Wind power generation capacity prediction method based on adaptive graph convolutional neural network and related device

Abstract

The invention discloses a wind power generation capacity prediction method based on an adaptive graph convolutional neural network and a related device, and the method comprises the steps: employing a default value processing technology for the wind power generation time series data to complement the missing data, carrying out the alignment of the data at different sampling times, and carrying out the standardization of each type of data; performing feature coding on the preprocessed wind power generation time sequence data, coding time information of the time sequence data and mapping the time information to a low-dimensional space; carrying out semantic graph convolution learning on the time features to obtain semantic similarity and the time features; carrying out convolutional learning on the spatial features of the geospatial map to obtain semantic similarity and time features; and fusing the semantic similarity, the time features, the geographic space dependence and the spatial-temporal features, and decoding the fused features into prediction of the power generation capacity of each wind driven generator in the future. According to the method, the mutual influence between the wind power generators is learned through the adaptive graph convolutional network, and the prediction precision and generalization ability of the wind power generation power can be improved.

Description

Technical Field

[0001] The present invention relates to the field of wind power generation, and specifically to a wind power generation capacity prediction method and related device based on an adaptive graph convolutional neural network. Background Art

[0002] With the advancement of the energy development plan, the development of clean energy has become an important pillar of the development of China's energy industry. As a widely existing clean energy, wind power generation has advantages such as cleanness and renewability. However, due to the significant characteristics of intermittency, randomness, and volatility of wind, its power generation has large fluctuations, which puts higher requirements on the stability and reliability of the power system. To address these challenges, how to effectively predict the wind power generation capacity has become an important issue faced by the wind power industry.

[0003] Currently, wind power generation prediction methods are mainly divided into the following types:

[0004] (1) Methods based on physical models. Such methods mainly predict the wind power generation output by simulating the physical characteristics of the wind power generation system and the influence of meteorological factors. Based on the power curve determined during the manufacture of the wind power generation system, a model of the output power of the wind turbine with meteorological variables such as wind speed, air pressure, and temperature and humidity is established. However, in the real scenario, the meteorological conditions are complex and changeable, and the prediction accuracy of the method based on the physical model is greatly affected by the weather prediction model.

[0005] (2) Methods based on statistical models. Such methods analyze the relationship between wind power generation and related meteorological variables in historical data and establish a regression model, usually using autoregressive and moving average theories to establish a statistical model. This method is relatively simple to calculate and requires a large amount of historical data, but the prediction accuracy of the model is greatly affected by the data stationarity.

[0006] (3) Machine learning models. Such methods automatically learn the complex relationship between wind power generation and meteorological data in a data-driven manner and are commonly used to handle the non-linear, time-series, and high-dimensional feature problems of wind power generation. Commonly used machine learning methods mainly include long short-term memory networks (LSTM), convolutional neural networks (CNN), and their variants. The machine learning method can effectively process non-linear data, but the computational complexity of the model training process is relatively high, and the interpretability of the model is relatively low.

[0007] Wind power generation systems are usually deployed relatively concentratedly in one place, and there are mutual influences among different wind turbines. For example, environmental factors such as wind speed and wind direction will generate spatial correlations among wind turbines at different locations. However, existing wind power generation prediction methods mostly focus on the independent prediction of a single wind turbine, ignoring the spatial correlation and spatio-temporal variation law among wind turbines. This makes the traditional prediction model unable to fully utilize the interaction among multiple wind turbines in the wind power system, resulting in low prediction accuracy. Summary of the Invention

[0008] The present invention provides a wind power generation capacity prediction method and related device based on an adaptive graph convolutional neural network, which predicts the power generation of multiple wind turbines in a region, and learns the mutual influence among wind turbines through an adaptive graph convolutional network, so as to improve the prediction accuracy and generalization ability of wind power generation power.

