Single-machine wind power ultra-short-term prediction method and device based on WNet

Abstract

The invention provides a single-machine wind power ultra-short-term prediction method and device based on WNet, and relates to the technical field of wind power generation prediction.The method comprises the steps that preprocessed sample data of a target unit is obtained, and the sample data comprises historical power data, historical meteorological data and numerical weather forecast data; constructing a double-branch input structure based on the preprocessed sample data; constructing a WNet model suitable for wind power prediction based on the double-branch input structure, and obtaining a historical data feature vector corresponding to the first branch input structure and a numerical weather forecast feature vector corresponding to the second branch input structure output by the WNet model; splicing the historical data feature vector and the numerical weather forecast feature vector to obtain a combined feature; and performing deep fusion and transformation on the combined features through a multi-layer perceptron network, and outputting a wind power prediction result. Through the prediction method provided by the invention, accurate prediction of the wind power can be realized.

Description

Technical Field

[0001] The present application relates to the technical field of wind power generation prediction, and in particular to a method and device for ultra-short-term prediction of single-machine wind power based on WNet. Background Art

[0002] With the acceleration of global urbanization and the continued growth of energy demand, the overuse of traditional fossil fuels has led to serious environmental problems. Against this backdrop, wind energy, as a zero-emission, renewable, clean energy source, has garnered widespread attention worldwide. Wind energy not only reduces dependence on fossil fuels, reduces air pollution, and improves air quality, but also enhances energy supply security and reduces reliance on imported energy, thereby promoting sustainable economic and social development.

[0003] The accuracy of wind power forecasts is crucial to the scheduling and operation of power systems. Accurate forecasts can help power system operators better plan power generation, optimize resource allocation, and reduce reserve capacity requirements, thereby improving the economic efficiency and reliability of the power system. In power systems with a high proportion of renewable energy, the volatility of wind power has a more significant impact on system stability. Therefore, improving the accuracy of wind power forecasts is particularly important. Summary of the Invention

[0004] In view of this, the present application proposes a single-machine wind power ultra-short-term prediction method and device based on WNet.

[0005] In a first aspect, the present application provides a single-machine wind power ultra-short-term prediction method based on WNet, comprising: Obtaining pre-processed sample data of the target unit, the sample data including historical power data, historical meteorological data, and numerical weather forecast data, the pre-processing including elimination of abnormal data and standardization to eliminate dimensional differences; Constructing a dual-branch input structure based on the preprocessed sample data, wherein the first branch input structure combines the preprocessed historical power data and the historical meteorological data into a joint input matrix, and the second branch input structure inputs the preprocessed numerical weather forecast data; Building a WNet model suitable for wind power forecasting based on the dual-branch input structure, and obtaining a feature processing result output by the WNet model, wherein the feature processing result includes a historical data feature vector corresponding to the first branch input structure and a numerical weather forecast feature vector corresponding to the second branch input structure; splicing the historical data feature vector and the numerical weather forecast feature vector in the feature dimension to obtain a combined feature; The combined features are deeply fused and transformed through a multi-layer perceptron network to output wind power prediction results.

[0006] In one embodiment, the method further comprises: The Huber loss function is used as the optimization target, and the Adam optimizer is used to update the parameters of the multilayer perceptron network.

[0007] In one embodiment, obtaining the pre-processed sample data of the target unit includes: Obtain sample data within a predetermined time series length; Incomplete daily data with consecutive missing data exceeding the preset time threshold are removed as abnormal data, and data with consecutive missing data within the preset time threshold are filled using linear interpolation to obtain sample data with uniform resolution; The sample data of uniform resolution is standardized to obtain the pre-processed sample data.

[0008] In one embodiment, constructing a WNet model suitable for wind power prediction based on the dual-branch input structure and obtaining a feature processing result output by the WNet model include: Adjusting the dual-branch input structure to adjust the height and width of the input data to be an even number and not less than a preset value; Extracting historical data features through a first branch input structure to obtain the historical data feature vector; wherein the first branch input structure is an encoder-decoder structure, the upsampling operation of each layer of the decoder corresponds to the downsampling operation in the encoder, and the first branch input structure combines the multi-scale features extracted by the encoder with the upsampling results of the decoder through a skip connection; Numerical weather forecast features are extracted through the second branch input structure to obtain the numerical weather forecast feature vector. The second branch input structure is an encoder-decoder structure. The second branch input structure combines the multi-scale features extracted by the encoder with the upsampling results of the decoder through a skip connection.

[0009] In one embodiment, adjusting the dual-branch input structure to adjust the height and width of the input data to be an even number and not less than a preset value includes: A padding operation is performed on the input data of the dual-branch input structure to adjust the height and width of the input data to an even number and not less than 8, wherein the padding operation is performed using zeros.

[0010] In one embodiment, extracting historical data features through the first branch input structure to obtain the historical data feature vector includes: The encoder of the first branch input structure performs feature extraction through a two-dimensional convolutional layer and obtains multi-scale feature representation in combination with a downsampling operation; The bottleneck layer of the first branch input structure extracts and compresses the feature map after multiple convolutions output by the encoder, and transmits the corresponding high-dimensional information to the decoder of the first branch input structure, wherein the bottleneck layer is used to connect the encoder and the decoder; Based on the output data of the bottleneck layer of the first branch input structure, the decoder of the first branch input structure implements feature fusion through a two-dimensional convolution layer and up-sampling using bilinear interpolation to output first initial feature data; According to the target size, the height and width of the first initial feature data are adjusted by bilinear interpolation to obtain the historical data feature vector.

