Wind condition recognition prediction method and device based on double-branch conditional generative adversarial network

Abstract

The application discloses a wind condition identification and prediction method and device based on a double-branch conditional generative adversarial network, and belongs to the technical field of wind turbine wind condition identification and prediction. The method obtains historical wind condition data of a wind turbine and performs time sequence feature extraction to generate conditional input data; the conditional input data and random disturbance are respectively input into two branches to output wind speed prediction values and wind direction prediction values; the prediction values are taken as generated wind condition data, and the historical true values are taken as real wind condition data, which are respectively paired with the conditional input data and then input into discriminators; a loss function is calculated according to a discrimination result, two branches are jointly optimized, and wind condition identification and prediction results are output. The application can realize joint prediction of wind speed and wind direction under complex wind condition, meanwhile, the stability of wind speed prediction and the boundary continuity of wind direction angle are ensured, so that the prediction results are close to real wind condition in terms of numerical accuracy and distribution consistency.

Description

Technical Field

[0001] This invention relates to the field of wind condition identification and prediction technology for wind turbines, specifically to a wind condition identification and prediction method and apparatus based on a dual-branch conditional generative adversarial network. Background Technology

[0002] In existing wind turbine monitoring systems, wind condition identification mostly relies on nacelle anemometers, SCADA operational data, traditional physical models, or simplified statistical methods. These methods are applicable to certain stable wind conditions; however, under complex wind conditions such as complex terrain, wake superposition, terrain obstruction, and rapid wind direction changes, they are prone to problems such as delayed wind condition identification, inaccurate characterization of wind speed fluctuations, and unstable judgment of wind direction trends, thereby affecting the foresight and stability of yaw control strategies.

[0003] Current wind condition forecasting methods primarily employ single regression or physical models. Their core flaw lies in the fact that wind speed and wind direction, as two meteorological variables with distinct physical properties, cannot be simultaneously accounted for by a single model, given the continuous scalar characteristics of wind speed and the cyclic variable characteristics of wind direction. For wind speed forecasting, existing methods directly predict the absolute value of future wind speeds, exhibiting excessive sensitivity at low wind speeds, small fluctuations, and local abrupt changes, resulting in insufficient forecast stability. For wind direction forecasting, existing methods treat the wind direction angle as a simple linear variable, ignoring the physically adjacent cyclic characteristics of 0° and 360°, leading to unreasonable large errors at angle boundaries and failing to accurately identify the incoming flow direction and its stability trend required for yaw control. Furthermore, existing models only focus on the point-value error between predicted and actual values, failing to constrain the consistency between predicted results and actual wind conditions at the distribution level, causing the statistical characteristics of predicted wind conditions to deviate from the actual wind condition distribution.

[0004] Therefore, how to achieve joint prediction of wind speed and wind direction under complex wind conditions and approximate the actual wind conditions has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a wind condition identification and prediction method and device based on a dual-branch conditional generative adversarial network, so as to overcome the problems in the prior art where a single regression model cannot simultaneously take into account the continuous scalar characteristics of wind speed and the cyclic variable characteristics of wind direction, the insufficient stability of wind speed prediction, the amplification of boundary errors of wind direction angle, and the poor consistency of prediction result distribution.

[0006] The present invention solves the above-mentioned technical problems through the following technical solution: This invention provides a wind condition identification and prediction method based on a dual-branch conditional generative adversarial network, comprising the following steps: Acquire historical wind data of wind turbines, extract time series features from historical wind data, and generate conditional input data; A two-branch conditional generative adversarial network (GAN) model is constructed, consisting of a wind speed generation branch and a wind direction generation branch. Conditional input data and random perturbations are input into the wind speed generation branch, which processes the conditional input data and random perturbations and outputs the predicted wind speed value at the future time. Conditional input data and random perturbations are input into the wind direction generation branch, which processes the conditional input data and random perturbations and outputs the predicted wind direction value at the future time. The wind speed and wind direction prediction values ​​are used as generated wind condition data, and the corresponding real values ​​in the historical wind condition data are used as real wind condition data. The generated wind condition data and real wind condition data are paired with the conditional input data and then input into the discriminator. The discriminator determines whether the input sample is real wind condition data or model-generated wind condition data. The loss function is calculated based on the discrimination result of the discriminator. The wind speed generation branch and the wind direction generation branch are jointly optimized based on the loss function, and the wind condition identification and prediction result for future time is output.

[0007] A further improvement of the present invention is that the historical wind data includes at least historical wind speed data and historical wind direction data; Time series feature extraction is performed on historical wind data to generate conditional input data, specifically: A sliding window approach is used to extract time-series features from historical wind data to generate conditional input data. The conditional input data includes historical wind speed sequences, historical wind direction sequences, and relevant operating condition parameters. The historical wind speed sequences, historical wind direction sequences, and relevant operating condition parameters are then concatenated to generate the conditional input data.