[0009] A wind power generation capacity prediction method based on an adaptive graph convolutional neural network includes the following steps:

[0010] Step S1: Data preprocessing: Use the missing value processing technology to supplement the missing data in the wind power generation time series data, align the data with different sampling times, and standardize each type of data;

[0011] Step S2: Feature encoding: Perform feature encoding on the wind power generation time series data preprocessed in step S1, and at the same time encode the time information of the time series data and map it to a low-dimensional space;

[0012] Step S3: Semantic graph convolution to learn time features: Use an adaptive semantic graph convolutional recurrent network to learn the semantic similarity and time features among the power generations of each wind turbine in the data processed in step S2;

[0013] Step S4: Geospatial graph convolution to learn spatial features: Use an adaptive geospatial graph convolutional network to learn the geospatial dependence and spatio-temporal features of each wind turbine in the original wind power generation time series data;

[0014] Step S5: Feature decoding prediction: Integrate the semantic similarity and time features learned in step S3 and the geospatial dependence and spatio-temporal features learned in step S4, and decode the integrated features into the prediction of the future power generation capacity of each wind turbine.

[0015] Further, the data preprocessing in step S1 includes the following steps:

[0016] S1-1: Data cleaning: The missing values in the original wind power time series data are processed using the default value handling method, and the missing values are assigned the mean value within the interval of the field at the missing position using the interval mean method; for the data that only exists in some wind turbines, according to its physical meaning and the impact on the system, if it has a greater impact, the data of other wind turbines is processed using default values, and the mode, median or mean value is assigned according to the field meaning; if the impact is small, the data of this field is discarded.

[0017] S1-2: Data alignment: The data of the units that do not meet the sampling frequency is adjusted according to the sampling frequency used by most generators in the data, and according to the physical meaning of each field of the data, the data of all generators is adjusted to the same sampling frequency by summing, taking the mean and averaging.

[0018] S1-3: Data standardization: For continuous data, zero-mean standardization is used; for categorical data, label encoding is used for processing.

[0019] Furthermore, the feature encoding in step S2 includes the following steps:

[0020] S2-1: Wind power data feature encoding: The wind power data obtained after preprocessing in step S1 is expressed as where N represents the number of wind turbines, T represents the number of time steps in the time series, C represents the length of the observation vector of the wind turbine, and the feature encoder uses stacked multi-layer residual blocks to encode the wind power data, and uses F to represent the encoded wind power data features.

[0021] S2-2: Time data feature encoding: The time information is expressed as where T represents the number of time steps in the time series, d represents the length of the time data vector, and the time information is feature-encoded to enhance the model's perception ability of time information.

[0022] Furthermore, the step S3 semantic graph convolution to learn time features includes the following steps:

[0023] S3-1: Construct a semantic graph structure: Set a node embedding for each wind turbine and form a global node embedding matrix where N is the number of wind turbines, D is the dimension of the node embedding, and an adaptive semantic graph structure is constructed through the node embedding matrix E, and the formula is as follows:

[0024]

[0025] S3-2: Perform graph convolution operations on the semantic graph structure, take the encoded wind power time series data features F obtained in step S2 as the input, and perform graph convolution in the following way:

[0026]

[0027] Among them, I N represents the identity matrix, E represents the node embedding matrix, W represents the weight matrix parameter, b represents the bias parameter, and Z represents the extracted feature, which is used as the input for the S3-3 gated recurrent unit GRU;

[0028] S3-3: Use the gated recurrent unit GRU to obtain the semantic similarity and temporal features among the power generation powers of each wind turbine. The calculation method is as follows:

[0029]

[0030] Among them, t represents the t-th GRU unit, z t is the output of the update gate, r t is the output of the reset gate, and the output of this GRU unit is h t , σ represents the sigmoid() activation function, F represents encoding the features of the wind power generation time series data, E represents the node embedding matrix, W z , b z are the parameters of the update gate, W r , b r are the parameters of the reset gate, W h , b h are the parameters of the output gate, I N represents the identity matrix, and the h t output by the last layer of GRU units is denoted as which represents the semantic similarity and temporal features obtained by the adaptive semantic graph convolutional recurrent network.

