A hybrid network-based ultra-short-term wind power prediction method

Abstract

The application discloses a kind of based on hybrid network's ultra-short-term wind power prediction method.The method includes the following steps: based on data-driven network branch and physical driving network branch constructs ultra-short-term wind power prediction model;Data-driven network branch fuses time convolution neural network and Transformer encoder;Based on the historical meteorological time series data and historical SCADA state data of wind farm, the training set of ultra-short-term wind power prediction model is constructed, and the training set of ultra-short-term wind power prediction model is used to train the ultra-short-term wind power prediction model;Based on the trained ultra-short-term wind power prediction model, the ultra-short-term wind power prediction result is obtained.The application overcomes the defects that existing pure data-driven model is poor in explainability and is insufficient in effectiveness in data-scarce scenarios, realizes high-precision and conforms to the ultra-short-term wind power prediction of physical law, and provides reliable technical support for the safe and stable dispatching of power grid.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power prediction technology, in particular to a hybrid network-based ultra-short-term wind power prediction method. BACKGROUND

[0002] With the increasing proportion of wind power in the energy structure, the randomness and volatility of wind power pose challenges to the safe and stable operation of the power grid. The randomness and volatility of wind power are often more pronounced in the next 15 minutes to 4 hours of real-time dispatching and operation of the power grid. Therefore, high-precision ultra-short-term (time resolution of 15 minutes) wind power prediction is a key technology to address this challenge.

[0003] Existing prediction methods can be mainly divided into two categories: methods based on physical principles and methods based on data-driven. Although the methods based on physical principles are theoretically reliable, they are complex to model, computationally intensive, and sensitive to initial conditions. The data-driven methods, especially deep learning models, can learn complex nonlinear relationships from historical data and have made significant progress.

[0004] However, pure data-driven deep learning models have the following inherent defects: the model is like a "black box", and its prediction results lack physical basis and are difficult to explain; in the scenario of data scarcity or high noise, the model performance drops sharply, and even may produce prediction results that violate physical laws; although the standard Transformer model is good at capturing long-range dependencies, it has weak perception of local details and short-term fluctuation patterns in time series, and is not well adapted to power mutation scenarios.

[0005] The emergence of physical information neural networks provides a new approach to introducing physical laws into deep learning. However, how to effectively integrate physical information neural networks with advanced time series prediction models to build a hybrid prediction model with high accuracy, strong robustness, and good interpretability remains a technical problem to be solved. SUMMARY

[0006] To address the above deficiencies in the prior art, the present application provides a hybrid network-based ultra-short-term wind power prediction method.

[0007] To achieve the above-mentioned application purposes, the technical solutions adopted by the present application are as follows:

[0008] A hybrid network-based ultra-short-term wind power prediction method, comprising the following steps:

[0009] An ultra-short-term wind power prediction model is constructed based on a data-driven network branch and a physics-driven network branch. The data-driven network branch integrates a time convolutional neural network and a Transformer encoder;

[0010] construct a training set of the ultra-short-term wind power prediction model based on historical meteorological time series data and historical SCADA (Supervisory Control and Data Acquisition) state data of the wind farm, and train the ultra-short-term wind power prediction model by using the training set of the ultra-short-term wind power prediction model;

[0011] obtain an ultra-short-term wind power prediction result based on the trained ultra-short-term wind power prediction model and real-time meteorological time series data and real-time SCADA state data of the wind farm.

[0012] Further, the ultra-short-term wind power prediction model comprises a data-driven network branch, a physics-driven network branch, a splicing layer and a fully connected layer; an input end of the data-driven network branch serves as a first data input end, an output end of the data-driven network branch is connected to a first input port of the splicing layer, an input end of the physics-driven network branch serves as a second data input end, an output end of the physics-driven network branch is connected to a second input port of the splicing layer, an output end of the splicing layer is connected to an input end of the fully connected layer, and an output end of the fully connected layer serves as an output of the ultra-short-term wind power prediction model.

[0013] Further, the data-driven network branch comprises a data-driven input layer, a linear adaptation layer, a position encoding layer, a TCN-Transformer module and a pooling layer connected in sequence;

[0014] The data-driven input layer is configured to input the normalized meteorological time series data.

[0015] The linear adaptation layer is configured to adjust a feature dimension of the normalized meteorological time series data to obtain meteorological time series data meeting an input requirement of a Transformer encoder;

[0016] The position encoding layer is configured to perform position encoding on the meteorological time series data meeting the input requirement of the Transformer encoder to obtain hidden representation sequence data.