[0008] A further improvement of this invention is that the wind speed generation branch processes the conditional input data and random disturbances, and outputs the predicted wind speed value for future times, specifically: The wind speed generation branch adopts a residual prediction structure, which predicts the change in future wind speed relative to the current wind speed based on the conditional input data and random disturbances. The change is added to the current wind speed to obtain the predicted wind speed value at the future time. The residual prediction structure of the wind speed generation branch is represented as follows:

[0009] in, The current wind speed, This represents the predicted future change in wind speed. For the future Predicted wind speed.

[0010] A further improvement of this invention lies in that the wind direction generation branch processes the conditional input data and random disturbances, and outputs the predicted wind direction value for future times, specifically: The wind direction generation branch maps the wind direction angle to a two-dimensional vector on the unit circle using a sine-cosine encoding method. The two-dimensional vector is learned based on the conditional input data and random perturbation. At the output end, the two-dimensional vector is restored to the wind direction angle value through the arctangent function, thus obtaining the wind direction prediction value at future time. The wind direction generation branch maps the wind direction angle to a two-dimensional vector on the unit circle using a sine-cosine encoding method, as follows:

[0011]

[0012] in, For wind direction angle, and These represent the sine and cosine components of the wind direction, respectively; after the wind direction generation branch outputs the prediction vector, the predicted wind direction angle is recovered using the arctangent function:

[0013] in, and These are the sine and cosine components for predicting wind direction, respectively. For the future Wind direction forecast.

[0014] A further improvement of this invention lies in the loss function. Loss due to prediction error Combating losses Distribution consistency constraint loss and engineering constraint losses The weighted composition is as follows:

[0015] in, , , , These are the prediction error losses. Combating losses Distribution consistency constraint loss and engineering constraint losses The weights; Prediction error loss is used to constrain the point value error between the predicted wind speed and the actual wind speed; distribution consistency constraint loss is used to constrain the consistency between the predicted wind speed and the actual wind speed on the cumulative distribution function, Kolmogorov-Smirnov test statistic, and Weibull distribution; engineering constraint loss is used to suppress prediction results with negative wind speed, abnormal peaks, and those that exceed the engineering scope.

[0016] A further improvement of this invention is that the optimization objective of the dual-branch conditional generative adversarial network model is:

[0017] in, Enter data as a condition. For random perturbation, As a sample of real wind conditions, The output of the generator, The discriminator's output is the discrimination of the real sample under the given conditional input. This is the discriminator's output for judging the generated samples under the given input conditions.

[0018] This invention also provides a wind condition identification and prediction system based on a two-branch conditional generative adversarial network, comprising: The first module is used to acquire historical wind condition data of wind turbines, extract time series features from historical wind condition data, and generate conditional input data. The second module is used to construct a two-branch conditional generative adversarial network model. The two-branch conditional generative adversarial network model includes a wind speed generation branch and a wind direction generation branch. The conditional input data and random disturbances are input into the wind speed generation branch, which processes the conditional input data and random disturbances and outputs the predicted wind speed value at the future time. The conditional input data and random disturbances are input into the wind direction generation branch, which processes the conditional input data and random disturbances and outputs the predicted wind direction value at the future time. The third module is used to take the wind speed prediction value and wind direction prediction value as the generated wind condition data, take the corresponding real value in the historical wind condition data as the real wind condition data, and input the generated wind condition data and real wind condition data into the condition input data respectively after pairing them with the condition input data. The discriminator determines whether the input sample is real wind condition data or model generated wind condition data. The fourth module is used to calculate a loss function based on the discrimination result of the discriminator, perform joint optimization of the wind speed generation branch and the wind direction generation branch based on the loss function, and output the wind condition identification and prediction result for future time.

[0019] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the wind condition identification and prediction method based on a two-branch conditional generative adversarial network as described above.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the wind condition identification and prediction method based on a two-branch conditional generative adversarial network as described above.

[0021] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the wind condition identification and prediction method based on a two-branch conditional generative adversarial network as described above.

[0022] Compared with the prior art, the positive and progressive effects of the present invention are as follows: The wind condition recognition and prediction method based on a dual-branch conditional generative adversarial network (GAN) provided by this invention employs a dual-branch GAN architecture, setting independent generation branches for wind speed and wind direction. This allows two meteorological variables with different physical properties to each adopt the most suitable modeling method: the wind speed generation branch models the continuous scalar characteristics, while the wind direction generation branch models the cyclic variable characteristics, avoiding the deficiency of a single model in failing to simultaneously account for the characteristics of both types of variables. By inputting conditional input data and random perturbations into the two generation branches, the adversarial training mechanism of the GAN ensures that the prediction results not only focus on point value errors but also approximate the actual wind conditions at the distribution level, improving the consistency between the prediction results and the actual wind condition distribution. By pairing the generated wind condition data and the actual wind condition data with the conditional input data and then inputting them into the discriminator, the discriminator can distinguish between the actual wind conditions and the generated wind conditions given the same historical conditions, thereby driving the generator to learn the conditional distribution of the actual wind conditions. By calculating the loss function based on the discrimination results and jointly optimizing the two generation branches, the collaborative training of wind speed and wind direction prediction was realized. The final output wind condition recognition and prediction results provide a forward-looking wind condition input for the subsequent yaw control strategy, overcoming the shortcomings of existing technologies such as lagging wind condition recognition, inaccurate characterization of wind speed fluctuations, and unstable judgment of wind direction trends.