[0011] In one embodiment, extracting numerical weather forecast features through the second branch input structure to obtain the numerical weather forecast feature vector includes: The encoder of the second branch input structure performs feature extraction through a two-dimensional convolutional layer and obtains multi-scale feature representation in combination with a downsampling operation; The bottleneck layer of the second branch input structure extracts and compresses the feature map after multiple convolutions output by the encoder, and transmits the corresponding high-dimensional information to the decoder of the second branch input structure, wherein the bottleneck layer is used to connect the encoder and the decoder; Based on the output data of the bottleneck layer of the second branch input structure, the decoder of the second branch input structure implements feature fusion through a two-dimensional convolution layer and up-sampling using bilinear interpolation to output second initial feature data; According to the target size, the height and width of the second initial feature data are adjusted by bilinear interpolation to obtain the numerical weather forecast feature vector.

[0012] In one embodiment, based on the output data of the bottleneck layer of the first branch input structure, the decoder of the first branch input structure implements feature fusion through a two-dimensional convolution layer and uses bilinear interpolation for upsampling to output first initial feature data, including: S11: The encoder of the first branch input structure uses bilinear interpolation to perform upsampling and uses two-dimensional convolution to reduce the number of channels; wherein the corresponding correlation formula is as follows: in, is the upsampling operation, It is a two-dimensional convolution operation with a convolution kernel of 3×3; It is The input of the layer is The output of the layer has the following channels: , is the feature matrix after upsampling, where Corresponding to the number of downsampling layers, the number of channels is ,and , Is the feature matrix after two-dimensional convolution, the number of channels is ; The steps of upsampling using bilinear interpolation include: indexing the elements of the output feature matrix Element indices mapped to the input feature matrix , ; Based on the four nearest elements in the input feature matrix Perform interpolation, ; The interpolation formula is: in, ; S12: The result of step S11 With encoder l Feature matrix after layer 2D convolution Splicing is performed in the channel dimension to achieve a skip-step connection between the encoder and the decoder. The corresponding association formula is as follows: in, and The number of channels is , The number of channels is 2 ; S13: Perform two-dimensional convolution on the concatenated feature matrix to further fuse features and reduce the number of channels. The corresponding correlation formula is as follows: in, Number of channels ; S14: Repeat steps S11 to S13 three times to output the first initial feature data .

[0013] In one embodiment, based on the output data of the bottleneck layer of the second branch input structure, the decoder of the second branch input structure implements feature fusion through a two-dimensional convolution layer and uses bilinear interpolation for upsampling to output second initial feature data, including: S21: The encoder of the second branch input structure uses bilinear interpolation to upsample and two-dimensional convolution to reduce the number of channels; wherein the corresponding correlation formula is: in, is the upsampling operation, It is a two-dimensional convolution operation with a convolution kernel of 3×3; It is The input of the layer is The output of the layer has the following channels: , is the feature matrix after upsampling, where Corresponding to the number of downsampling layers, the number of channels is ,and , Is the feature matrix after two-dimensional convolution, the number of channels is ; The steps of upsampling using bilinear interpolation include: indexing the elements of the output feature matrix Element indices mapped to the input feature matrix , ; Based on the four nearest elements in the input feature matrix Perform interpolation , the interpolation formula is: in, ; S22: The result of step S21 With encoder Feature matrix after layer 2D convolution Splicing is performed in the channel dimension to achieve a skip-step connection between the encoder and the decoder. The corresponding association formula is as follows: in and The number of channels is , The number of channels is 2 ; S23: Perform two-dimensional convolution on the concatenated feature matrix to further fuse features and reduce the number of channels: in, Number of channels ; S24: Repeat steps S21 to S23 three times to output the second initial feature data .

[0014] In a second aspect, the present application provides a single-machine wind power ultra-short-term prediction device based on WNet, comprising: An acquisition module is used to obtain pre-processed sample data of the target unit, wherein the sample data includes historical power data, historical meteorological data, and numerical weather forecast data. The pre-processing includes the elimination of abnormal data and the standardization process to eliminate dimensional differences. A first construction module is configured to construct a dual-branch input structure based on the preprocessed sample data, wherein the first branch input structure combines the preprocessed historical power data and the historical meteorological data into a joint input matrix, and the second branch input structure inputs the preprocessed numerical weather forecast data; a second construction module, configured to construct a WNet model suitable for wind power forecasting based on the dual-branch input structure, and obtain a feature processing result output by the WNet model, wherein the feature processing result includes a historical data feature vector corresponding to the first branch input structure and a numerical weather forecast feature vector corresponding to the second branch input structure; A splicing module, configured to splice the historical data feature vector and the numerical weather forecast feature vector in a feature dimension to obtain a combined feature; The output module is used to deeply fuse and transform the combined features through a multi-layer perceptron network and output the wind power prediction result.

[0015] The WNet-based single-machine wind power ultra-short-term prediction method of this application has the following advantages over related technologies: 1. By preprocessing the sample data, abnormal data can be eliminated. The preprocessed sample data is highly reliable and can provide reliable data support for subsequent processing, which is conducive to improving prediction accuracy. In addition, the prediction method of this application comprehensively utilizes historical power data, historical meteorological data, and numerical weather forecast data. Through the dual-branch input structure, the characteristic information of multi-source data is fully mined, thereby further improving the prediction accuracy of single-machine wind power. At the same time, the dual-branch input structure can independently adjust the input data and model design, and can flexibly adapt to the characteristics of different data sources.