[0031] Furthermore, the step S4 of geospatial graph convolution to learn spatial features includes the following steps:

[0032] S4-1: Construct a geographical graph structure: According to the geographical locations of the wind turbines, calculate the Euclidean distance between any two wind turbines to obtain the geographical distance matrix G dis , and the closer the distance between wind turbines, the greater their mutual influence, which is reflected as an edge in the graph. The following formula is used to calculate whether the influence generated by the distance between any two wind turbines exceeds the threshold ∈:

[0033]

[0034] Among them, i and j represent any two wind turbines, represents the distance between wind turbines i and j, and σ is the standard deviation of the distance. Through A dis construct the spatial graph adjacency matrix

[0035]

[0036] S4-2: Perform temporal convolution: Use gated dilated convolution to learn the temporal feature T on the encoded features of each wind turbine obtained in S2 out , expressed as:

[0037] T out = tanh(TC1) ⊙ sigmoid(TC2)

[0038] where both TC1 and TC2 are the results of dilated convolution, expressed as:

[0039] TC i = cocat(K 1×2 * F, K 1×3 * F, K 1×6 * F, K 1×7 * F)

[0040] where * represents the convolution operation, K 1×2 represents a 1×2 convolution kernel, and concat() represents the concatenation operation of tensors;

[0041] S4-3: Perform spatial graph convolution: Perform spatial graph convolution on the temporal feature T obtained in S4-2 out to capture the spatial feature H (k) , expressed as:

[0042]

[0043] where β is a hyperparameter used to control the proportion of input features, and H (0) = T out , and the outputs of multiple layers of spatial convolution are fused to obtain the geospatial dependence and spatio-temporal features captured by the adaptive geospatial graph convolution network

[0044]

[0045] Furthermore, the feature decoding prediction in step S5 includes the following steps:

[0046] S5-1: Spatio-temporal feature fusion: Fuse the semantic similarity and temporal features obtained in step S3 and the geospatial dependence and spatio-temporal features obtained in S4 using a multi-layer perceptron to fuse the spatio-temporal features, expressed as:

[0047]

[0048] where L represents the number of spatial graph convolution modules used, represents the output of the k-th spatial graph convolution module;

[0049] S5-2: Decode spatio-temporal features: The decoder constructed using multiple layers of residual blocks decodes the fused spatio-temporal features to obtain the predicted values of the power generation capacity of future wind turbines. Where H represents the predicted number of future time steps, and C target represents the number of variables predicted for each wind turbine.

[0050] A wind power generation capacity prediction device based on an adaptive graph convolutional neural network, comprising:

[0051] A data preprocessing module, which is used to complete the missing data of the wind power generation time series data by using the default value processing technology, align the data with different sampling times, and standardize each type of data;

[0052] A feature encoding module, which is used to encode the features of the preprocessed wind power generation time series data, and at the same time encode the time information of the time series data and map it to a low-dimensional space;

[0053] A semantic graph convolutional learning time feature module, which is used to use an adaptive semantic graph convolutional recurrent network to learn the semantic similarity and time features between the power generation powers of each wind turbine in the data processed by the feature encoding module;

[0054] A geospatial graph convolutional learning spatial feature module, which is used to use an adaptive geospatial graph convolutional network to learn the geospatial dependence and spatio-temporal features of each wind turbine in the original wind power generation time series data;

[0055] A feature decoding prediction module, which is used to fuse the semantic similarity and time features and the geospatial dependence and spatio-temporal features, and decode the fused features into the prediction of the power generation capacity of each future wind turbine.

[0056] A wind power generation capacity prediction system based on an adaptive graph convolutional neural network, comprising: a computer-readable storage medium and a processor;

[0057] The computer-readable storage medium is used to store executable instructions;

[0058] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the wind power generation capacity prediction method based on the adaptive graph convolutional neural network:

[0059] A non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the wind power generation capacity prediction method based on the adaptive graph convolutional neural network.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] (1) Accurate prediction of the power generation capacity of wind turbines: The present invention models the wind power generation system as a graph structure, fully considering the spatio-temporal characteristics of the wind power generation system in the region, and realizes high-precision prediction of wind power generation power. Compared with the prior art, the present invention has significant advantages in prediction accuracy and stability.

