix, and the weight value being determined by the distance between each unit;

[0109] adding the spatial features of each wind turbine to the initial training set to obtain a target training set.

[0110] For example, a historical sample set with similar weather features to the to-be-predicted point is extracted from the historical sample set after data preprocessing using a grey correlation algorithm based on entropy weight method to construct a sample set. The weighted grey correlation method calculates the grey correlation degree by assigning appropriate weight values to each feature, finds the features with a large degree of influence, and realizes the comprehensive evaluation of the similarity between the historical sample and the to-be-predicted sample, and the specific steps are as follows:

[0111] selecting the to-be-predicted sample X0 as the mother sequence, and the first n historical samples X of the to-be-predicted sample as the subsequence, X ij represents the j-th feature in the i-th historical sample, m is the number of sample features, and X0 and X are as follows:

[0112] X0=[X 01 ,X 02 ,K X 0n ]

[0113]

[0114] calculating the grey correlation coefficient ξ between the mother sequence and the subsequenceik After that, the grey correlation matrix ξ is constructed, and the calculation formula is as follows:

[0115]

[0116] In the formula, ξ ik represents the grey correlation coefficient of the kth factor of the ith sample; p is a resolution coefficient for determining the distinguishing ability, and is usually taken as 0.5.

[0117]

[0118] In the formula, ξ is the grey correlation matrix between the sub-sequence and the mother sequence.

[0119] The entropy weight method is used to calculate the weight of each index, and a weight vector is formed, and the calculation formula is as follows:

[0120]

[0121]

[0122]

[0123] γ=[γ1,γ2,K,γ m ]

[0124] In the formula, r ij is the proportion of the jth feature of the ith sample; E j is the information entropy of the jth feature; γ j is the weight of the jth feature; and γ is the weight vector determined by the entropy weight method.

[0125] The weighted grey degree is obtained by ξ′=γξ, the weighted grey correlation degrees of the historical samples are arranged from large to small, and the top k historical samples in the order are selected as the similar historical sample set of the to-be-predicted sample.

[0126] According to the prediction scale requirement, a training set is constructed. If the input of the prediction model is a time sequence with a step length of T i , and the output is a time sequence with a step length of T o , each distributed wind turbine extracts the similar historical sample set of the to-be-tested t samples and the previous T i samples according to the above steps, arranges each sample set in the original time sequence, and samples one historical sample from each set to form a step length of (T oThe time series samples of the historical data similarity-based training set of each unit are constructed by sampling k times. Then, the spatial features of each unit are added according to the method steps of generating the basic data set, the sample features in the training set are expanded into a space-feature-time spatio-temporal tensor, and the historical data similarity-based training set is perfected.

[0127] In a possible implementation, before the step of importing the basic data set and the training set into the preset spatio-temporal graph neural network to obtain the distributed wind power prediction value, the method further includes:

[0128] The hyperparameters of the preset spatio-temporal graph neural network are optimized by adopting a Bayesian optimization method, and the optimal calculation dimension and hyperparameters are selected.

[0129] The preset spatio-temporal graph neural network is constructed based on the calculation dimension and the hyperparameters, and the preset spatio-temporal graph neural network includes a graph convolution network for extracting spatial features of each wind turbine, an LSTM model for extracting time features of each wind turbine, a residual layer for preventing network degradation, and a prediction convolution layer for outputting a distributed wind power prediction value.

[0130] For example, the graph convolution layer captures the spatial features between different nodes by using Chebyshev polynomials. In order to fully utilize the graph structure information in the topology network, a spectral graph-based graph convolution network is used to extract the spatial features of the distributed wind turbine power in each time step, and a Laplacian matrix is used to represent the corresponding topology graph, and the spectral graph convolution is defined as follows:

[0131] g θ * G x = Ug θ U T x

[0132] where g θ is a convolution kernel, G the graph convolution operation is represented by x, the input data is represented by U, and the Fourier basis is represented by U. The normalized Laplacian matrix of the distributed wind power topology network is shown as follows:

[0133]

[0134] where A is an adjacency matrix, Λ represents a diagonal matrix composed of eigenvalues of the basic data set in step two, I N is a unit matrix, D represents a degree matrix of a node, and D ∈ R N×N The spectral graph convolution defined on the distributed wind power topology network can be obtained as follows:

[0135] g θ * G x = g θ (L)x = g θ(UΛU T x=Ug θ (Λ)U T x

[0136] When the number of distributed wind turbines is too large, the required calculation memory and time consumption for eigenvalue decomposition of L are very large, so the present application adopts K-order Chebyshev polynomials for parameter simplification:

[0137]

[0138] In the formula, θ' is a vector composed of polynomial coefficients, is expressed as:

[0139]

[0140] λ max is the maximum eigenvalue of the Laplace matrix L. For Chebyshev polynomials Tk(x), it satisfies:

[0141] T0(x)=1

[0142] T1(x)=x

[0143] T k (x)=2xT k-1 (x)-T k-2 (x)

[0144] Based on the above, the graph convolution is extended to the topology network multi-channel graph signal input X t ∈R N×F , based on the graph convolution information propagation formula H (l) =f(H (l-1) ,A), the output of the lth layer is defined as:

[0145]

[0146] Where σ is an activation function, the spatial attribute features of the output are mapped in a nonlinear manner, so that the state of each node is updated based on the 0th- (K-1)th neighbor node of the node, so that the influence of the spatial features in the distributed wind turbine output on each node is extracted. The output of the topology network graph convolution module is:

[0147] Z=ReLU(g θ’ * G x A )

[0148] In the formula, c represents the number of output channels of the graph convolution layer, T r represents the time step.

[0149] The LSTM model is used to fully extract the hidden time characteristics of the power of each wind turbine, and the Z output by the graph convolution network is reduced in dimension to as the input of the LSTM model.

[0150] The input Z' of the current moment t and the hidden state h of the previous moment t are sent into the input gate I t , the forget gate F t and the output gate O t . After that, the values of the input gate, the forget gate and the output gate are obtained by processing through the fully connected layers with the sigmoid activation function, and the candidate cell state record new input information is obtained by processing through the fully connected layer with the tanh activation function, then the information from is selected to be retained through the input gate, the information from the cell state C t-1 of the previous moment is selected to be retained through the forget gate, and the current moment cell state is updated to realize long-term memory transmission, which is shown as follows:

[0151] i t = σ (W i · [h t-1 , Z'] t + b i )

[0152] f t = σ (W f · [h t-1 , Z'] t + b f )

[0153] o t = σ (W o · [h t-1 , Z'] t + b o )

[0154]

[0155]

[0156] h t = o t *tanh (C t )

[0157] wherein W i , W f , W o are the weight matrices in the input gate, the forget gate and the output gate respectively; b i , b f , b o are the bias parameters, and W cis a weight matrix; b c is a bias parameter.

[0158] The LSTM layer outputs a feature M at time step t, and then M is transformed to M e R f×N×τ , where f is the number of channels, that is, M is a spatio-temporal feature tensor extracted from the basic data set.

[0159] The residual module is introduced to fuse the tensor M' output by the training set based on historical data similarity and the time and space feature extraction module into the tensor M output by the basic data set, fully excavating the hidden time features in the historical data, and effectively avoiding the degradation problem in the model learning process. The feature tensor M' output by the training set based on historical data similarity is input into the convolution layer with a 1x1 convolution kernel, and the time feature M is connected by jumping, and the operation is defined as follows:

[0160] H st = ReLU(M + (W residual *M' + b residual )) e R f×N×τ

[0161] where H st is the modified spatio-temporal feature tensor, the ReLU activation function can make the network better fit the nonlinear characteristics in the distributed wind power data, W residual is the convolution kernel parameter, and b residual is the corresponding bias.

[0162] The normalization (LayerNorm) layer normalizes the channel dimension of the feature information to stabilize the data distribution, and is represented as follows:

[0163] H ln = f ln (H st )

[0164] where f ln represents the normalization function.

[0165] Finally, the application adopts a prediction convolution layer with a time step T o , and a 1x f convolution kernel to output the distributed wind power prediction:

[0166]

[0167] where W pre is the convolution kernel parameter, and b pre is the bias.