[0016] 2. The prediction method of this application constructs a WNet model suitable for wind power prediction based on a dual-branch input structure. Due to the use of the WNet model, multi-scale features can be extracted through two-dimensional convolution and downsampling operations, and multi-level feature fusion is achieved by combining upsampling and skip connection mechanisms, enhancing the model's feature expression capabilities. On this basis, the dual-branch feature extraction results are deeply fused and transformed through a multi-layer perceptron network. The powerful nonlinear fitting capabilities of the multi-layer perceptron network are utilized to flexibly capture the complex correlations between historical data and forecast data, thereby effectively integrating multi-source information, adapting to the prediction needs in different scenarios, and generating more accurate and stable wind power prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Schematic diagram of the process of a single-machine wind power ultra-short-term prediction method based on WNet in one embodiment of the present application; Figure 2 Schematic diagram of the principle of a single-machine wind power ultra-short-term prediction method based on WNet in one embodiment of the present application; Figure 3 This is a schematic diagram of the principle of a single-machine wind power ultra-short-term prediction method based on WNet in another embodiment of the present application; Figure 4 This is a flow chart of step S101 in one embodiment of the present application; Figure 5 This is a flow chart of step S103 in one embodiment of the present application; Figure 6 This is a flow chart of step S502 in one embodiment of the present application; Figure 7 This is a flow chart of step S503 in one embodiment of the present application; Figure 8 Schematic diagram of the structure of a single-machine wind power ultra-short-term prediction device based on WNet in one embodiment of the present application. DETAILED DESCRIPTION

[0019] The following will be combined with the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0020] In some embodiments, as Figure 1 As shown, the present application provides a single-machine wind power ultra-short-term prediction method based on WNet, and the single-machine wind power ultra-short-term prediction method based on WNet includes the following steps S101 to S105.

[0021] S101: Obtain pre-processed sample data of the target unit. The sample data includes historical power data, historical meteorological data, and numerical weather forecast data. The pre-processing includes removing abnormal data and standardizing to eliminate dimensional differences.

[0022] A wind turbine can be selected as the target unit, and historical power data, historical meteorological data, and numerical weather forecast data for the target unit over a predetermined time series length can be collected as sample data. After obtaining the sample data, the data integrity of each time series (for example, a single day can be considered a time series) needs to be checked. If the data integrity of a time series is low, the data in that time series is abnormal and needs to be removed to prevent it from negatively impacting the wind power forecast accuracy.

[0023] By standardizing the sample data, historical power data, historical meteorological data, and numerical weather forecast data can be converted into data of a unified scale, ensuring the comparability of historical power data, historical meteorological data, and numerical weather forecast data, avoiding misleading model weights or result interpretation due to differences in numerical magnitude, thereby avoiding analytical bias caused by dimensional differences and improving the rationality of data visualization and feature engineering.

[0024] S102: Construct a dual-branch input structure based on the preprocessed sample data. The first branch input structure combines the preprocessed historical power data and historical meteorological data into a joint input matrix, while the second branch input structure inputs the preprocessed numerical weather forecast data. The first branch input structure, which combines the historical power data and historical meteorological data into a joint input matrix, primarily represents historical trend change information, while the second branch input structure, which uses the numerical weather forecast data as input, represents future fluctuations.

[0025] It can be understood that the first branch input structure is the historical power data and historical meteorological data branch, and the second branch input structure is the numerical weather forecast data branch. The dual-branch input structure also includes the historical power data and historical meteorological data branches, as well as the numerical weather forecast data branch, which can make full use of various types of power and meteorological information for prediction.

[0026] For example, to meet the actual ultra-short-term prediction requirements, use The data of the first 24 hours at a moment is used to predict the data of the next 4 hours. That is, when the time resolution is 15 minutes, the data of the first 96 time steps are used to predict the data of the next 16 time steps. They are represented by the following vectors:

[0027] in represents the concatenation of two matrices in the feature dimension, Indicates the Time point Historical meteorological elements and power generation. At the same time, in order to adapt to the two-dimensional convolution operation, the input data is reshaped into ,in is the number of channels (feature number), is the height (time step) and 1 is the width, so each feature is regarded as an independent channel. The subsequent two-dimensional convolution kernel captures the relationship between features by sliding on the channel. Similarly, the numerical weather forecast data input can be obtained: S103: Constructing a WNet model suitable for wind power forecasting based on the dual-branch input structure, and obtaining feature processing results output by the WNet model, wherein the feature processing results include a historical data feature vector corresponding to the first branch input structure and a numerical weather forecast feature vector corresponding to the second branch input structure.

[0028] It can be understood that based on the dual-branch input structure in step S102, historical data is used as the input for the first branch input structure (left branch), and numerical weather forecast data is used as the input for the second branch input structure (right branch). Both the first and second branch input structures consist of an encoder-decoder. The encoder extracts features through two-dimensional convolutional layers and combines downsampling operations to obtain multi-scale feature representations. The decoder achieves feature fusion through two-dimensional convolutional layers, uses bilinear interpolation for upsampling to improve resolution, and introduces a skip connection mechanism to enhance feature transfer capabilities. A WNet model suitable for wind power forecasting is constructed based on the dual-branch input structure. The WNet model ultimately outputs a historical data feature vector corresponding to the first branch input structure and a numerical weather forecast feature vector corresponding to the second branch input structure.

[0029] S104: Concatenate the historical data feature vector and the numerical weather forecast feature vector in the feature dimension to obtain a combined feature. The combined feature can be obtained by aligning the resolutions of the historical data feature vector and the numerical weather forecast feature vector and then concatenating them in the feature dimension.

[0030] For example, the historical data feature vector , 96 is the height, 1 is the width; numerical weather forecast feature vector , 16 for height and 1 for width. The resolution of historical data feature vector and numerical weather forecast feature vector is not high, so it is necessary to convert them into Dimensionality reduction: in, is the weight matrix, is the bias vector, and then and Concatenate on the feature dimension:

[0031] Among them, the combined features .