[0062] (2) Modeling the mutual influence between different wind turbines: The present invention models multiple wind turbines as a topological graph, and uses an adaptive graph convolutional neural network to learn the spatio-temporal characteristics in the system and model the mutual influence between different wind turbines. Compared with existing methods, the present invention models the spatial correlation and spatio-temporal variation law between wind turbines, and has stronger generalization ability between different wind turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flowchart of a method for predicting wind power generation capacity based on an adaptive graph convolutional neural network according to an embodiment of the present invention.

[0064] Figure 2 is a schematic diagram of encoded features according to an embodiment of the present invention.

[0065] Figure 3 is a schematic diagram of the structure of an adaptive semantic graph convolutional recurrent network according to an embodiment of the present invention.

[0066] Figure 4 is a schematic diagram of the structure of an adaptive geospatial graph convolutional network according to an embodiment of the present invention.

[0067] Figure 5 is a schematic diagram of the structure of a feature decoding module according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. 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 protection scope of the present invention.

[0069] Please refer to Figure 1 , the first aspect of the present invention provides a method for predicting wind power generation capacity based on an adaptive graph convolutional neural network, which is used for predicting the power generation power of a wind power generation system. The method includes the following steps:

[0070] S1: Data preprocessing: The missing data in the wind power generation time series data is supplemented by using the default value processing technology. For data with different sampling times, relevant methods are used for alignment, and each type of data is standardized to reduce the impact of non-standard data on the model training process.

[0071] S2: Feature encoding: The wind power generation time series data preprocessed in step S1 is feature-encoded. At the same time, the time information of the time series data is encoded and mapped to a low-dimensional space.

[0072] S3: Semantic graph convolution to learn time features: An adaptive semantic graph convolution recurrent network is used to learn the semantic similarity and time features among the power generation powers of each wind turbine in the data processed in step S2.

[0073] S4: Geospatial graph convolution to learn spatial features: An adaptive graph convolution network is used to learn the geospatial dependence and spatio-temporal features of each wind turbine in the original wind power generation time series data.

[0074] S5: Feature decoding prediction: The semantic similarity and time features learned in step S3 and the geospatial dependence and spatio-temporal features learned in step S4 are fused, and the fused features are decoded into predictions of the future power generation capabilities of each wind turbine.

[0075] Furthermore, the data preprocessing in step S1 specifically includes the following steps:

[0076] S1-1: Data cleaning. The missing values in the original wind power generation time series data are processed by using the default value processing method, and the missing values are assigned the mean value within the interval of the field at the missing position by using the interval mean method. For data that only exists in some wind turbines, according to its physical meaning and the impact on the system, if it has a greater impact, default processing is used for the data of other wind turbines, and the mode, median or mean value is assigned according to the field meaning; if the impact is small, the data of this field is considered to be discarded.

[0077] S1-2: Data alignment. Due to differences in data acquisition systems, the time intervals of the time series data sampled by different wind turbines may vary. The data of the units that do not meet this sampling frequency is adjusted according to the sampling frequency used by most generators in the data. According to the physical meaning of each field in the data, methods such as summation, taking the mean value, and equal division are used to adjust the data of all generators to the same sampling frequency.

[0078] S1-3: Data standardization. To eliminate the differences in dimension and numerical range between different features and improve the convergence speed and prediction performance of the model, the original data is standardized. For continuous data, zero-mean standardization is used; for categorical data, label encoding is used for processing.

[0079] Furthermore, for the feature encoding in step S2, the encoder structure is as Figure 2 shown. It specifically includes the following steps:

[0080] S2-1: Wind power data feature encoding. The wind power generation data obtained after preprocessing in step S1 can be expressed as where N represents the number of wind turbines, T represents the number of time steps in the time series, and C represents the length of the observation vector of the wind turbine. The feature encoder uses stacked multi-layer residual blocks to encode the wind power generation data, and uses F to represent the feature of the encoded wind power data. The main data fields used are shown in Table 1.

