A context-aware wind turbine power curve reconstruction method based on generative models

Abstract

This invention discloses a context-aware wind turbine power curve reconstruction method based on a generative model, relating to the fields of artificial intelligence and wind power generation. The reconstruction method includes the following steps: S1, acquiring the original operating data stream of the wind turbine; S2, extracting temporal features from the original operating data stream to obtain a high-dimensional temporal feature tensor; S3, generating a context embedding vector representing the current operating state; S4, generating a reconstructed power value based on the context embedding vector and a preset noise vector using a condition generator; S5, applying physical consistency constraints to the reconstructed power value; S6, performing post-processing verification on the power sequence, and outputting the reconstructed power data if the verification passes. By constructing a feature extraction and encoding module that integrates multi-source context information, the context-aware mechanism of this invention enables the reconstruction results to better adapt to the power response characteristics of the wind turbine under complex operating conditions, thereby improving the accuracy and dynamic adaptability of the reconstruction.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and wind power generation technology, and in particular to a context-aware wind turbine power curve reconstruction method based on generative models. Background Technology

[0002] Against the backdrop of the large-scale development of renewable energy, wind power has become a crucial pillar of the clean energy system. The wind turbine power curve, as a core performance indicator characterizing the mapping relationship between wind speed and active power, is not only a key basis for the factory acceptance and on-site verification of wind turbine units, but also a fundamental data source for short-term power forecasting, equipment health status assessment, and early fault warning. Therefore, ensuring the integrity, accuracy, and physical consistency of power curve data is a prerequisite for improving the refined management level of wind farms.

[0003] Chinese invention patent CN114065618A discloses a method for fitting wind turbine power curves using a parametric model based on differential evolution. It employs a five-parameter logistic function to construct a prior analytical model and combines it with a differential evolution algorithm to globally optimize cleaned effective samples, thereby fitting a smooth curve that approximates the theoretical power characteristics. This method exhibits good convergence and computational efficiency under standard operating conditions, and effectively suppresses overfitting risk through structured function constraints. Chinese invention patent CN114968991A discloses a wind power data recovery method based on a combination of change-point grouping, two-step quantiles, and DCNN. This method attempts to integrate change-point detection, quantile regression, and deep convolutional neural networks to construct a multi-stage anomaly identification and missing data imputation framework. This scheme initially screens reliable data subsets using statistical methods and then utilizes a deep convolutional architecture to capture local nonlinear features, significantly improving its adaptability to complex noise patterns and demonstrating superior recovery accuracy compared to traditional regression models in cases of small to medium-scale missing data.

[0004] However, as wind power systems evolve towards high-density deployment, adaptation to complex terrain, and intelligent operation and maintenance throughout their entire lifecycle, the two mainstream methods mentioned above have gradually revealed inherent contradictions at the principle level. Parametric model fitting methods, limited by the expressive power of predefined function forms, cannot fully characterize the multimodal, non-monotonic, and even hysteretic power response characteristics of wind turbines under non-steady-state conditions. Furthermore, they completely ignore the operational context information in the time dimension, leading to a sharp decline in model generalization ability in sparse data regions or under conditions with a high proportion of anomalies. While deep learning methods possess powerful feature extraction capabilities, they generally follow a discriminative modeling paradigm, learning only a deterministic mapping from input to output. They lack the ability to learn and generate the data distribution itself, and are prone to problems such as oversmoothing, physical distortion, or mode collapse when faced with large segments of continuous missing data or extremely rare operating conditions.