[0032] S105: Deeply fuse and transform the combined features through a multi-layer perceptron network to output the wind power prediction result. The principle of using the single-machine wind power ultra-short-term prediction method based on WNet to predict wind power can be found in Figure 2 , Figure 2 The future meteorological data in is determined based on numerical weather forecast data.

[0033] In this embodiment, the MLP network consists of multiple fully connected layers and activation functions: in, It is a multi-layer perceptron method. is the prediction result. Specifically, Expand into a one-dimensional vector , the final prediction result is obtained through the quartic linear mapping and Gelu activation function .

[0034] The WNet-based single-machine wind power ultra-short-term prediction method can eliminate abnormal data by preprocessing sample data. The preprocessed sample data has high reliability and can provide reliable data support for subsequent processing, which is conducive to improving prediction accuracy. In addition, the prediction method of this application comprehensively utilizes historical power data, historical meteorological data, and numerical weather forecast data. Through a dual-branch input structure, it fully mines the characteristic information of multi-source data, thereby further improving the prediction accuracy of single-machine wind power. At the same time, the dual-branch input structure can independently adjust the input data and model design, and can flexibly adapt to the characteristics of different data sources.

[0035] It can also be understood that the prediction method of this application constructs a WNet model suitable for wind power prediction based on a dual-branch input structure. Due to the use of the WNet model, multi-scale features can be extracted through two-dimensional convolution and downsampling operations, and multi-level feature fusion can be achieved by combining upsampling and skip connection mechanisms, thereby enhancing the model's feature expression capabilities. On this basis, the dual-branch feature extraction results are deeply fused and transformed through a multi-layer perceptron network. The powerful nonlinear fitting capabilities of the multi-layer perceptron network are utilized to flexibly capture the complex correlation between historical data and forecast data, thereby effectively integrating multi-source information, adapting to the prediction needs in different scenarios, and generating more accurate and stable wind power prediction results.

[0036] In some embodiments, the WNet-based single-machine wind power ultra-short-term prediction method further includes: using the Huber loss function as the optimization target and using the Adam optimizer to update the parameters of the multi-layer perceptron network. In combination with the above embodiments, the process of the wind turbine power generation prediction method of the embodiment of the present application can be as follows: Figure 3 shown.

[0037] As you can understand, the Huber loss combines the advantages of both the mean squared error (MSE) and the mean absolute error (MAE). When the error is small, the squared loss is used to accelerate convergence, while when the error is large, the linear loss is used to enhance robustness to outliers. This example uses the Huber loss function as the optimization objective, minimizes the Huber loss using the Adam optimizer, and gradually adjusts the model parameters to ultimately achieve optimal model prediction capabilities.

[0038] The calculation formula of Huber loss is: in, is the true value, Is the threshold parameter used to control the switching point between square loss and linear loss (usually The Huber loss function has the advantages of being sensitive to small errors, accelerating model convergence, and being robust to large errors, reducing the impact of outliers on model training. Using the Adam optimizer for parameter updates, combined with dynamic learning rate adjustment, early stopping, and model saving, can gradually optimize the model and achieve optimal predictive capabilities.

[0039] In some embodiments, as Figure 4 As shown, in step S101, the sample data of the target unit after preprocessing is obtained, including the following steps S401 to S403.

[0040] S401: Obtain sample data within a predetermined time series length.

[0041] Get the sample wind turbine in the time series length Historical power generation data P Historical weather data And the corresponding numerical weather forecast data F , forming sample data, where the historical meteorological data contains M meteorological elements and the numerical weather forecast data contains N meteorological elements, which are represented by the following vectors respectively:

[0042] in, is the length of the time series, Indicates the Historical power generation data at a given point in time, Indicates the The time point meteorological elements, No. The time point A meteorological element.

[0043] S402: Incomplete daily data whose consecutive missing time exceeds a preset time threshold is eliminated as abnormal data, and data whose consecutive missing time does not exceed the preset time threshold is filled using linear interpolation to obtain sample data with uniform resolution.

[0044] Among them, for the historical power data P , historical meteorological data , and numerical weather forecast data , check whether there is continuous missing data exceeding a preset time threshold (for example, 2 hours) in each time series, and if so, delete the data for that day.

[0045] The data that are missing continuously but do not exceed the preset time threshold are filled by linear interpolation. For example, for missing values , and the nearest non-missing value is and (in ),but: The processing of historical meteorological data and numerical weather forecast data is similar. At the same time, the data is resampled using the above linear interpolation method to make the resolution uniform, for example, uniform to 15 minutes, and the processed historical power data P and historical meteorological data are obtained. and numerical weather forecast data .

[0046] S403: Standardize the sample data with uniform resolution to obtain pre-processed sample data.

[0047] The normalization process involves centralizing and scaling the data. For example, the normalization formula is:

[0048] in for mean, is the standard deviation, and similarly, the standardized historical meteorological data is obtained and numerical weather forecast data .

[0049] In some embodiments, as Figure 5 As shown, in step S103, a WNet model suitable for wind power prediction is constructed based on a dual-branch input structure, and a feature processing result output by the WNet model is obtained, including the following steps S501 to S503.

[0050] S501: Adjust the dual-branch input structure to adjust the height and width of the input data to an even number and not less than a preset value.

[0051] Among them, the left and right branches of the dual-branch input structure are both composed of an encoder-decoder, and the encoder part performs feature extraction through a two-dimensional convolution layer. Two-dimensional convolution is a convolution for image data, and its operation is similar to one-dimensional convolution, but its window and convolution kernel are both two-dimensional. This application migrates two-dimensional convolution to time series prediction, and uses the number of features as the number of channels for convolution. In order to ensure that the input data dimension is sufficient for subsequent convolution and downsampling operations, it is necessary to first perform a padding operation on the input data of the dual-branch input structure, and adjust the height and width of the input data to an even number and not less than the preset value, so that it can be evenly downsampled in subsequent pooling operations.