[0081] Table 1 Main fields of wind power generation data

[0082]

[0083] S2-2: Time data feature encoding. The time of the time series data has a certain impact on the wind power generation capacity, which is obvious in aspects such as daily, monthly, seasonal, and weather cycle changes. The time information can be expressed as where T represents the number of time steps in the time series, and d represents the length of the time data vector. To fully explore the impact of time features on the prediction results, the time information is encoded to enhance the model's perception ability of time information and further improve the accuracy of wind power generation capacity prediction. The main time data feature fields used are shown in Table 2.

[0084] Table 2 Time data feature fields

[0085]

[0086]

[0087] Furthermore, in step S3, the semantic graph convolution learns time features, and the adopted model structure is as Figure 3 shown. It specifically includes the following steps:

[0088] S3-1: Construct a semantic graph structure. In the present invention, a node embedding is set for each wind turbine, and a global node embedding matrix is formed, where N is the number of wind turbines and D is the dimension of the node embedding. An adaptive semantic graph structure is constructed through the node embedding matrix E, and the formula is as follows:

[0089]

[0090] S3-2: Perform graph convolution operations on the semantic graph structure, and its model structure is Figure 3The graph convolutional structure in it. Taking the encoded wind power generation time series data feature F obtained in S2 as the input, perform graph convolution in the following manner:

[0091]

[0092] Among them, I N represents the identity matrix, E represents the node embedding matrix, W represents the weight matrix parameter, b represents the bias parameter, and Z represents the extracted feature, which is used as the input for the GRU in S3-3.

[0093] S3-3: Use the gated recurrent unit (GRU) to obtain the semantic similarity and temporal features among the power generation powers of each wind turbine. The calculation method is as follows:

[0094]

[0095] Among them, t represents the t-th GRU unit, z t is the update gate output, r t is the reset gate output, and the output of this GRU unit is h t , σ represents the sigmoid() activation function, F represents the encoded wind power generation time series data feature, E represents the node embedding matrix, W z , b z are the parameters of the update gate, W r , b r are the parameters of the reset gate, W h , b h are the parameters of the output gate, I N represents the identity matrix. The h t output by the last layer of GRU units is denoted as which represents the semantic similarity and temporal features obtained by the adaptive semantic graph convolutional recurrent network.

[0096] Furthermore, in the step S4, the geospatial graph convolution learns spatial features, and the model used is as Figure 4 shown. Specifically, it includes the following steps:

[0097] S4-1: Construct a geographical graph structure. According to the geographical locations of the wind turbines, calculate the Euclidean distance between any two wind turbines to obtain the geographical distance matrix G dis , and the closer the distance between the wind turbines, the greater their mutual influence, which is reflected as an edge in the graph. The following formula can be used to calculate whether the influence generated by the distance between any two wind turbines exceeds the threshold ∈:

[0098]

[0099] Among them, i and j represent any two wind turbines, denote the distance between wind turbines i and j, and σ is the standard deviation of the distance. Through A dis Construct the spatial graph adjacency matrix

[0100]

[0101] S4-2: Perform temporal convolution. Use gated dilated convolution on the encoded features of each wind turbine obtained in S2 to learn the temporal feature T out , and its model structure is as shown in Figure 4 the gated dilated convolution part in, which can be expressed as:

[0102] T out = tanh(TC1) ⊙ sigmoid(TC2)

[0103] where both TC1 and TC2 are the results of dilated convolution, which can be expressed as:

[0104] TC i = cocat(K 1×2 *F, K 1×3 *F, K 1×6 *F, K 1×7 *F)

[0105] where * represents the convolution operation, K 1×2 represents a 1×2 convolution kernel, and concat() represents the concatenation operation of tensors.