[0023] Furthermore, the sliding window can preserve the temporal dependencies of wind data, and the splicing operation integrates multi-source conditional information into a unified input representation, providing a complete conditional input including wind speed, wind direction, and operating conditions for subsequent bi-branch prediction.

[0024] Furthermore, due to the strong persistence of short-term wind speeds, directly predicting the absolute value of future wind speeds is too sensitive to low wind speeds, small fluctuations, and local abrupt changes. The residual prediction structure fully utilizes the persistence of short-term wind speeds, reducing the model's sensitivity to local fluctuations and improving the stability of wind speed predictions.

[0025] Furthermore, since wind direction is a circular variable, 0° and 360° are physically adjacent. Treating it directly as a linear variable would lead to an unreasonable amplification of errors at the boundary. Sine-cosine encoding maps the angle onto a unit circle, making the two ends of the boundary continuous on the unit circle. This fundamentally avoids the problem of amplified boundary errors in the prediction method, making the prediction results more consistent with the wind direction trend recognition requirements in yaw control scenarios.

[0026] Furthermore, the model is specifically defined with a distribution consistency constraint loss to ensure consistency between predicted and actual wind speeds in terms of the cumulative distribution function, Kolmogorov-Smirnov test statistic, and Weibull distribution; and an engineering constraint loss to suppress negative wind speeds, anomalous spikes, and predictions exceeding engineering limits. The weighted combination of these four losses optimizes the model simultaneously in terms of point value accuracy, distribution consistency, and physical plausibility. In particular, the distribution consistency constraint loss ensures that the predicted wind conditions maintain the statistical characteristics of the actual wind conditions, while the engineering constraint loss ensures that the prediction results conform to physical laws.

[0027] Furthermore, by defining the optimization objective of the two-branch conditional generative adversarial network model, and through a minimax game between the generator and the discriminator, the generator learns to generate predictions that more closely approximate the actual wind distribution, while the discriminator learns to better distinguish between the actual and generated wind conditions. This optimization objective provides a theoretical framework for the joint training of the two branches, enabling the two generative branches to collaboratively optimize during adversarial training. Attached Figure Description

[0028] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0029] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the structure of the dual-branch conditional generative adversarial network (CGAN) wind condition recognition and prediction model of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0032] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0033] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0034] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0035] In wind turbine monitoring systems, current wind condition identification largely relies on nacelle anemometers, SCADA data, traditional physical models, or simplified statistical methods. Under complex wind conditions such as terrain variations, wake superposition, terrain obstruction, and rapid wind direction changes, these methods are prone to problems such as delayed wind condition identification, inaccurate characterization of wind speed fluctuations, and unstable wind direction trend judgments. Existing single regression models or physical models struggle to simultaneously account for the continuous scalar characteristics of wind speed and the cyclic variable characteristics of wind direction. Directly predicting the absolute value of wind speed is sensitive to local abrupt changes, treating the wind direction angle as a simple linear variable introduces unreasonable errors at the 0° and 360° boundary, and focusing solely on point value errors leads to deviations in the predicted wind condition statistical characteristics from the actual wind condition distribution.

[0036] Based on the above background, this invention provides a wind condition recognition and prediction method based on a dual-branch conditional generative adversarial network. It generates conditional input data by extracting time-series features from historical wind condition data, constructs wind speed generation branches and wind direction generation branches to output predicted wind speed and wind direction values ​​respectively, and pairs the generated wind condition data with real wind condition data before inputting it into a discriminator for adversarial discrimination. Based on the discrimination results, a loss function is calculated, and the wind speed generation branches and wind direction generation branches are jointly optimized. This enables joint prediction of wind speed and wind direction under complex wind conditions, making the predicted wind conditions approximate the real wind conditions in terms of both point values ​​and distribution.

[0037] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0038] Example 1 like Figure 1 As shown, this embodiment provides a wind condition identification and prediction method based on a dual-branch conditional generative adversarial network, including the following steps: Step S1: Acquire historical wind condition data for the wind turbine, extract time-series features from the historical wind condition data, and generate conditional input data. Historical wind condition data refers to wind-related data collected during the operation of the wind turbine, which may include wind speed data, wind direction data, and operating condition parameters. Extracting time-series features from the historical wind condition data transforms the raw historical wind condition data into a conditional input format that can be used by subsequent generation branches, enabling the wind speed and wind direction generation branches to predict future wind conditions based on the historical wind condition time-series features. The conditional input data contains the time-series feature information of historical wind conditions, which is used to constrain the prediction direction of the generation branches, giving the prediction process a historical basis.