[0005] Therefore, it is necessary to improve upon the shortcomings of existing technologies in order to solve the above problems. Summary of the Invention

[0006] This invention overcomes the shortcomings of the prior art and provides a context-aware wind turbine power curve reconstruction method based on a generative model.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: This invention provides a context-aware wind turbine power curve reconstruction method based on a generative model, comprising the following steps: S1. Obtain the raw operating data stream of the wind turbine, which includes the three-dimensional wind speed vector at the hub height, active power output value, ambient temperature, atmospheric pressure, relative humidity, blade pitch angle, yaw angle, generator speed, pitch motor current, tower vibration acceleration and geographical coordinates. S2. Temporal features are extracted from the original running data stream through the context feature extraction module to obtain a high-dimensional temporal feature tensor. The context feature extraction module includes a multi-level temporal convolutional unit, and each level of temporal convolutional unit includes a one-dimensional causal convolutional layer, a batch normalization layer, and an activation function. S3. Input the high-dimensional temporal feature tensor into the context encoder to generate a context embedding vector representing the current running state; the context encoder adopts a bidirectional gated cyclic unit structure; S4. During the time period when power data is missing, a reconstructed power value is generated by a condition generator based on the context embedding vector and a preset noise vector; the condition generator is a multi-layer fully connected neural network. S5. Apply physical consistency constraints to the reconstructed power value, including constraining the power coefficient to not exceed the relevant threshold of the Bates limit and constraining the power change rate to not exceed the maximum allowable slope of the wind turbine. S6. Perform post-processing verification on the generated power sequence. After the verification passes, output the reconstructed power data.

[0008] In a preferred embodiment of the present invention, in step S2, the context feature extraction module includes four cascaded temporal convolutional units. The number of input channels of each temporal convolutional unit is the original variable dimension, 64, 128, and 256, respectively, and the number of output channels is 128, 256, and 512, respectively. The convolutional kernel size is 7, the stride is 1, and causal padding is used.

[0009] In a preferred embodiment of the present invention, in step S4, the condition generator includes a four-layer fully connected neural network with the number of neurons in each layer being 512, 256, 128, and 1 respectively, and all layers except the last layer use exponential linear units as activation functions; the input of the condition generator is the concatenation vector of the context embedding vector and the 64-dimensional noise vector.

[0010] In a preferred embodiment of the present invention, in step S4, if the length of the power missing segment does not exceed 20 sampling periods, the context embedding vector is directly used to drive the condition generator; if the length of the missing segment exceeds 20 sampling periods, the context embedding vector is extracted at both ends of the missing segment, and linear interpolation is performed in its embedding space to obtain the intermediate context state sequence, and then the condition generator is driven point by point.

[0011] In a preferred embodiment of the present invention, in step S5, the power coefficient threshold of the physical consistency constraint is dynamically set according to the rated power coefficient corresponding to the wind turbine model; the maximum allowable power slope is determined according to the wind turbine rotational inertia, transmission system damping and converter response capability.

[0012] In a preferred embodiment of the present invention, in step S6, the post-processing verification includes: Verify whether the power value is between the upper and lower envelopes of the theoretical power curve, which are obtained by fitting historical normal data based on a five-parameter logistic function; Verify whether the spectral characteristics of the power sequence are consistent with the typical spectrum under similar operating conditions; Verify the smooth transition between the power sequence and neighboring normal data points.

[0013] In a preferred embodiment of the present invention, it further includes: During the training phase, an adversarial training mechanism is adopted, in which a discriminator distinguishes between the generated power value and the actual power value, and joint optimization is performed in combination with the physical consistency constraint loss. During the inference phase, the noise vector is set to zero to make the generated result deterministic.

[0014] In a preferred embodiment of the present invention, an incremental learning step is also included: when new high-quality operating data is added, the standardized parameters of the context feature extraction module are updated, and the condition generator is fine-tuned using an elastic weight solidification algorithm to adapt to the drift of wind turbine operating characteristics.

[0015] In a preferred embodiment of the present invention, the invention further includes an output reconstruction confidence index, which is calculated based on the discriminant score of the generated sample and is used to identify the reliability of the reconstruction result.

[0016] In a preferred embodiment of the present invention, the original operating data stream is accessed in real time and transmitted via the Industrial Ethernet protocol; the reconstructed power data is verified and then transmitted back to the wind farm monitoring system.