[0052] S502: Extract historical data features through the first branch input structure to obtain a historical data feature vector. The first branch input structure is an encoder-decoder structure, where the upsampling operation of each decoder layer corresponds to the downsampling operation in the encoder. The first branch input structure combines the multi-scale features extracted by the encoder with the upsampling results of the decoder through a skip connection.

[0053] The first branch input structure is an encoder-decoder architecture. The encoder extracts features through two-dimensional convolutional layers and combines them with downsampling to obtain multi-scale feature representations, capturing the complex correlation between wind power variations and meteorological conditions. The decoder then uses bilinear interpolation for upsampling to increase resolution. A skip connection mechanism is introduced to effectively retain and transfer key features extracted at multiple scales, enhancing feature transfer capabilities and thus extracting appropriate feature vectors for historical data.

[0054] S503: Numerical weather forecast features are extracted through the second branch input structure to obtain a numerical weather forecast feature vector. The second branch input structure is an encoder-decoder structure that combines the multi-scale features extracted by the encoder with the upsampling results of the decoder through a skip connection.

[0055] Similarly, the second branch's input structure also uses an encoder-decoder architecture. The encoder captures the multi-dimensional features of the numerical weather forecast data through layers of two-dimensional convolution and downsampling. The decoder combines bilinear interpolation upsampling with a skip connection mechanism, enabling this branch to improve resolution while preserving multi-scale features of future weather conditions, providing strong support for short-term wind power forecasting.

[0056] In some embodiments, the dual-branch input structure is adjusted to adjust the height and width of the input data to an even number and not less than a preset value, including: performing a padding operation on the input data of the dual-branch input structure to adjust the height and width of the input data to an even number and not less than 8, wherein the padding operation uses zeros for padding.

[0057] In this embodiment, the height and width of the input data are adjusted to an even number and not less than 8, so that they can be evenly downsampled in the subsequent pooling operation. , the height after filling for: ,in express If the remainder after dividing by 2 is still less than 8, the height is directly padded to 8. The same is true for the width, and zero padding is used when filling.

[0058] In some embodiments, as Figure 6 As shown, in step S502, historical data feature extraction is performed through the first branch input structure to obtain a historical data feature vector, including the following steps S601 to S604.

[0059] S601: The encoder of the first branch input structure performs feature extraction through a two-dimensional convolutional layer and obtains a multi-scale feature representation in combination with a downsampling operation.

[0060] Specifically, S601 may include step 601a and step 601b.

[0061] 601a: Multi-scale feature extraction is performed through two-dimensional convolution and downsampling to capture the complex relationship between wind power changes and meteorological conditions. As an example, the input is the output feature matrix of the previous layer. , and then use the two-dimensional convolution kernel to extract features: in, represents a two-dimensional convolution, represents the maximum pooling operation, It is The output of the layer is The input of the layer has the following channels: , It is the feature matrix after two-dimensional convolution and is used as the subsequent skip connection. The number of channels is , It is The output of the layer has the following channels: ,in That is, the number of channels is doubled after the two-dimensional convolution operation. The convolution kernel size of the two-dimensional convolution is 3×3, and one row / column is padded with zeros around the input matrix to ensure that the size of the feature matrix output after convolution is the same as the input matrix size, avoiding the loss of edge information. The pooling kernel size of the maximum pooling is 2×2, and the stride is 2. The specific steps are as follows: in, and is the element position index of the output matrix, and is the element position index within the pooling kernel.

[0062] 601b: Repeat step 601a three times to obtain the final encoder output .

[0063] S602: The bottleneck layer of the first branch input structure extracts and compresses the feature map after multiple convolutions output by the encoder, and transmits the corresponding high-dimensional information to the decoder of the first branch input structure. The bottleneck layer is used to connect the encoder and the decoder.

[0064] Among them, the bottleneck layer is the key part connecting the encoder and the decoder. The bottleneck layer is used to perform deep feature extraction and compression on the feature map output by the encoder after multiple convolutions, and pass the corresponding high-dimensional information to the decoder. The corresponding formula is: S603: Based on the output data of the bottleneck layer of the first branch input structure, the decoder of the first branch input structure realizes feature fusion through a two-dimensional convolution layer, and uses bilinear interpolation for upsampling to output first initial feature data.

[0065] S603 may include steps S11 to S14.

[0066] S11: The encoder of the first branch input structure uses bilinear interpolation to upsample and uses two-dimensional convolution to reduce the number of channels; wherein the corresponding association formula is as follows: in, is the upsampling operation, It is a two-dimensional convolution operation with a convolution kernel of 3×3; It is The input of the layer is The output of the layer has the following channels: , is the feature matrix after upsampling, where Corresponding to the number of downsampling layers, the number of channels is ,and , Is the feature matrix after two-dimensional convolution, the number of channels is ; The steps of upsampling using bilinear interpolation include: indexing the elements of the output feature matrix Element indices mapped to the input feature matrix , ; Based on the four nearest elements in the input feature matrix Perform interpolation, ; The interpolation formula is: in, .

[0067] S12: The result of step S11 With encoder l Feature matrix after layer 2D convolution Splicing is performed in the channel dimension to achieve a skip-step connection between the encoder and the decoder. The corresponding association formula is as follows: in, and The number of channels is , The number of channels is 2 .

[0068] S13: Perform two-dimensional convolution on the concatenated feature matrix to further fuse features and reduce the number of channels. The corresponding correlation formula is as follows: in, Number of channels .

[0069] S14: Repeat steps S11 to S13 three times to output the first initial feature data .