[0039] Step S2: Construct a two-branch conditional generative adversarial network (GAN) model, which includes a wind speed generation branch and a wind direction generation branch. The conditional input data and random perturbations are input into the wind speed generation branch, which processes the input data and perturbations and outputs a predicted wind speed value for the future time. Similarly, the conditional input data and random perturbations are input into the wind direction generation branch, which processes the input data and perturbations and outputs a predicted wind direction value for the future time. The random perturbation can be random noise, used to increase the diversity of the generated results and avoid overly deterministic generation. The wind speed generation branch generates the predicted wind speed value based on the conditional input data and random perturbations, while the wind direction generation branch generates the predicted wind direction value. The wind speed and wind direction generation branches can each use independent generator structures. The two branches share the conditional input data and random perturbation input logic, but perform predictions for wind speed and wind direction separately. By using a dual-branch setup, the continuous scalar characteristics of wind speed and the circular characteristics of wind direction can be processed separately in their respective branches, avoiding mutual interference caused by a single model handling two different types of physical quantities at the same time.

[0040] Step S3: The predicted wind speed and direction are used as generated wind condition data, and the corresponding real values ​​from historical wind condition data are used as real wind condition data. The generated and real wind condition data are paired with the conditional input data and then input into the discriminator. The discriminator determines whether the input sample is real wind condition data or model-generated wind condition data. Here, the real values ​​refer to the historical actual wind speed and actual wind direction values ​​corresponding to the prediction target time. The discriminator receives the combined input of conditional input data and wind condition samples, and outputs a discrimination result to determine whether the input sample is real wind condition data or model-generated wind condition data. Through the adversarial discrimination process, the discriminator can continuously learn the differences between real and generated wind conditions, providing optimization signals for the generation branch.

[0041] Step S4: Calculate the loss function based on the discriminator's judgment result. Then, jointly optimize the wind speed and wind direction generation branches based on the loss function, outputting the wind condition prediction result for future timeframes. The loss function measures the difference between the generated wind condition data and the actual wind condition data, reflecting the discriminator's judgment result on the generated samples. Jointly optimizing the wind speed and wind direction generation branches based on the loss function means using the loss function feedback to adjust the parameters of the two generation branches, gradually bringing the predicted wind speed and wind direction values ​​closer to the actual wind conditions. After optimization, the wind condition prediction result for future timeframes is output. The wind condition prediction result can include predicted wind speed and wind direction values ​​for future timeframes, and can be used in scenarios such as wind turbine yaw control.

[0042] This embodiment generates wind speed and wind direction separately through a dual-branch conditional generative adversarial network, and through adversarial training and joint optimization of the discriminator, the generated wind data simultaneously approximates the real wind speed and wind direction. During the adversarial training process, it gradually approaches the real wind distribution, thereby providing effective support for wind condition identification and prediction under complex wind conditions.

[0043] Example 2 Based on Example 1, this example further explains the composition of historical wind data, the generation method of conditional input data, the specific implementation methods of wind speed generation branch and wind direction generation branch, as well as the loss function and optimization objective.

[0044] In one embodiment of this invention, historical wind data includes at least historical wind speed data and historical wind direction data. Time-series feature extraction is performed on the historical wind data to generate conditional input data. Specifically, a sliding window approach is used to extract time-series features from the historical wind data to generate conditional input data. The conditional input data includes historical wind speed sequences, historical wind direction sequences, and related operating condition parameters. The historical wind speed sequences, historical wind direction sequences, and related operating condition parameters are then concatenated to generate the conditional input data. The sliding window approach divides the original time series into multiple continuous windows, each window corresponding to a set of input features, enabling the model to use wind condition change information from continuous historical periods for future predictions. Related operating condition parameters may include wind turbine operating status parameters, environmental parameters, etc., to provide auxiliary constraints for prediction.

[0045] In one embodiment of this invention, the wind speed generation branch processes the conditional input data and random disturbances to output a predicted wind speed value for future times. Specifically, the wind speed generation branch employs a residual prediction structure, predicts the change in future wind speed relative to the current wind speed based on the conditional input data and random disturbances, adds the change to the current wind speed, and obtains the predicted wind speed value for future times. The residual prediction structure of the wind speed generation branch is represented as follows:

[0046] in, The current wind speed, This represents the predicted future change in wind speed. For the future Step-by-step wind speed prediction. This structure makes full use of the persistence of short-term wind speed, allowing the model to focus on learning the amount of wind speed change rather than the absolute wind speed value. This reduces the model's sensitivity to low wind speeds, small fluctuations, and local abrupt changes, and improves the stability of wind speed prediction.