[0017] The TCN-Transformer module comprises three identical TCN-Transformer layers connected in sequence; the TCN-Transformer module is configured to process the hidden representation sequence data by a time convolutional neural network and a Transformer encoder to obtain multi-layer fusion data.

[0018] The pooling layer is configured to perform average pooling on the multi-layer fusion data to obtain global time sequence feature data.

[0019] Further, the data processing process of the TCN-Transformer layer is specifically as follows:

[0020] processing the hidden representation sequence data by a temporal convolutional neural network to obtain local trend feature data;

[0021] adding the hidden representation sequence data and the local trend feature data by a residual connection layer to obtain locally enhanced sequence data;

[0022] performing self-attention calculation on the locally enhanced sequence data by a Transformer encoder to output fusion data of the TCN-Transformer layer.

[0023] Further, the physical driving network branch comprises a physical driving input layer, a PhyNet layer and a physical quantity calculation layer connected in sequence;

[0024] The physical driving input layer is configured to input a tip speed ratio and a pitch angle of the wind turbine generator set;

[0025] The PhyNet layer is configured to predict a wind energy utilization coefficient of the wind turbine generator set according to the tip speed ratio and the pitch angle of the wind turbine generator set;

[0026] The physical quantity calculation layer is configured to calculate a super-short-term wind power theoretical value of the wind turbine generator set according to the wind energy utilization coefficient of the wind turbine generator set, and convert the super-short-term wind power theoretical value into physical feature data meeting input requirements of the Transformer encoder.

[0027] Further, an expression of the physical quantity calculation layer is as follows:

[0028]

[0029] wherein: is a super-short-term wind power theoretical value of the wind turbine generator set, is air density, is a circular constant, is a blade radius of the wind turbine generator set, is a wind speed, is a wind energy utilization coefficient of the wind turbine generator set, is a tip speed ratio of the wind turbine generator set, is a pitch angle of the wind turbine generator set.

[0030] Further, types of the meteorological time series data include a wind speed, a wind direction, a temperature, a humidity and an air pressure; types of the SCADA state data include a rotor speed and a pitch angle of the wind turbine generator set.

[0031] Further, a training set of the super-short-term wind power prediction model is constructed based on historical meteorological time series data and historical SCADA state data of a wind farm, and a specific process is as follows:

[0032] Minimum-maximum normalization was performed on the historical meteorological time-series data and historical SCADA status data of the wind farm to obtain normalized historical meteorological time-series data and historical SCADA status data.

[0033] The tip speed of the wind turbine generator is calculated based on the wind speed in the normalized historical meteorological time series data and the rotor speed in the historical SCADA status data. The training set of the ultra-short-term wind power prediction model is constructed based on the tip speed of the wind turbine generator, the pitch angle of the wind turbine generator in the normalized historical SCADA status data, and the normalized historical meteorological time series data.

[0034] Furthermore, the ultra-short-term wind power prediction model is trained using the training set of the ultra-short-term wind power prediction model. The specific process is as follows:

[0035] We construct data-driven loss, physical data error loss, and physical consistency loss, and based on these, we construct the loss function for the ultra-short-term wind power prediction model, the expression of which is:

[0036]

[0037] in: For the loss of the ultra-short-term wind power prediction model, For data-driven loss, The weighting coefficients for physical data error loss. For physical data error loss, The weighting coefficients for the physical consistency loss are... This represents a loss of physical consistency.

[0038] Based on the loss function of the ultra-short-term wind power prediction model, the ultra-short-term wind power prediction model is trained using the training set of the ultra-short-term wind power prediction model.

[0039] Furthermore, the expression for data-driven loss is:

[0040]

[0041] in: This represents the number of samples in the training set. The sample IDs for the training set. For the first The actual value of ultra-short-term wind power of a sample of wind turbine generators. For the first Ultra-short-term wind power forecast values ​​for individual samples of wind turbine generator sets;

[0042] The expression for physical data error loss is:

[0043]

[0044] wherein: is the actual value of the ultra-short-term wind power rotation speed of the wind turbine set for the first sample, is the predicted value of the ultra-short-term wind power rotation speed of the wind turbine set for the first sample;

[0045] The expression of the physical consistency loss is:

[0046]

[0047] wherein: is the moment of inertia, is the angular acceleration of the wind turbine set for the first sample, is the theoretical value of the ultra-short-term wind power of the wind turbine set for the first sample.