[0017] This invention addresses the shortcomings of the prior art and has the following beneficial effects: (1) This invention provides a context-aware wind turbine power curve reconstruction method based on generative models. It constructs a feature extraction and encoding module that integrates multi-source context information to fully capture multi-dimensional information during wind turbine operation. Through the combination of temporal convolutional units and bidirectional gated recurrent units, it effectively mines the long-term and short-term dependencies in the time dimension, realizes deep modeling of the evolution trend of historical operating conditions, and the generated context embedding vector can accurately represent the comprehensive operating state of the wind turbine at different times. Compared with the shortcomings of existing parameterization methods that completely ignore context information and thus have insufficient generalization ability, the context-aware mechanism of this invention enables the reconstruction results to better adapt to the power response characteristics of the wind turbine under complex operating conditions, thereby improving the accuracy and dynamic adaptability of reconstruction.

[0018] (2) In this invention, a conditional generative adversarial network is used as the core reconstruction model. Through adversarial training between the generator and the discriminator, the generator can learn the distribution characteristics of real power data and generate power response trajectories that conform to physical laws and have reasonable diversity. The discriminator continuously optimizes the generator's generation ability by distinguishing between generated samples and real samples, and can achieve high-fidelity reconstruction of missing or abnormal power data. Compared with the existing discriminative deep learning methods, which are prone to over-smoothing or mode collapse due to lack of generation ability, this invention can restore the inherent random fluctuation characteristics in power response, thereby providing a fundamental guarantee for outputting high-fidelity and physically real power trajectories, and effectively enhancing the realism and credibility of the reconstruction results.

[0019] (3) The present invention introduces a physical consistency constraint mechanism. By constraining the power coefficient and the power time derivative, it ensures that the generated power data strictly follows the aerodynamic laws and the operating characteristics of the wind turbine. The power coefficient constraint prevents the generated power value from exceeding the theoretical limit, while the power time derivative constraint ensures the smoothness and rationality of power changes. This effectively avoids physical distortion or unreasonable sudden changes in the reconstruction results, making the reconstruction results not only numerically accurate but also physically reasonable, thus providing a high-quality data foundation for subsequent power prediction, state assessment and other applications.

[0020] (4) In this invention, by integrating missing pattern perception and adaptive reconstruction strategy, the length of missing segments can be automatically identified. For short missing segments and long missing segments, two different strategies are adopted: local context-driven and context state transition interpolation. The model can adaptively adjust the generation logic according to the severity of data missing, balance local detail restoration and long-term state coherence, and thus effectively handle various missing scenarios from instantaneous anomalies to long-term communication interruptions. It can maintain the smooth transition of the reconstruction power sequence in time and the natural evolution of the state, thereby effectively solving the problem of context state drift in the reconstruction process of long missing segments and improving the robustness of the model in complex missing scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a flowchart illustrating a preferred embodiment of the present invention of a context-aware wind turbine power curve reconstruction method based on a generative model. Figure 2 This is a schematic diagram of the data flow and post-processing verification process in the inference stage of the reconstruction method of the present invention. Detailed Implementation

[0022] 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, and 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.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0024] like Figure 1 As shown, a context-aware wind turbine power curve reconstruction method based on a generative model includes the following steps: S1. Obtain the raw operating data stream of the wind turbine; S2. Temporal features are extracted from the original running data stream through the context feature extraction module to obtain a high-dimensional temporal feature tensor; S3. Input the high-dimensional temporal feature tensor into the context encoder to generate a context embedding vector representing the current running state; S4. During the time period when power data is missing, a reconstructed power value is generated by a condition generator based on the context embedding vector and a preset noise vector; the condition generator is a multi-layer fully connected neural network. S5. Apply physical consistency constraints to the reconstructed power value, including constraining the power coefficient to not exceed the relevant threshold of the Bates limit and constraining the power change rate to not exceed the maximum allowable slope of the wind turbine. S6. Perform post-processing verification on the generated power sequence. After the verification passes, output the reconstructed power data.