[0047] In one embodiment of this invention, the wind direction generation branch processes the conditional input data and random perturbations to output the predicted wind direction value for future times. Specifically, the wind direction generation branch maps the wind direction angle to a two-dimensional vector on the unit circle using a sine-cosine encoding method. It then learns from the two-dimensional vector based on the conditional input data and random perturbations. At the output end, it uses an arctangent function to restore the two-dimensional vector to the wind direction angle value, thus obtaining the predicted wind direction value for future times. The wind direction generation branch maps the wind direction angle to a two-dimensional vector on the unit circle using a sine-cosine encoding method, as shown below:

[0048]

[0049] in, For wind direction angle, and These represent the sine and cosine components of the wind direction, respectively; after the wind direction generation branch outputs the prediction vector, the predicted wind direction angle is recovered using the arctangent function:

[0050] in, and These are the sine and cosine components for predicting wind direction, respectively. For the future Wind direction prediction. By using a sine-cosine coding method, the physically adjacent circular feature of 0° and 360° wind direction is preserved, avoiding unreasonable errors at the 0° / 360° boundary when treating wind direction as a normal linear variable.

[0051] In one embodiment of this example, the complete process of calculating the loss function based on the discriminator's discrimination result, jointly optimizing the wind speed generation branch and the wind direction generation branch based on the loss function, and outputting the wind condition identification and prediction result for future time moments is further explained.

[0052] The discriminator receives real wind condition data and generated wind condition data paired with the conditional input data, respectively. For each input sample, it outputs a discrimination score, representing the probability that the discriminator classifies the input sample as real wind condition data. For real wind condition data, the discriminator outputs a real sample discrimination score; for generated wind condition data, it outputs a generated sample discrimination score. The real sample discrimination score and the generated sample discrimination score together constitute the discriminator's discrimination result, serving as the direct basis for subsequent loss function calculation.

[0053] The loss function is calculated based on the discriminator's discrimination results. Loss due to prediction error Combating losses Distribution consistency constraint loss and engineering constraint losses The weighted composition is as follows:

[0054] in, , , , These are the prediction error losses. Combating losses Distribution consistency constraint loss and engineering constraint losses The weighted structure includes: a prediction error loss to constrain the point value error between the predicted and actual wind speeds; a distribution consistency constraint loss to constrain the consistency between the predicted and actual wind speeds on the cumulative distribution function, Kolmogorov-Smirnov test statistic, and Weibull distribution; and an engineering constraint loss to suppress negative wind speeds, abnormal peaks, and predictions exceeding engineering limits. This multi-loss weighted structure ensures the accuracy of point value predictions while constraining the consistency between the predicted and actual wind conditions at the distribution level, preventing the statistical characteristics of the predicted wind conditions from deviating from the actual wind condition distribution.

[0055] Combat losses The results are directly derived from the discriminator's judgment. Substituting the generated sample's discrimination score into the adversarial loss calculation formula, the generator side minimizes the adversarial loss. The goal is to generate samples with a discrimination score close to 1, meaning the discriminator is expected to misclassify the generated wind data as real wind data; the discriminator aims to maximize its resistance to loss. The goal is for the discriminant score of real samples to approach 1 and the discriminant score of generated samples to approach 0, meaning the discriminator is expected to correctly distinguish between real and generated wind condition data. (This is an example of) adversarial loss. The difference in distribution between generated wind data and real wind data in the discriminator's feature space is reflected, and it is the core component connecting the discriminator's discrimination results with the generator's optimization direction.

[0056] Prediction error loss The prediction wind speed is calculated based on the output of the wind speed generation branch and the corresponding actual wind speed values ​​in historical wind data. The absolute error between the predicted wind speed at each future time and the corresponding actual wind speed is calculated point-by-point. Then, the mean is calculated over all prediction steps and all training samples to obtain the prediction error loss. Prediction error loss Used to constrain the point value error between the predicted wind speed value and the actual wind speed value.

[0057] Distribution Consistency Constraint Loss The calculation is based on the statistical distributions of predicted and actual wind speeds. The cumulative distribution functions (CDFs) of the predicted and actual wind speed sequences are calculated separately, and the difference between the two CDFs is taken as the CDF consistency term. The Kolmogorov-Smirnov test statistic between the predicted and actual wind speeds is calculated as the distribution difference measure. Weibull distributions are fitted to the predicted and actual wind speeds respectively, and the parameter difference or distribution fitting error between the two Weibull distributions is taken as the Weibull distribution consistency term. The above terms are then weighted and combined to obtain the distribution consistency constraint loss. Distribution consistency constraint loss This is used to constrain the consistency between predicted and actual wind speeds on the cumulative distribution function, Kolmogorov-Smirnov test statistic, and Weibull distribution.