[0070] S604: According to the target size, adjust the height and width of the first initial feature data by bilinear interpolation to obtain the historical data feature vector. For example, the target size is , then adjust the height and width of the first initial feature data by bilinear interpolation, and output the historical data feature vector .

[0071] In some embodiments, as Figure 7 As shown, in step S503, numerical weather forecast feature extraction is performed through the second branch input structure to obtain a numerical weather forecast feature vector, including the following steps S701 to S704.

[0072] S701: The encoder of the second branch input structure performs feature extraction through a two-dimensional convolution layer and obtains a multi-scale feature representation in combination with a downsampling operation. The encoder of the second branch input structure uses a two-dimensional convolution layer to perform deep feature extraction on historical meteorological data and obtains a multi-scale feature representation in combination with a downsampling operation. The encoder systematically extracts the spatiotemporal features of the input data through layer-by-layer convolution and pooling operations, while providing multi-level feature information for subsequent skip connections through a feature retention mechanism. Specifically, S701 may include steps 701a and 701b.

[0073] 701a: Multi-scale feature extraction through 2D convolution and downsampling to capture the relationship between numerical weather forecast data. As an example, the input is the output feature matrix of the previous layer. , and then use 2D convolution and downsampling for feature extraction: in, represents a two-dimensional convolution, represents the maximum pooling operation, It is The output of the layer is The input of the layer has the following channels: , It is the feature matrix after two-dimensional convolution and is used as the subsequent skip connection. The number of channels is , It is The output of the layer has the following channels: ,in That is, the number of channels is doubled after the two-dimensional convolution operation. The convolution kernel size of the two-dimensional convolution is 3×3, and one row / column is padded with zeros around the input image to ensure that the size of the feature matrix output after convolution is the same as the input matrix size. The pooling kernel size of the maximum pooling is 2×2, and the stride is 2. The specific steps are as follows:

[0074] in, and is the element position index of the output matrix, and is the element position index within the pooling kernel.

[0075] 701b: Repeat step 701a three times to obtain the final encoder output .

[0076] S702: The bottleneck layer of the second branch input structure extracts and compresses the feature map after multiple convolutions of the encoder output, and passes the corresponding high-dimensional information to the decoder of the second branch input structure. The bottleneck layer is used to connect the encoder and the decoder. Similar to the bottleneck layer of the first branch input structure, the bottleneck layer of the second branch input structure further extracts and compresses the final output of the encoder, and then uses the result as the input of the decoder. The operation formula of the bottleneck layer is as follows: in, The number of channels is twice as much.

[0077] S703: Based on the output data of the bottleneck layer of the second branch input structure, the decoder of the second branch input structure realizes feature fusion through a two-dimensional convolution layer, and uses bilinear interpolation for upsampling to output second initial feature data.

[0078] S21: The encoder of the second branch input structure uses bilinear interpolation to upsample and two-dimensional convolution to reduce the number of channels; wherein the corresponding association formula is: in, is the upsampling operation, It is a two-dimensional convolution operation with a convolution kernel of 3×3; It is The input of the layer is The output of the layer has the following channels: , is the feature matrix after upsampling, where Corresponding to the number of downsampling layers, the number of channels is ,and , Is the feature matrix after two-dimensional convolution, the number of channels is ; The steps of upsampling using bilinear interpolation include: indexing the elements of the output feature matrix Element indices mapped to the input feature matrix , ; Based on the four nearest elements in the input feature matrix Perform interpolation , the interpolation formula is: in, .

[0079] S22: The result of step S21 With encoder Feature matrix after layer 2D convolution Splicing is performed in the channel dimension to achieve a skip-step connection between the encoder and the decoder. The corresponding association formula is as follows: in and The number of channels is , The number of channels is 2 .

[0080] S23: Perform two-dimensional convolution on the concatenated feature matrix to further fuse features and reduce the number of channels: in, Number of channels .

[0081] S24: Repeat steps S21 to S23 three times to output the second initial feature data .

[0082] S704: According to the target size, adjust the height and width of the second initial feature data by bilinear interpolation to obtain the numerical weather forecast feature vector. For example, the target size is , then the height and width of the second initial feature data are adjusted by bilinear interpolation, and the numerical weather forecast feature vector is output. .

[0083] Based on the method of the above embodiment, a unit of a certain MW wind farm station is used as the target unit and verification unit. This embodiment is verified by predicting the power generation of a unit of the wind farm station on July 25, 2024. The historical power data, historical meteorological data and numerical weather forecast data of the unit from March 1, 2023 to August 22, 2024 are collected. Among them, the historical meteorological data and numerical weather forecast data are the data of the wind tower of the station; the incomplete daily data with missing values ​​for more than 2 consecutive hours are eliminated, and the other missing data are filled with linear interpolation; considering that the time resolution of the actual power generation forecast report is 15 minutes. The single-machine wind power ultra-short-term prediction method based on WNet has a certain improvement in effect compared with other prediction models. The specific situation is shown in the table below.

[0084]

[0085] In some embodiments, as Figure 8 As shown, the present application also provides a single-machine wind power ultra-short-term prediction device 800 based on WNet, including: an acquisition module 801, a first construction module 802, a second construction module 803, a splicing module 804 and an output module 805.

[0086] The acquisition module 801 is used to obtain pre-processed sample data of the target unit. The sample data includes historical power data, historical meteorological data and numerical weather forecast data. The pre-processing includes the elimination of abnormal data and the standardization process to eliminate dimensional differences.

[0087] The first construction module 802 is used to construct a dual-branch input structure based on the preprocessed sample data, wherein the first branch input structure combines the preprocessed historical power data and the historical meteorological data into a joint input matrix, and the second branch input structure inputs the preprocessed numerical weather forecast data.