[0168] In one possible implementation, the step of adopting a Bayesian optimization method to optimize the hyperparameters of the preset spatio-temporal graph neural network comprises:

[0169] Initial sample points are randomly generated within the set hyperparameter range and input into the Gaussian process to obtain the objective function loss value;

[0170] The Gaussian model is modified based on the objective function loss value until the objective function loss value meets the preset requirements, and the hyperparameter values ​​when the objective function loss value meets the preset requirements are obtained.

[0171] For example, Bayesian optimization is a data-driven optimization algorithm that mainly includes a Gaussian process and a sampling process. The Gaussian process is used to fit a high-dimensional function of input and output, while the sampling process determines the next most promising sampling point through a sampling function, and iterative sampling continuously optimizes the algorithm.

[0172] Specifically, Gaussian process:

[0173] A Gaussian process refers to a set of random variables, where the set is the set of different parameters in the model. Any linear combination of several samples has a joint Gaussian distribution, which can be expressed in the following form:

[0174] f(x)~GP(μ(x),k(x,x'))

[0175] In the formula, μ(x) is the mean function; k(x,x') is the covariance function representing x; μ(x) = E(f(x)), where E(f(x)) is the expected value of f(x), and the mean function is usually set to 0; f(x) represents the mean absolute error.

[0176] Let D be the historical information obtained from the model. 1:t ={x 1:t ,f 1:t}, where f t =f(x) t We need to find the next value x. t+1 The covariance matrix K is:

[0177]

[0178] According to the properties of Gaussian processes, f t and f t+1 All of them follow a joint Gaussian distribution. Assuming their mean is 0, the joint Gaussian distribution can be expressed as:

[0179]

[0180] Where k is:

[0181] k = [k(x t+1 ,x1)·k(x t+1 ,x2)L k(x t+1 ,x t )]

[0182] By the edge density function, we can get f t+1 The posterior probability of f

[0183]

[0184] In the formula, μ t (x t+1 ) and The calculation formula is as follows:

[0185] μ t (x t+1 ) = k T K -1 f 1:t

[0186]

[0187] After the above calculation, the normal distribution of x t+1 at any value can be estimated, and then the next more optimal sample point is located by using the sampling function.

[0188] Sampling function:

[0189] The Expected Improvement (EI) function is used as the sampling function, and the EI acquisition function is:

[0190]

[0191] In the formula, Φ(·) is the probability density function of the standard normal distribution, and φ(·) is the distribution function of the standard normal distribution, where Z can be represented as:

[0192]

[0193] Select hyperparameters:

[0194] First, randomly generate initial sample points within the set hyperparameter range, and input them into the Gaussian process. The loss value output by the model objective function is used to correct the Gaussian model, so that the Gaussian model is close to the true function distribution. Then the next set of sample points x i to be evaluated is selected by using the sampling function, and the model is trained to obtain the new output value y i of the objective function, so as to update the sample set D = {(x1, y1), (x2, y2), L, (x t , y t )} and the Gaussian model. Finally, if the loss value of the selected sample point corresponding to the objective function meets the requirements, the currently selected optimal hyperparameter is output, if it does not meet the requirements, the sample set D t is updated, and the Gaussian model is continuously corrected until the requirements are met.

[0195] In a possible implementation, as shown in Figure 3 , input feature selection and hyperparameter optimization are performed based on Bayesian theory, a sample set is constructed based on similarity, and wind power prediction is performed using a graph neural network. First, according to Bayesian theory, factors that have a greater impact on power are selected from the set of influencing factors to form an input vector, input features are screened, and Bayesian optimization is used to optimize hyperparameters. Then, data preprocessing is performed based on the screened input features, a basic data set that includes time features and space features is constructed, a weighted gray correlation method is used to find similar samples to the features of the point to be predicted in the historical sample set, and a weighted interleaved training set is constructed. Finally, a space-time graph neural network is constructed, the spatial features of each wind turbine are extracted through a graph convolution network, the node information is updated and then input into an LSTM model to extract time features, a residual layer is used to prevent network degradation and optimize training efficiency, and finally, the dispersed wind power prediction value is output through a prediction convolution layer.

[0196] In a possible implementation, as shown in Figure 4 , the application provides a dispersed wind turbine power prediction device, and the maintenance device comprises:

[0197] The selection module 201 is configured to select, according to Bayesian theory, a target factor that has the greatest impact on dispersed wind power from the set of influencing factors.