[0058] Engineering constraint loss Engineering rationality calculation based on prediction results. A penalty term is applied to the portion of the predicted wind speed less than zero to suppress negative wind speeds; a penalty term is applied to the portion of the predicted wind speed exceeding the preset upper limit for engineering conditions to suppress predictions exceeding the engineering scope; a penalty term is applied to the portion of the predicted wind speed where the change in wind speed between adjacent time points exceeds a preset peak threshold to suppress abnormal peaks. Combining these factors yields the engineering constraint loss. Engineering constraint losses It is used to suppress negative wind speeds, abnormal peaks, and prediction results that exceed the engineering scope.

[0059] Obtain the total loss function Then, based on the total loss function Joint optimization is performed on the wind speed and wind direction generation branches. This joint optimization is achieved through the backpropagation algorithm: the total loss function is... The gradients of the network parameters for the wind speed generation branch and the wind direction generation branch are calculated separately, and the network parameters of both branches are updated simultaneously using the gradient descent algorithm. (Due to adversarial loss...) It also involves the wind speed prediction output from the wind speed generation branch and the wind direction prediction output from the wind direction generation branch, with the total loss function... The gradient can be backpropagated to both branches simultaneously, enabling the two branches to coordinately adjust parameters under a unified optimization objective, thereby achieving joint optimization of the wind speed generation branch and the wind direction generation branch.

[0060] While jointly optimizing the two generative branches, the discriminator's network parameters are updated via backpropagation based on the discriminator's discrimination results. The discriminator's optimization objective is to distinguish between real and generated wind data as much as possible, and its parameter update direction forms an adversarial game with the generator's parameter update direction. The generator and discriminator iteratively optimize alternately until the dual-branch conditional generative adversarial network model converges.

[0061] The optimization objective of the two-branch conditional generative adversarial network model is:

[0062] in, Enter data as a condition. For random perturbation, As a sample of real wind conditions, The output of the generator, The discriminator's output is the discrimination of the real sample under the given conditional input. This represents the discriminator's output based on the generated samples under conditional input. The optimization objective is to make the discriminator distinguish between real and generated samples as much as possible, while simultaneously enabling the generator to generate samples that can be distinguished by the discriminator. This allows the generated wind conditions to gradually approach the real wind condition distribution during adversarial training, ensuring that the generated results closely approximate the real wind condition distribution under conditional constraints. Through iterative optimization of the generator and discriminator, the two-branch conditional generative adversarial network model gradually converges.

[0063] After the model converges, the wind speed and wind direction generation branches, trained using the conditional input data and random disturbance inputs at the time to be predicted, are used. The wind speed generation branch outputs the predicted wind speed value for the future time, and the wind direction generation branch outputs the predicted wind direction value for the future time. The wind speed and wind direction prediction values ​​are combined to output the wind condition identification and prediction result for the future time. This wind condition identification and prediction result contains joint prediction information of short-term wind speed and wind direction, providing a wind condition basis for the forward-looking and stability design of wind turbine yaw control strategies.

[0064] Optionally, the prediction error of the wind direction generation branch can be calculated using the definition of circular error to correctly handle the error at the boundary between 0° and 360°. The circular error is expressed as:

[0065] in, To predict wind direction, To reflect the actual wind direction, This refers to the circumferential wind direction error. This error definition can correctly handle the 0° / 360° boundary problem. For example, when the actual wind direction is 359° and the predicted wind direction is 1°, the physical error is 2° instead of 358°, which is more in line with the actual requirements of yaw control for wind direction deviation.

[0066] Optionally, the loss function for the wind direction generation branch may include a circular wind direction error loss, used to constrain the circular error between the predicted and actual wind direction values. This circular wind direction error loss can be used in conjunction with adversarial loss and distribution constraints, and the specific weights can be set according to actual needs.

[0067] like Figure 2As shown, the dual-branch conditional generative adversarial network model in this embodiment can generally include a conditional input module, a wind speed generation branch, a wind direction generation branch, a discriminator module, and a prediction result evaluation module. The conditional input module generates conditional input data; the wind speed generation branch outputs predicted wind speed values ​​for future times; the wind direction generation branch outputs predicted wind direction values ​​for future times; the discriminator module distinguishes whether the input sample is real wind data or model-generated wind data; and the prediction result evaluation module calculates the loss function based on the discrimination result and optimizes the generation branch.

[0068] Example 3 This embodiment provides a wind condition identification and prediction system based on a two-branch conditional generative adversarial network. This system can be deployed in a wind turbine monitoring system or a SCADA system. The system includes a first module, a second module, a third module, and a fourth module.