[0106] S4-3: Perform spatial graph convolution. Perform spatial graph convolution on the basis of the temporal feature T out obtained in S4-2 to capture the spatial feature H (k) , and its model structure is as shown in Figure 4 the spatial graph convolution part in, which can be expressed as:

[0107]

[0108] where β is a hyperparameter used to control the proportion of input features, and H (0) = T out . Fuse the outputs of multiple layers of spatial convolution to obtain the geospatial dependencies and spatio-temporal features captured by the adaptive geospatial graph convolution network

[0109]

[0110] Furthermore, for the step S5, feature decoding and prediction, the model structure is as shown in Figure 5 shown. Specifically, it includes the following steps:

[0111] S5-1: Spatio-temporal feature fusion. Combine the semantic similarity and temporal features obtained in S3 The geospatial dependence and spatio-temporal characteristics obtained from S4 Using a multi-layer perceptron to fuse spatio-temporal features can be expressed as:

[0112]

[0113] where L represents the number of spatial graph convolution modules used, represents the output of the k-th spatial graph convolution module.

[0114] S5-2: Decoding spatio-temporal features. A decoder constructed using multi-layer residual blocks decodes the fused spatio-temporal features to obtain the predicted value of the future power generation capacity of the wind turbines where H represents the number of future time steps predicted, and C target represents the number of variables predicted for each wind turbine.

[0115] The embodiments of the present invention were verified on a domestic wind power prediction dataset. The performance of the wind power generation capacity prediction algorithm based on the adaptive graph convolutional neural network proposed by this method is significantly better than traditional algorithms, and the results are shown in Table 3. In the table, MSE (Mean Square Error) represents the mean square error between the predicted value and the true value of the wind power generation capacity of the algorithm, and MAE (Mean Absolute Error) represents the mean absolute error between the predicted value and the true value of the wind power generation capacity of the algorithm.

[0116] Table 3 Comparison of the performance of wind power generation prediction algorithms

[0117] Algorithm The method of the present invention LSTM MTGNN SegRNN MSE 0.152 0.183 0.176 0.178 MSE improvement / 16.9％ 13.6％ 14.6％ MAE 0.184 0.224 0.209 0.212 MAE improvement / 17.9％ 12.0％ 13.2％

[0118] The second aspect of the present invention also provides a wind power generation capacity prediction device based on an adaptive graph convolutional neural network, including:

[0119] A data preprocessing module for complementing the missing data in the wind power generation time series data using the default value processing technology, aligning the data with different sampling times, and standardizing each type of data;

[0120] A feature encoding module for encoding the features of the preprocessed wind power generation time series data, and at the same time encoding the time information of the time series data and mapping it to a low-dimensional space;

[0121] A semantic graph convolution learning time feature module for using an adaptive semantic graph convolutional recurrent network to learn the semantic similarity and time features between the power generation powers of each wind turbine in the data processed by the feature encoding module;

[0122] The geospatial graph convolution learning spatial feature module is used to learn the geospatial dependence and spatio-temporal features of each wind turbine in the original wind power generation time series data by using an adaptive geospatial graph convolution network;

[0123] The feature decoding and prediction module is used to fuse semantic similarity, temporal features, geospatial dependence and spatio-temporal features, and decode the fused features into predictions of the power generation capabilities of future wind turbines.

[0124] On the other hand, the present invention provides a wind power generation capacity prediction system based on an adaptive graph convolution neural network, including: a computer-readable storage medium and a processor;

[0125] The computer-readable storage medium is used to store executable instructions;

[0126] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the wind power generation capacity prediction method based on the adaptive graph convolution neural network described in the first aspect.

[0127] On the other hand, the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the wind power generation capacity prediction method based on the adaptive graph convolution neural network described in the first aspect.

[0128] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the specified functions in one Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0130] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction device that implements the functions specified in one or more of the flow Figure 1 steps or a plurality of steps and / or boxes Figure 1 specified in one or more of the boxes or a plurality of boxes.

[0131] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the flow Figure 1 steps or a plurality of steps and / or boxes Figure 1 specified in one or more of the boxes or a plurality of boxes.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific embodiments of the present invention, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention..