[0048] The beneficial effects of the present application are:

[0049] (1) In the construction of the ultra-short-term wind power prediction model, the data-driven network branch fusing the time convolutional neural network and the Transformer encoder effectively captures the local details and global rules of the meteorological time series data, and the physical priori knowledge is introduced through the physical-driven network branch, which together ensures the high accuracy of the prediction results.

[0050] (2) By introducing the physical-driven network branch, the ultra-short-term wind power prediction model can still give reasonable prediction according to the physical law in the scene of data scarcity or large noise interference, avoiding the overfitting and performance collapse problem of the pure data-driven model in the data scarcity scene, and the prediction result of the ultra-short-term wind power prediction model not only depends on the historical data fitting, but also is constrained by the explicit physical equation, making the prediction behavior more interpretable and conforming to the engineering practice.

[0051] (3) In the ultra-short-term wind power prediction, there are the following situations: power climbing or falling caused by a sudden wind speed change; power high-frequency oscillation caused by continuous turbulence; power response produced by specific control actions (such as pitch angle fine-tuning) in a short time; therefore, if the ultra-short-term wind power prediction model cannot capture these local details, the prediction performance in the ultra-short-term scene will be greatly reduced, so the present application introduces the time convolutional neural network to make up for the insufficient perception of local mutation information of the Transformer, so that the prediction curve of the ultra-short-term wind power prediction model is more smooth and stable in the mutation scene of power rapid climbing or falling, reducing the overreaction to short-term violent fluctuations. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A flowchart of a hybrid network-based ultra-short-term wind power prediction method;

[0053] Figure 2 A schematic diagram of ultra-short-term wind power prediction results of different models;

[0054] Figure 3 A schematic diagram of ultra-short-term wind power prediction results of different models in a data-scarce scenario. DETAILED DESCRIPTION

[0055] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application as defined and determined by the appended claims, and all applications utilizing the concept of the present application are within the scope of protection.

[0056] As shown in Figure 1 A hybrid network-based ultra-short-term wind power prediction method includes steps S1-S3, specifically as follows:

[0057] S1, constructing an ultra-short-term wind power prediction model based on a data-driven network branch and a physics-driven network branch; the data-driven network branch integrates a time convolutional neural network and a Transformer encoder.

[0058] In an optional embodiment of the present application, the ultra-short-term wind power prediction model includes a data-driven network branch, a physics-driven network branch, a concatenation layer, and a fully connected layer; the input end of the data-driven network branch is a first data input end, the output end of the data-driven network branch is connected to a first input port of the concatenation layer, the input end of the physics-driven network branch is a second data input end, the output end of the physics-driven network branch is connected to a second input port of the concatenation layer, the output end of the concatenation layer is connected to the input end of the fully connected layer, and the output end of the fully connected layer is the output of the ultra-short-term wind power prediction model.

[0059] The data-driven network branch comprises, which are connected in sequence, a data-driven input layer, a linear adaptation layer, a position encoding layer, a TCN-Transformer module and a pooling layer. The data-driven input layer is used for inputting normalized meteorological time series data. The linear adaptation layer is used for adjusting the feature dimension of the normalized meteorological time series data to obtain meteorological time series data meeting the input requirements of the Transformer encoder. The position encoding layer is used for position encoding the meteorological time series data meeting the input requirements of the Transformer encoder to obtain hidden representation sequence data. The TCN-Transformer module comprises three identical TCN-Transformer layers connected in sequence; the TCN-Transformer module is used for processing the hidden representation sequence data through a time convolutional neural network and a Transformer encoder to obtain multi-layer fusion data. The data processing process of the TCN-Transformer layer is specifically: processing the hidden representation sequence data through a time convolutional neural network to obtain local trend feature data; adding the hidden representation sequence data and the local trend feature data through a residual connection layer to obtain locally enhanced sequence data; performing self-attention calculation on the locally enhanced sequence data through a Transformer encoder to output fusion data of the TCN-Transformer layer. The pooling layer is used for average pooling the multi-layer fusion data to obtain global time series feature data.

[0060] The physical-driven network branch comprises, which are connected in sequence, a physical-driven input layer, a PhyNet layer and a physical quantity calculation layer. The physical-driven input layer is used for inputting the tip speed ratio and the pitch angle of the wind turbine generator. The PhyNet layer is used for predicting the wind energy utilization coefficient of the wind turbine generator according to the tip speed ratio and the pitch angle of the wind turbine generator. The physical quantity calculation layer is used for calculating the ultra-short-term wind power theoretical value of the wind turbine generator according to the wind energy utilization coefficient of the wind turbine generator, and converting it into physical feature data meeting the input requirements of the Transformer encoder.