[0069] The first module acquires historical wind data from wind turbines, extracts time-series features from the historical wind data, and generates conditional input data. The second module constructs a two-branch conditional generative adversarial network (GAN) model, consisting of a wind speed generation branch and a wind direction generation branch. Conditional input data and random disturbances are input to the wind speed generation branch, which processes the conditional input data and random disturbances, outputting a predicted wind speed value for the future time. Similarly, conditional input data and random disturbances are input to the wind direction generation branch, which processes the conditional input data and random disturbances, outputting a predicted wind direction value for the future time. The third module uses the predicted wind speed and wind direction values ​​as generated wind data and the corresponding real values ​​from historical wind data as real wind data. The generated and real wind data are then paired with the conditional input data and input into a discriminator, which determines whether the input sample is real wind data or model-generated wind data. The fourth module is used to calculate the loss function based on the discriminator's discrimination result, and to jointly optimize the wind speed generation branch and the wind direction generation branch based on the loss function, and output the wind condition identification and prediction results for future time.

[0070] The coordination relationship between the modules is as follows: the first module provides conditional input data to the second module; the wind speed prediction value and wind direction prediction value generated by the second module are output to the third module; the third module pairs the generated wind condition data and the actual wind condition data and inputs them into the discriminator, and provides the discrimination result to the fourth module; the fourth module calculates the loss function based on the discrimination result, and optimizes the wind speed generation branch and wind direction generation branch in the second module based on the loss function.

[0071] In a complete working process of this embodiment, historical wind condition data of the wind turbine is first acquired. A sliding window method is used to extract historical wind speed sequences, historical wind direction sequences, and related operating condition parameters. These sequences are then concatenated to generate conditional input data. The conditional input data and random disturbances are input into the wind speed generation branch and wind direction generation branch, respectively. The wind speed generation branch uses a residual prediction structure to output predicted wind speed values ​​for future times; the wind direction generation branch uses sine-cosine encoding to output predicted wind direction values ​​for future times. The predicted wind speed and wind direction values ​​are used as generated wind condition data, and the corresponding real values ​​from the historical wind condition data are used as real wind condition data. The generated and real wind condition data are then paired with the conditional input data and input into a discriminator. After the discriminator outputs the discrimination result, a loss function is calculated based on the result. For example, the loss function is a weighted sum of prediction error loss, adversarial loss, distribution consistency constraint loss, and engineering constraint loss. The wind speed generation branch and wind direction generation branch are then jointly optimized based on the loss function. After training, real-time historical wind data is input into the model, which outputs predicted short-term wind speed and direction. The predicted wind speed is used to determine whether the current wind conditions meet the effective yaw conditions, and the predicted wind direction is used to determine the future direction of the incoming flow and its stability. Together, they constitute the wind condition prediction input for short-term wind speed, wind direction, and stability, which is used for yaw dead zone threshold, hysteresis constraint, stability window, and motion limit design.

[0072] Based on the same inventive concept, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a wind condition identification and prediction method based on a two-branch conditional generative adversarial network. The memory may include main memory, such as high-speed random access memory, or it may also include non-volatile memory, such as at least one disk storage device. The processor, network interface, and memory are interconnected via an internal bus, which may be an industry-standard architecture bus, a peripheral component interconnection standard bus, an extended industry-standard architecture bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory stores the program; specifically, the program may include program code, which includes computer operation instructions. The memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0073] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the wind condition identification and prediction method based on a two-branch conditional generative adversarial network. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include RAM (Random Access Memory) and / or cache memory, etc. The non-volatile memory may include ROM (Read-Only Memory), hard disk, flash memory, optical disk, magnetic disk, etc.

[0074] Based on the same inventive concept, this application provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer device, cause the computer device to perform the steps of the above-described wind condition identification and prediction method based on a two-branch conditional generative adversarial network.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM (Compact Disc Read-Only Memory), optical storage, etc.) containing computer-usable program code.

[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer apparatus or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer device or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer device or other programmable data processing equipment to cause a series of operational steps to be performed on the computer device or other programmable equipment to produce a process implemented by the computer device, thereby providing instructions that execute on the computer device or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0080] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A wind condition identification and prediction method based on a dual-branch conditional generative adversarial network, characterized in that, Includes the following steps: Acquire historical wind data of wind turbines, extract time series features from historical wind data, and generate conditional input data; A two-branch conditional generative adversarial network model is constructed, which includes a wind speed generation branch and a wind direction generation branch. The conditional input data and random disturbances are input into the wind speed generation branch, which processes the conditional input data and random disturbances and outputs the wind speed prediction value at future time. The conditional input data and random disturbances are input into the wind direction generation branch. The wind direction generation branch processes the conditional input data and random disturbances and outputs the wind direction prediction value for future time. The wind speed and wind direction prediction values ​​are used as generated wind condition data, and the corresponding real values ​​in the historical wind condition data are used as real wind condition data. The generated wind condition data and real wind condition data are paired with the conditional input data and then input into the discriminator. The discriminator determines whether the input sample is real wind condition data or model-generated wind condition data. The loss function is calculated based on the discrimination result of the discriminator. The wind speed generation branch and the wind direction generation branch are jointly optimized based on the loss function, and the wind condition identification and prediction result for future time is output.