[0198] The construction module 202 is configured to determine a target feature based on the target factor, establish a basic data set including time features and space features according to the target feature, and select, from a historical sample set, a target sample having a similarity higher than a threshold value to the features of a point to be predicted, and construct a training set.

[0199] The prediction module 203 is configured to import the basic data set and the training set into a preset space-time graph neural network to obtain a dispersed wind power prediction value.

[0200] In a possible implementation, as shown in Figure 5 , the electronic device 300 provided by the application embodiment comprises a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and capable of running on the processor 320. When the processor 320 executes the computer program 311, the following steps are implemented: selecting, according to Bayesian theory, a target factor that has the greatest impact on dispersed wind power from the set of influencing factors; determining a target feature based on the target factor, establishing a basic data set including time features and space features according to the target feature, and selecting, from a historical sample set, a target sample having a similarity higher than a threshold value to the features of a point to be predicted, and constructing a training set; and importing the basic data set and the training set into a preset space-time graph neural network to obtain a dispersed wind power prediction value.

[0201] In a possible implementation, as shown in Figure 6 The embodiment of the present application provides a computer readable storage medium 400, which stores a computer program 411, and the computer program 411 is executed by a processor to realize the following steps: selecting a target factor which has the greatest influence on the distributed wind power from influence factors according to Bayesian theory; determining a target feature based on the target factor, establishing a basic data set including time features and space features according to the target feature, and screening a target sample with a feature similar to a to-be-predicted point feature from a historical sample set to construct a training set; and importing the basic data set and the training set into a preset spatio-temporal graph neural network to obtain a distributed wind power prediction value.

[0202] The computer readable signal medium can include a data signal transported, for example, in baseband or as a carrier wave by a propagated signal, such as those used in all or some of computers and computer networks. The propagated signal can take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. Computer readable signal medium can be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate or transport programming for use by or in connection with an instruction execution system, apparatus, or device.

[0203] Program code embodied on a computer readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wire line, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0204] Computer program code for carrying out operations for aspects of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). These implementations can provide all or a portion of the functions described herein.

[0205] Those skilled in the art should understand that the modules or steps of the present application described above can be realized by general computing devices, which can be centralized on a single computing device or distributed on a network composed of multiple computing devices, and optionally, they can be realized by program codes executable by the computing devices, so that they can be stored in storage devices and executed by the computing devices, or they can be respectively manufactured into individual integrated circuit modules, or multiple modules or steps among them can be manufactured into a single integrated circuit module to realize. Thus, the present application is not limited to any specific combination of hardware and software.

[0206] It is noted that the above merely describes the preferred embodiments of the present application and the principles of the applied technology. Those skilled in the art will understand that the present application is not limited to the specific embodiments described above, and various obvious changes, re-adjustments and substitutions can be made by those skilled in the art without departing from the scope of the present application. Therefore, although the present application has been described in detail through the above embodiments, the present application is not limited to the above embodiments, and more other equivalent embodiments can be included without departing from the concept of the present application, and the scope of the present application is determined by the appended claims.

[0207] The above only discloses the preferred embodiments of the present application, and of course cannot limit the scope of the rights of the present application, so equivalent changes made according to the claims of the present application are still within the scope of the present application.

Claims

1. A distributed wind turbine power prediction method, characterized in that, The method comprises the following steps: selecting a target factor having the greatest influence on the distributed wind power from the influence factor set according to Bayesian theory; determining a target feature based on the target factor, establishing a basic data set comprising time features and space features according to the target feature, and screening target samples having a similarity higher than a threshold to a feature of a to-be-predicted point from a historical sample set to construct a training set; and importing the basic data set and the training set into a preset spatio-temporal graph neural network to obtain a distributed wind power prediction value.

2. A decentralized wind turbine power prediction method according to claim 1, characterized in that, The step of selecting a target factor having the greatest influence on the distributed wind power from the influence factor set according to Bayesian theory comprises the following steps: selecting different factors influencing wind power to form the influence factor set; determining a prior probability of the influence factor based on historical power data by using a probability statistical method; randomly selecting influence factors from the influence factor set to form an input feature vector; calculating an effectiveness verification error by using an N-fold cross-validation algorithm to obtain a likelihood function of the input feature vector; calculating a posterior probability of the input feature vector by using a variational inference method based on the likelihood function; comparing the posterior probabilities of at least two groups of input feature vectors, and selecting an input feature vector having the greatest posterior probability as a target input feature vector; selecting a target factor having the greatest influence on the distributed wind power according to the target input feature vector.