[0061] The expression of the physical quantity calculation layer is:

[0062]

[0063] Wherein: is the ultra-short-term wind power theoretical value of the wind turbine generator, is the air density, is the circular constant, is the blade radius of the wind turbine generator, is the wind speed, is the wind energy utilization coefficient of the wind turbine generator, is the tip speed ratio of the wind turbine generator, A pitch angle of a wind turbine.

[0064] The concatenation layer is used for concatenating the global timing feature data and the physical feature data meeting the input requirements of the Transformer encoder to obtain the final fusion feature data.

[0065] The full connection layer is used for mapping the final fusion feature data into two output scalars, i.e., an ultra-short-term wind power prediction value of the wind turbine and an ultra-short-term wind speed prediction value of the wind turbine.

[0066] S2, based on the historical meteorological timing data and the historical SCADA state data of the wind farm, a training set of the ultra-short-term wind power prediction model is constructed, and the training set of the ultra-short-term wind power prediction model is used to train the ultra-short-term wind power prediction model.

[0067] In an optional embodiment of the present application, the types of meteorological timing data include wind speed, wind direction, temperature, humidity and air pressure; and the types of SCADA state data include rotor speed and pitch angle of the wind turbine.

[0068] The present application constructs a training set of an ultra-short-term wind power prediction model based on historical meteorological timing data and historical SCADA state data of a wind farm, and the specific process is as follows: the historical meteorological timing data and the historical SCADA state data of the wind farm are subjected to minimum-maximum normalization to obtain normalized historical meteorological timing data and historical SCADA state data; the tip speed of the wind turbine is calculated based on the wind speed in the normalized historical meteorological timing data and the rotor speed in the historical SCADA state data; and the training set of the ultra-short-term wind power prediction model is constructed based on the tip speed of the wind turbine, the pitch angle of the wind turbine in the normalized historical SCADA state data, and the normalized historical meteorological timing data.

[0069] The present application trains the ultra-short-term wind power prediction model using the training set of the ultra-short-term wind power prediction model, and the specific process is as follows:

[0070] A data-driven loss, a physical data error loss and a physical consistency loss are constructed, and a loss function of the ultra-short-term wind power prediction model is constructed based on the data-driven loss, the physical data error loss and the physical consistency loss, and the expression is as follows:

[0071]

[0072] wherein: is the loss of the ultra-short-term wind power prediction model, is the data-driven loss, is the weight coefficient of the physical data error loss, For physical data error loss, The weighting coefficients for the physical consistency loss are... This represents a loss of physical consistency.

[0073] The expression for data-driven loss is:

[0074]

[0075] in: This represents the number of samples in the training set. The sample IDs for the training set. For the first The actual value of ultra-short-term wind power of a sample of wind turbine generators. For the first Ultra-short-term wind power forecast values ​​for a sample of wind turbine generators.

[0076] The expression for physical data error loss is:

[0077]

[0078] in: For the first The actual ultra-short-term wind power speed of a sample of wind turbine generators. For the first Ultra-short-term wind turbine rotational speed prediction values ​​for a sample of wind turbine generators.

[0079] The expression for the physical consistency loss is:

[0080]

[0081] in: For rotational inertia, For the first The angular acceleration of a sample of wind turbine generator sets. For the first The theoretical value of ultra-short-term wind power of a sample of wind turbine generators.

[0082] Based on the loss function of the ultra-short-term wind power prediction model, the ultra-short-term wind power prediction model is trained using the training set of the ultra-short-term wind power prediction model.

[0083] S3. Based on the trained ultra-short-term wind power prediction model and the real-time meteorological time series data and real-time SCADA status data of the wind farm, obtain the ultra-short-term wind power prediction results.

[0084] In an optional embodiment of the present application, the present application obtains the tip speed of the wind turbine generator, the pitch angle of the wind turbine generator in the normalized real-time SCADA state data, and the normalized real-time meteorological time series data based on the real-time meteorological time series data and the real-time SCADA state data of the wind farm, and inputs the tip speed of the wind turbine generator, the pitch angle of the wind turbine generator in the normalized real-time SCADA state data, and the normalized real-time meteorological time series data into the trained ultra-short-term wind power prediction model to obtain an ultra-short-term wind power prediction result.