2. The wind condition identification and prediction method based on a dual-branch conditional generative adversarial network according to claim 1, characterized in that, Historical wind data should include at least historical wind speed data and historical wind direction data; Time series feature extraction is performed on historical wind data to generate conditional input data, specifically: A sliding window approach is used to extract time-series features from historical wind data to generate conditional input data. The conditional input data includes historical wind speed sequences, historical wind direction sequences, and relevant operating condition parameters. The historical wind speed sequences, historical wind direction sequences, and relevant operating condition parameters are then concatenated to generate the conditional input data.

3. The wind condition identification and prediction method based on a dual-branch conditional generative adversarial network according to claim 1, characterized in that, The wind speed generation branch processes the conditional input data and random disturbances, and outputs the predicted wind speed values ​​for future times, specifically: The wind speed generation branch adopts a residual prediction structure, which predicts the change in future wind speed relative to the current wind speed based on the conditional input data and random disturbances. The change is added to the current wind speed to obtain the predicted wind speed value at the future time. The residual prediction structure of the wind speed generation branch is represented as follows: in, The current wind speed, This represents the predicted future change in wind speed. For the future Predicted wind speed.

4. The wind condition identification and prediction method based on a dual-branch conditional generative adversarial network according to claim 1, characterized in that, The wind direction generation branch processes the conditional input data and random disturbances, and outputs the predicted wind direction value for future time moments, specifically: The wind direction generation branch maps the wind direction angle to a two-dimensional vector on the unit circle using a sine-cosine encoding method. The two-dimensional vector is learned based on the conditional input data and random perturbation. At the output end, the two-dimensional vector is restored to the wind direction angle value through the arctangent function, thus obtaining the wind direction prediction value at future time. The wind direction generation branch maps the wind direction angle to a two-dimensional vector on the unit circle using a sine-cosine encoding method, as follows: in, For wind direction angle, and These represent the sine and cosine components of the wind direction, respectively; after the wind direction generation branch outputs the prediction vector, the predicted wind direction angle is recovered using the arctangent function: in, and These are the sine and cosine components for predicting wind direction, respectively. For the future Wind direction forecast.

5. The wind condition identification and prediction method based on a dual-branch conditional generative adversarial network according to claim 1, characterized in that, loss function Loss due to prediction error Combating losses Distribution consistency constraint loss and engineering constraint losses The weighted composition is as follows: in, , , , These are the prediction error losses. Combating losses Distribution consistency constraint loss and engineering constraint losses The weights; Prediction error loss is used to constrain the point value error between the predicted wind speed and the actual wind speed; distribution consistency constraint loss is used to constrain the consistency between the predicted wind speed and the actual wind speed on the cumulative distribution function, Kolmogorov-Smirnov test statistic, and Weibull distribution; engineering constraint loss is used to suppress prediction results with negative wind speed, abnormal peaks, and those that exceed the engineering scope.

6. The wind condition identification and prediction method based on a dual-branch conditional generative adversarial network according to claim 1, characterized in that, The optimization objective of the two-branch conditional generative adversarial network model is: in, Enter data as a condition. For random perturbation, As a sample of real wind conditions, The output of the generator, The discriminator's output is the discrimination of the real sample under the given conditional input. This is the discriminator's output for judging the generated samples under the given conditions.

7. A wind condition identification and prediction system based on a dual-branch conditional generative adversarial network, characterized in that, include: The first module is used to acquire historical wind condition data of wind turbines, extract time series features from the historical wind condition data, and generate conditional input data. The second module is used to construct a two-branch conditional generative adversarial network model. The two-branch conditional generative adversarial network model includes a wind speed generation branch and a wind direction generation branch. The conditional input data and random disturbances are input into the wind speed generation branch, which processes the conditional input data and random disturbances and outputs the wind speed prediction value for future time. The conditional input data and random disturbances are input into the wind direction generation branch. The wind direction generation branch processes the conditional input data and random disturbances and outputs the wind direction prediction value for future time. The third module is used to take the wind speed prediction value and wind direction prediction value as the generated wind condition data, take the corresponding real value in the historical wind condition data as the real wind condition data, and input the generated wind condition data and real wind condition data into the condition input data respectively after pairing them with the condition input data. The discriminator determines whether the input sample is real wind condition data or model generated wind condition data. The fourth module is used to calculate a loss function based on the discrimination result of the discriminator, perform joint optimization of the wind speed generation branch and the wind direction generation branch based on the loss function, and output the wind condition identification and prediction result for future time.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the wind condition identification and prediction method based on a dual-branch conditional generative adversarial network as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the wind condition identification and prediction method based on a two-branch conditional generative adversarial network as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the wind condition identification and prediction method based on a two-branch conditional generative adversarial network as described in any one of claims 1 to 6.

Images