[0088] The second construction module 803 is used to construct a WNet model suitable for wind power prediction based on a dual-branch input structure, and obtain the feature processing results output by the WNet model, wherein the feature processing results include the historical data feature vector corresponding to the first branch input structure and the numerical weather forecast feature vector corresponding to the second branch input structure.

[0089] The splicing module 804 is used to splice the historical data feature vector and the numerical weather forecast feature vector in the feature dimension to obtain a combined feature.

[0090] The output module 805 is used to perform deep fusion and transformation on the combined features through a multi-layer perceptron network and output the wind power prediction result.

[0091] In some embodiments, the WNet-based single-machine wind power ultra-short-term prediction device further includes: an optimization module, which is used to use the Huber loss function as the optimization target and use the Adam optimizer to update the parameters of the multi-layer perceptron network.

[0092] It should be noted that the WNet-based single-machine wind power ultra-short-term prediction device 800 provided in the embodiment of the present application and the WNet-based single-machine wind power ultra-short-term prediction method provided in the embodiment of the present application are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned WNet-based single-machine wind power ultra-short-term prediction method, and the repeated parts will not be repeated.

[0093] In some embodiments, an embodiment of the present application further provides an electronic device comprising a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the above-mentioned WNet-based single-machine wind power ultra-short-term prediction method.

[0094] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or related chipsets and / or a dedicated microprocessor (such as an application-specific integrated circuit (ASIC)), and the like. The processor may also include on-board memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to the embodiment of the present application. The memory may, for example, be any medium capable of containing, storing, conveying, propagating or transmitting instructions. For example, the memory may include, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared or semiconductor system, device, device or propagation medium. Specific examples of memory include: magnetic storage devices, such as magnetic tape or hard disk (HDD); optical storage devices, such as compact disc (CD-ROM); it may also be random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0095] The present application also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned WNet-based single-machine wind power ultra-short-term prediction method.

[0096] According to an embodiment of the present application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device, or component. In the present application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries a computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, radio frequency signals, or any suitable combination thereof.

[0097] Those skilled in the art will understand that the features described in the various embodiments and / or claims of the present application may be combined and / or combined in a variety of ways, even if such combinations or combinations are not explicitly described in the present application. In particular, without departing from the spirit and teachings of the present application, the features described in the various embodiments and / or claims of the present application may be combined and / or combined in a variety of ways. All of these combinations and / or combinations fall within the scope of the present application. Therefore, the scope of the present application should not be limited to the above-mentioned embodiments, but should be determined not only by the attached claims, but also by the equivalents of the attached claims. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A single-machine wind power ultra-short-term prediction method based on WNet, characterized by: include: Obtaining pre-processed sample data of the target unit, the sample data including historical power data, historical meteorological data, and numerical weather forecast data, the pre-processing including elimination of abnormal data and standardization to eliminate dimensional differences; Constructing a dual-branch input structure based on the preprocessed sample data, wherein the first branch input structure combines the preprocessed historical power data and the historical meteorological data into a joint input matrix, and the second branch input structure inputs the preprocessed numerical weather forecast data; Building a WNet model suitable for wind power forecasting based on the dual-branch input structure, and obtaining a feature processing result output by the WNet model, wherein the feature processing result includes a historical data feature vector corresponding to the first branch input structure and a numerical weather forecast feature vector corresponding to the second branch input structure; splicing the historical data feature vector and the numerical weather forecast feature vector in the feature dimension to obtain a combined feature; The combined features are deeply fused and transformed through a multi-layer perceptron network to output wind power prediction results.

2. The WNet-based single-machine wind power ultra-short-term prediction method according to claim 1, characterized in that: The method further comprises: The Huber loss function is used as the optimization target, and the Adam optimizer is used to update the parameters of the multilayer perceptron network.

3. The WNet-based single-machine wind power ultra-short-term prediction method according to claim 1, characterized in that: The obtaining of pre-processed sample data of the target unit includes: Obtain sample data within a predetermined time series length; Incomplete daily data with consecutive missing data exceeding the preset time threshold are removed as abnormal data, and data with consecutive missing data within the preset time threshold are filled using linear interpolation to obtain sample data with uniform resolution; The sample data of uniform resolution is standardized to obtain the pre-processed sample data.

4. The WNet-based single-machine wind power ultra-short-term prediction method according to claim 1, characterized in that: The step of constructing a WNet model suitable for wind power prediction based on the dual-branch input structure and obtaining a feature processing result output by the WNet model includes: Adjusting the dual-branch input structure to adjust the height and width of the input data to be an even number and not less than a preset value; Extracting historical data features through a first branch input structure to obtain the historical data feature vector; wherein the first branch input structure is an encoder-decoder structure, the upsampling operation of each layer of the decoder corresponds to the downsampling operation in the encoder, and the first branch input structure combines the multi-scale features extracted by the encoder with the upsampling results of the decoder through a skip connection; Numerical weather forecast features are extracted through the second branch input structure to obtain the numerical weather forecast feature vector. The second branch input structure is an encoder-decoder structure. The second branch input structure combines the multi-scale features extracted by the encoder with the upsampling results of the decoder through a skip connection.

5. The WNet-based single-machine wind power ultra-short-term prediction method according to claim 4, characterized in that: The step of adjusting the dual-branch input structure to adjust the height and width of the input data to be an even number and not less than a preset value includes: A padding operation is performed on the input data of the dual-branch input structure to adjust the height and width of the input data to an even number and not less than 8, wherein the padding operation is performed using zeros.