Claims

1. A wind power generation capacity prediction method based on an adaptive graph convolutional neural network, characterized in that It includes the following steps: Step S1: Data preprocessing: Use the default value processing technology to complete the missing data in the wind power generation time series data, align the data with different sampling times, and standardize each type of data; Step S2: Feature encoding: Encode the wind power generation time series data preprocessed in Step S1, and at the same time encode the time information of the time series data and map it to a low-dimensional space; Step S3: Semantic graph convolution to learn time features: Use the adaptive semantic graph convolution recurrent network to learn the semantic similarity and time features between the power generation powers of each wind turbine in the data processed in Step S2; Step S4: Geospatial graph convolution to learn spatial features: Use the adaptive geospatial graph convolution network to learn the geospatial dependence and spatio-temporal features of each wind turbine in the original wind power generation time series data; Step S5: Feature decoding prediction: Integrate the semantic similarity and time features learned in Step S3 and the geospatial dependence and spatio-temporal features learned in Step S4, and decode the integrated features into predictions of the future power generation capabilities of each wind turbine.

2. The wind power generation capacity prediction method based on the adaptive graph convolutional neural network according to claim 1, wherein: The data preprocessing in Step S1 includes the following steps: S1-1: Data cleaning: Use the default value processing method to process the missing values in the original wind power generation time series data, and use the interval mean method to assign the mean value of the field within the interval at the missing position to the missing value; for the data that only exists in some wind turbines, according to its physical meaning and the impact on the system, if it has a greater impact, use the default processing for the data of other wind turbines, and assign the mode, median or mean according to the field meaning; if the impact is small, discard the field data; S1-2: Data alignment: Adjust the data of the units that do not meet the sampling frequency according to the sampling frequency used by most generators in the data, and adjust all generator data to the same sampling frequency by summing, taking the mean and averaging according to the physical meaning of each field of the data; S1-3: Data standardization: For continuous data, use zero-mean standardization; for categorical data, use the label encoding method for processing.

3. The wind power generation capacity prediction method based on an adaptive graph convolutional neural network according to claim 1, characterized in that: The feature encoding in Step S2 includes the following steps: S2-1: Wind power data feature encoding: The wind power generation data obtained after preprocessing in step S1 is expressed as where N represents the number of wind turbines, T represents the number of time steps in the time series, C represents the length of the observation vector of the wind turbine, and the feature encoder uses stacked multi-layer residual blocks to encode the wind power generation data, and F is used to represent the wind power data features after encoding; S2-2: Temporal Data Feature Encoding: The temporal information is represented as where T represents the number of time steps in the time series, d represents the length of the temporal data vector, and the temporal information is feature-encoded to enhance the model's perception ability of temporal information.

4. The wind power generation capacity prediction method based on the adaptive graph convolutional neural network according to claim 3, wherein: The semantic graph convolution in Step S3 to learn time features includes the following steps: S3-1: Construct a semantic graph structure: Set a node embedding for each wind turbine and form a global node embedding matrix where N is the number of wind turbines and D is the dimension of the node embedding. An adaptive semantic graph structure is constructed through the node embedding matrix E, and the formula is as follows: S3-2: Perform graph convolution operations on the semantic graph structure, take the encoded wind power generation time series data feature F obtained in Step S2 as the input, and perform graph convolution in the following way: Among them, I N represents the identity matrix, E represents the node embedding matrix, W represents the weight matrix parameter, b represents the bias parameter, and Z represents the extracted feature, which is used as the input for the S3-3 gated recurrent unit GRU; S3-3: Use the gated recurrent unit GRU to obtain the semantic similarity and time features between the power generation powers of each wind turbine, and the calculation method is as follows: Among them, t represents the t-th GRU unit, and z t is the output of the update gate, and r t is the output of the reset gate. The output of this GRU unit is h t , σ represents the sigmoid() activation function, F represents encoding the features of the wind power generation time series data, E represents the node embedding matrix, W z , b z are the parameters of the update gate, W r , b r are the parameters of the reset gate, W h , b h are the parameters of the output gate, I N represents the identity matrix. The h t output by the last GRU unit is denoted as which represents the semantic similarity and temporal features obtained by the adaptive semantic graph convolutional recurrent network.