3. A distributed wind turbine power prediction method as claimed in claim 1, characterized in that, Before the step of establishing a basic data set comprising time features and space features according to the target feature, and screening target samples having a similarity higher than a threshold to a feature of a to-be-predicted point from a historical sample set to construct a training set, the method further comprises the following steps: performing outlier removal on the target feature based on a local outlier factor algorithm to obtain an initial target feature; performing missing value reconstruction on the initial target feature according to a missing value filling method of a generative adversarial network to obtain an intermediate target feature; performing normalization processing on the intermediate target feature to obtain a final target feature.

4. A distributed wind turbine power prediction method as claimed in claim 1, characterized in that, The step of establishing a basic data set comprising time features and space features according to the target feature comprises the following steps: taking the target feature and wind turbine output historical data as time features of each wind turbine; representing relative relationships of spatial positions of each distributed wind turbine by using an undirected topological net to obtain space features of each wind turbine, the undirected topological net having an adjacency matrix with weights determined by distances between wind turbines; constructing the basic data set based on the time features and the space features.

5. A distributed wind turbine power prediction method as claimed in claim 1, characterized in that, The step of screening target samples having a similarity higher than a threshold to a feature of a to-be-predicted point from a historical sample set to construct a training set comprises the following steps: extracting a target sample set having a similarity higher than a threshold to a feature of a to-be-predicted point from the historical sample set by using a grey correlation algorithm based on an entropy weight method; arranging the target sample set according to an original time sequence, and sequentially sampling one sample from each target sample set without replacement to form a time sequence sample with a preset step length, and constructing an initial training set according to the time sequence sample; The relative relationship of the spatial positions of each distributed wind turbine is represented by an undirected topological net, and the spatial characteristics of each wind turbine are obtained, the undirected topological net has a weight adjacency matrix, and the weight value is determined by the distance between each turbine; The spatial characteristics of each wind turbine are added to the initial training set to obtain a target training set.

6. A distributed wind turbine power prediction method as claimed in claim 1, characterized in that, Before the step of importing the base data set and the training set into a preset spatio-temporal graph neural network to obtain a distributed wind power prediction value, the method further comprises: Optimizing hyperparameters of the preset spatio-temporal graph neural network by adopting a Bayesian optimization method, and selecting optimal calculation dimensions and hyperparameters; Based on the calculation dimensions and hyperparameters, the preset spatio-temporal graph neural network is constructed, which includes a graph convolution network for extracting spatial characteristics of each wind turbine, an LSTM model for extracting time characteristics of each wind turbine, a residual layer for preventing network degradation, and a prediction convolution layer for outputting a distributed wind power prediction value.

7. The distributed wind turbine power prediction method of claim 6, wherein, The step of optimizing hyperparameters of the preset spatio-temporal graph neural network by adopting the Bayesian optimization method comprises: Randomly generating initial sample points within a set hyperparameter range and inputting them into a Gaussian process to obtain a target function loss value; Based on the target function loss value, the Gaussian model is corrected until the target function loss value meets the preset requirement, and the hyperparameter value when the target function loss value meets the preset requirement is obtained.

8. A distributed wind turbine power prediction apparatus, characterized by, The prediction device comprises: A selection module configured to select a target factor having the greatest impact on the distributed wind power from the influence factor set according to the Bayesian theory; A construction module configured to determine a target feature based on the target factor, establish a base data set including time characteristics and spatial characteristics based on the target feature, and select target samples having a similarity higher than a threshold to a to-be-predicted point feature from a historical sample set to construct a training set; A prediction module configured to import the base data set and the training set into a preset spatio-temporal graph neural network to obtain a distributed wind power prediction value.

9. An electronic device, comprising: It comprises: A processor, a memory, and a bus, the memory stores machine-readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to perform the steps of the distributed wind turbine power prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the computer program is executed by the processor to perform the steps of the distributed wind turbine power prediction method according to any one of claims 1 to 7.

Images