6. The WNet-based single-machine wind power ultra-short-term prediction method according to claim 4, characterized in that: The extracting historical data features through the first branch input structure to obtain the historical data feature vector includes: The encoder of the first branch input structure performs feature extraction through a two-dimensional convolutional layer and obtains multi-scale feature representation in combination with a downsampling operation; The bottleneck layer of the first branch input structure extracts and compresses the feature map after multiple convolutions output by the encoder, and transmits the corresponding high-dimensional information to the decoder of the first branch input structure, wherein the bottleneck layer is used to connect the encoder and the decoder; Based on the output data of the bottleneck layer of the first branch input structure, the decoder of the first branch input structure implements feature fusion through a two-dimensional convolution layer and performs upsampling using bilinear interpolation to output first initial feature data; According to the target size, the height and width of the first initial feature data are adjusted by bilinear interpolation to obtain the historical data feature vector.

7. The WNet-based single-machine wind power ultra-short-term prediction method according to claim 4, characterized in that: The extracting numerical weather forecast features through the second branch input structure to obtain the numerical weather forecast feature vector includes: The encoder of the second branch input structure performs feature extraction through a two-dimensional convolutional layer and obtains multi-scale feature representation in combination with a downsampling operation; The bottleneck layer of the second branch input structure extracts and compresses the feature map after multiple convolutions output by the encoder, and transmits the corresponding high-dimensional information to the decoder of the second branch input structure, wherein the bottleneck layer is used to connect the encoder and the decoder; Based on the output data of the bottleneck layer of the second branch input structure, the decoder of the second branch input structure implements feature fusion through a two-dimensional convolution layer and up-sampling using bilinear interpolation to output second initial feature data; According to the target size, the height and width of the second initial feature data are adjusted by bilinear interpolation to obtain the numerical weather forecast feature vector.

8. The WNet-based single-machine wind power ultra-short-term prediction method according to claim 6, characterized in that: The output data of the bottleneck layer based on the first branch input structure, the decoder of the first branch input structure implements feature fusion through a two-dimensional convolution layer, and uses bilinear interpolation for upsampling to output first initial feature data, including: S11: The encoder of the first branch input structure uses bilinear interpolation to perform upsampling and uses two-dimensional convolution to reduce the number of channels; wherein the corresponding correlation formula is as follows: in, is the upsampling operation, It is a two-dimensional convolution operation with a convolution kernel of 3×3; It is The input of the layer is The output of the layer has the following channels: , is the feature matrix after upsampling, where Corresponding to the number of downsampling layers, the number of channels is ,and , Is the feature matrix after two-dimensional convolution, the number of channels is ; The steps of upsampling using bilinear interpolation include: indexing the elements of the output feature matrix Element indices mapped to the input feature matrix , ; Based on the four nearest elements in the input feature matrix Perform interpolation, ; The interpolation formula is: in, ; S12: The result of step S11 With encoder l Feature matrix after layer 2D convolution Splicing is performed in the channel dimension to achieve a skip-step connection between the encoder and the decoder. The corresponding association formula is as follows: in, and The number of channels is , The number of channels is 2 ; S13: Perform two-dimensional convolution on the concatenated feature matrix to further fuse features and reduce the number of channels. The corresponding correlation formula is as follows: in, Number of channels ; S14: Repeat steps S11 to S13 three times to output the first initial feature data .

9. The WNet-based single-machine wind power ultra-short-term prediction method according to claim 7, characterized in that: The output data of the bottleneck layer based on the second branch input structure, the decoder of the second branch input structure implements feature fusion through a two-dimensional convolution layer, and uses bilinear interpolation for upsampling to output second initial feature data, including: S21: The encoder of the second branch input structure uses bilinear interpolation to upsample and two-dimensional convolution to reduce the number of channels; wherein the corresponding correlation formula is: in, is the upsampling operation, It is a two-dimensional convolution operation with a convolution kernel of 3×3; It is The input of the layer is The output of the layer has the following channels: , is the feature matrix after upsampling, where Corresponding to the number of downsampling layers, the number of channels is ,and , Is the feature matrix after two-dimensional convolution, the number of channels is ; The steps of upsampling using bilinear interpolation include: indexing the elements of the output feature matrix Element indices mapped to the input feature matrix , ; Based on the four nearest elements in the input feature matrix Perform interpolation , the interpolation formula is: in, ; S22: The result of step S21 With encoder Feature matrix after layer 2D convolution Splicing is performed in the channel dimension to achieve a skip-step connection between the encoder and the decoder. The corresponding association formula is as follows: in and The number of channels is , The number of channels is 2 ; S23: Perform two-dimensional convolution on the concatenated feature matrix to further fuse features and reduce the number of channels: in, Number of channels ; S24: Repeat steps S21 to S23 three times to output the second initial feature data .

10. A single-machine wind power ultra-short-term prediction device based on WNet, characterized in that: include: An acquisition module is used to obtain pre-processed sample data of the target unit, wherein the sample data includes historical power data, historical meteorological data, and numerical weather forecast data. The pre-processing includes the elimination of abnormal data and the standardization process to eliminate dimensional differences. A first construction module is configured to construct a dual-branch input structure based on the preprocessed sample data, wherein the first branch input structure combines the preprocessed historical power data and the historical meteorological data into a joint input matrix, and the second branch input structure inputs the preprocessed numerical weather forecast data; a second construction module, configured to construct a WNet model suitable for wind power forecasting based on the dual-branch input structure, and obtain a feature processing result output by the WNet model, wherein the feature processing result includes a historical data feature vector corresponding to the first branch input structure and a numerical weather forecast feature vector corresponding to the second branch input structure; A splicing module, configured to splice the historical data feature vector and the numerical weather forecast feature vector in a feature dimension to obtain a combined feature; The output module is used to deeply fuse and transform the combined features through a multi-layer perceptron network and output the wind power prediction result.

Images