5. The wind power generation capacity prediction method based on the adaptive graph convolutional neural network according to claim 1, characterized in that: The geospatial graph convolution in Step S4 to learn spatial features includes the following steps: S4-1: Construct the geographical map structure: According to the geographical locations of the wind turbines, calculate the Euclidean distance between any two wind turbines to obtain the geographical distance matrix G dis , the closer the distance between wind turbines, the greater their mutual influence, which is represented as an edge in the graph. Use the following formula to calculate whether the influence generated by the distance between any two wind turbines exceeds the threshold ∈: where i and j represent any two wind turbines represents the distance between wind turbines i and j, and σ is the standard deviation of the distance. Through A dis construct a spatial graph adjacency matrix S4-2: Perform temporal convolution: Use gated dilated convolution to learn the temporal feature T on the encoded features of each wind turbine obtained in S2, expressed as: out , denoted as: T out = tanh(TC1) ⊙ sigmoid(TC2) Where both TC1 and TC2 are the results of dilated convolution, expressed as: TC i = concat(K 1×2 * F, K 1×3 * F, K 1×6 * F, K 1×7 * F) where * represents the convolution operation, and K 1×2 represents a 1×2 convolution kernel, and concat() represents the concatenation operation of tensors; S4-3: Perform spatial graph convolution: Perform spatial graph convolution on the temporal feature T obtained in S4-2 out to capture the spatial feature H (k) , expressed as: where β is a hyperparameter used to control the proportion of input features, H (0) = T out , and the outputs of the multi-layer spatial convolutions are fused to obtain the geospatial dependencies and spatio-temporal features captured by the adaptive geospatial graph convolutional network 6. The wind power generation capacity prediction method based on the adaptive graph convolutional neural network according to claim 1, characterized in that: The feature decoding prediction in Step S5 includes the following steps: S5-1: Spatiotemporal Feature Fusion: Fuse the semantic similarity and temporal features obtained in step S3 and the geospatial dependence and spatiotemporal features obtained in S4 to fuse the spatiotemporal features using a multi-layer perceptron, expressed as: where L represents the number of spatial graph convolutional modules used, represents the output of the k-th spatial graph convolutional module; S5-2: Decoding Spatiotemporal Features: The decoder constructed using multiple layers of residual blocks decodes the fused spatiotemporal features to obtain the predicted values of the future power generation capacity of the wind turbines. where H represents the predicted number of future time steps, and C target represents the number of variables predicted for each wind turbine.

7. A wind power generation capacity prediction device based on an adaptive graph convolutional neural network, characterized in that It includes: A data preprocessing module for using the default value processing technology to complete the missing data in the wind power generation time series data, aligning the data with different sampling times, and standardizing each type of data; A feature encoding module, configured to perform feature encoding on the preprocessed wind power generation time series data, and at the same time encode the time information of the time series data and map it to a low-dimensional space; A semantic graph convolutional learning time feature module, configured to use an adaptive semantic graph convolutional recurrent network to learn the semantic similarity and time features among the power generation powers of each wind turbine in the data processed by the feature encoding module; A geospatial graph convolutional learning spatial feature module, configured to use an adaptive geospatial graph convolutional network to learn the geospatial dependence and spatio-temporal features of each wind turbine in the original wind power generation time series data; A feature decoding prediction module, configured to fuse the semantic similarity and time features and the geospatial dependence and spatio-temporal features, and decode the fused features into predictions of the future power generation capabilities of each wind turbine.

8. A wind power generation capacity prediction system based on an adaptive graph convolutional neural network, comprising: A computer-readable storage medium and a processor; The computer-readable storage medium is configured to store executable instructions; The processor is configured to read the executable instructions stored in the computer-readable storage medium and execute the wind power generation capacity prediction method based on an adaptive graph convolutional neural network according to any one of claims 1-6.

9. A non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the wind power generation capacity prediction method based on an adaptive graph convolutional neural network according to any one of claims 1-6 is implemented.

Images