[0085] Simulation experiment:

[0086] In order to verify the effectiveness of the ultra-short-term wind power prediction model proposed in the present application in the offshore wind power prediction task, the LSTM model, the GRU model, the Transformer model, and the model proposed in the present application are respectively selected to be trained under the same data distribution, and a 48-hour prediction comparison with a prediction duration of 15 minutes is performed, and the ultra-short-term wind power prediction results of different models are as shown in Figure 2 The blue curve is the prediction curve of the ultra-short-term wind power prediction model proposed in the present application, and the red curve is the prediction curve of the LSTM model, the green curve is the prediction curve of the GRU model, and the purple curve is the prediction curve of the Transformer model. Figure 2 It can be seen that when the wind power appears a significant fluctuation, that is, a mutation of power climbing or falling, each model can effectively predict the real value of the wind power. However, in the power peak area, it can be seen that the LSTM model, the GRU model, and the Transformer model often exhibit prediction values that exceed the actual power peak value. In comparison, the ultra-short-term wind power prediction model proposed in the present application effectively avoids the extreme overestimation of the actual power, which indicates that the method provided in the present application can help to improve the stability and robustness of the prediction, make the prediction effect more smooth, and reduce the excessive reaction to short-term drastic fluctuations.

[0087] In order to more intuitively compare the prediction effects of different models, the MAE (mean absolute error), the RMSE (root mean square error), and the R 2 The index test results are as shown in Table 1:

[0088]

[0089] As shown in Table 1, the ultra-short-term wind power prediction model proposed in this invention exhibits the best prediction performance. Its MAE and RMSE are the lowest among the compared methods, indicating that its predicted values ​​are closest to the actual power values ​​with the smallest error. Compared to the LSTM model, GRU model, and Transformer model, the prediction accuracy of the model proposed in this paper is significantly improved, with MAE decreasing by 11.06% to 24.18%, RMSE decreasing by 6.69% to 17.08%, and R² increasing by 3.19% to 6.07%. This demonstrates the superiority of the ultra-short-term wind power prediction model proposed in this invention for ultra-short-term wind power prediction.

[0090] To verify the performance of the proposed ultra-short-term wind power prediction model in a data-scarce scenario, this paper sets the training set to a 3:7 ratio with the test set to simulate a data-scarce situation. Under this scenario, LSTM, GRU, Transformer models, and the proposed model are trained and used for prediction. The ultra-short-term wind power prediction results of different models are shown below. Figure 3 As shown in Table 2:

[0091]

[0092] from Figure 3 As can be seen, the predicted curves of the LSTM model, GRU model, and Transformer model differ significantly from the actual power curves, and none of them can fit them well. However, the ultra-short-term wind power prediction model proposed in this invention, due to the introduction of physical constraints, can still fit the predicted curve to a certain extent and maintain higher prediction accuracy even in scenarios with scarce data.

[0093] As shown in Table 2, the proposed model outperforms other models in terms of MAE, RMSE, and R2, demonstrating the effectiveness of the proposed ultra-short-term wind power prediction model in scenarios with scarce data.

[0094] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for ultra-short-term wind power prediction based on hybrid networks, characterized in that, Includes the following steps: An ultra-short-term wind power prediction model is constructed based on a data-driven network branch and a physical-driven network branch. The data-driven network branch integrates a temporal convolutional neural network and a Transformer encoder. The ultra-short-term wind power prediction model includes a data-driven network branch, a physical-driven network branch, a stitching layer, and a fully connected layer. The input of the data-driven network branch serves as the first data input, and the output of the data-driven network branch is connected to the first input port of the stitching layer. The input of the physical-driven network branch serves as the second data input, and the output of the physical-driven network branch is connected to the second input port of the stitching layer. The output of the stitching layer is connected to the input of the fully connected layer, and the output of the fully connected layer serves as the output of the ultra-short-term wind power prediction model. The data-driven network branch consists of a data-driven input layer, a linear adaptation layer, a position encoding layer, a TCN-Transformer module, and a pooling layer, connected in sequence. The data-driven input layer is used to input normalized meteorological time-series data. The linear adaptation layer is used to adjust the feature dimensions of the normalized meteorological time-series data to obtain meteorological time-series data that meets the input requirements of the Transformer encoder. The position encoding layer is used to encode the meteorological time-series data that meets the input requirements of the Transformer encoder to obtain hidden representation sequence data. The TCN-Transformer module consists of three identical TCN-Transformer layers connected in sequence. The TCN-Transformer module is used to process the hidden representation sequence data through a temporal convolutional neural network and a Transformer encoder to obtain multi-layer fused data. The pooling layer is used to perform average pooling on the multi-layer fused data to obtain global temporal feature data. The physical drive network branch consists of a physical drive input layer, a PhyNet layer, and a physical quantity calculation layer connected in sequence. The physical drive input layer is used to input the tip speed ratio and pitch angle of the wind turbine generator. The PhyNet layer predicts the wind energy utilization coefficient of the wind turbine generator based on the tip speed ratio and pitch angle. The physical quantity calculation layer is used to calculate the theoretical value of the ultra-short-term wind power of the wind turbine generator based on the wind energy utilization coefficient of the wind turbine generator and convert it into physical feature data that meets the input requirements of the Transformer encoder. A training set for an ultra-short-term wind power prediction model was constructed based on historical meteorological time-series data and historical SCADA status data of wind farms, and the ultra-short-term wind power prediction model was trained using the training set of the ultra-short-term wind power prediction model. Based on the trained ultra-short-term wind power prediction model, as well as real-time meteorological time-series data and real-time SCADA status data of wind farms, the ultra-short-term wind power prediction results are obtained.

2. The ultra-short-term wind power prediction method based on a hybrid network according to claim 1, characterized in that, The specific data processing procedure of the TCN-Transformer layer is as follows: The hidden representation sequence data is processed by a temporal convolutional neural network to obtain local trend feature data; The hidden representation sequence data and local trend feature data are added together by the residual connection layer to obtain locally enhanced sequence data; The Transformer encoder performs self-attention calculation on the locally enhanced sequence data to output the fused data of the TCN-Transformer layer.

3. The ultra-short-term wind power prediction method based on a hybrid network according to claim 1, characterized in that, The expression for the physical quantity calculation layer is: ； in: This represents the theoretical value of ultra-short-term wind power output of the wind turbine generator set. air density, Pi The radius of the wind turbine blade. For wind speed, The wind energy utilization coefficient of the wind turbine generator set. The tip speed ratio of the wind turbine generator. This refers to the pitch angle of the wind turbine generator set.

4. The ultra-short-term wind power prediction method based on a hybrid network according to claim 1, characterized in that, Meteorological time-series data includes wind speed, wind direction, temperature, humidity, and air pressure; SCADA status data includes rotor speed and pitch angle of wind turbine generators.

5. The ultra-short-term wind power prediction method based on a hybrid network according to claim 4, characterized in that, The training set for constructing an ultra-short-term wind power prediction model is built based on historical meteorological time-series data and historical SCADA status data of wind farms. The specific process is as follows: Minimum-maximum normalization was performed on the historical meteorological time-series data and historical SCADA status data of the wind farm to obtain normalized historical meteorological time-series data and historical SCADA status data. The tip speed of the wind turbine generator is calculated based on the wind speed in the normalized historical meteorological time series data and the rotor speed in the historical SCADA status data. The training set of the ultra-short-term wind power prediction model is constructed based on the tip speed of the wind turbine generator, the pitch angle of the wind turbine generator in the normalized historical SCADA status data, and the normalized historical meteorological time series data.

6. The ultra-short-term wind power prediction method based on hybrid networks according to claim 1, characterized in that, The ultra-short-term wind power prediction model is trained using the training set of the model. The specific process is as follows: We construct data-driven loss, physical data error loss, and physical consistency loss, and based on these, we construct the loss function for the ultra-short-term wind power prediction model, the expression of which is: ； in: For the loss of the ultra-short-term wind power prediction model, For data-driven loss, The weighting coefficients for physical data error loss. For physical data error loss, The weighting coefficients for the physical consistency loss are... This represents a loss of physical consistency. Based on the loss function of the ultra-short-term wind power prediction model, the ultra-short-term wind power prediction model is trained using the training set of the ultra-short-term wind power prediction model.

7. The ultra-short-term wind power prediction method based on a hybrid network according to claim 6, characterized in that, The expression for data-driven loss is: ； in: This represents the number of samples in the training set. The sample IDs for the training set. For the first The actual value of ultra-short-term wind power of a sample of wind turbine generators. For the first Ultra-short-term wind power forecast values ​​for individual samples of wind turbine generator sets; The expression for physical data error loss is: ； in: For the first The actual ultra-short-term wind power speed of a sample of wind turbine generators. For the first Ultra-short-term wind turbine rotation speed prediction values ​​for a sample of wind turbine generator sets; The expression for the physical consistency loss is: ； in: For rotational inertia, For the first The angular acceleration of a sample of wind turbine generator sets. For the first The theoretical value of ultra-short-term wind power of a sample of wind turbine generators.

Images