[0025] In some specific implementations, in step S1, the raw operational data stream serves as the input source for the entire system, including the three-dimensional wind speed vector at hub height, active power output value, ambient temperature, atmospheric pressure, relative humidity, blade pitch angle, yaw angle, generator speed, pitch motor current, tower vibration acceleration, and geographic location coordinates provided by the Global Navigation Satellite System receiver. All variables are synchronously acquired at a fixed sampling frequency, with a typical sampling period of 10 s, forming a time-aligned multidimensional time series.

[0026] It should be noted that before entering the context feature extraction module, each variable is independently standardized. Its mean and standard deviation are determined by the statistical moments of the corresponding variable in the historical operation database of the wind farm over a time span of at least one year, and are updated regularly to adapt to distribution drift caused by seasonal changes or equipment aging.

[0027] In some specific implementations, in step S2, the context feature extraction module consists of four cascaded temporal convolutional units. Each temporal convolutional unit includes a one-dimensional causal convolutional layer, a batch normalization layer, and a modified linear unit activation function. The kernel size of the causal convolutional layer is 7, the stride is 1, and the padding method ensures that the length of the output sequence is consistent with the input and does not introduce future information leakage. The number of input channels of the first-level temporal convolutional unit is equal to the dimension of the original variable. The number of input channels of subsequent levels are 64, 128, and 256 respectively, and the number of output channels are 128, 256, and 512 respectively.

[0028] Furthermore, by stacking a four-level structure, the receptive field of this module covers 360 consecutive sampling periods, i.e., a total of 1 hour of historical data window, which is sufficient to capture multi-scale dynamic features from transient gusts to diurnal meteorological evolution. All convolutional operations are performed in the temporal dimension, while the spatial dimension, i.e., the variable dimension, is processed independently to preserve the semantic independence of each physical quantity. The output of the context feature extraction module is a high-dimensional temporal feature tensor with the same time step as the original input, and each time step corresponds to a 512-dimensional feature vector.

[0029] In some specific implementations, in step S3, the high-dimensional temporal feature tensor is fed into the context encoder, which adopts a bidirectional gated recurrent unit structure, with each of the forward and backward paths containing a gated recurrent unit layer with a hidden state dimension of 256.

[0030] Among them, in time step t At this point, the forward path outputs the hidden state. The back path outputs the hidden state. The two are concatenated along the feature dimension to form a 512-dimensional joint state, which is then processed by a weight matrix. and bias vector Perform linear projection to generate a 128-dimensional context embedding vector. This vector As a compact representation of the overall operating status of the wind turbine at the current moment, it integrates information from multiple sources such as meteorological input, mechanical response, control commands and external disturbances, and has the ability to remember the evolution trend of historical operating conditions for a long time.

[0031] In some specific implementation methods, such as Figure 2 As shown, in step S4, the condition generator module uses a random noise vector z With context embedding vector For joint input. During the training phase, z It is a 64-dimensional standard normally distributed random vector; during the inference phase, z It is fixed as a zero vector to ensure output determinism.

[0032] It should be noted that the condition generator consists of a four-layer fully connected neural network, with the input layer receiving... z and The concatenated vector has a total dimension of 192. The first hidden layer contains 512 neurons, the second hidden layer contains 256 neurons, the third hidden layer contains 128 neurons, and the output layer is a single neuron that directly outputs the reconstructed power value. Except for the output layer, all hidden layers use the exponential linear unit activation function, which is mathematically expressed as: ; Where α takes the value 1.0; e It is the base of the natural logarithm. The output layer has no activation function and directly outputs the linear combination result to ensure the unboundedness of the power value, which conforms to the actual physical dimensions.

[0033] Furthermore, the discriminator module receives a pair of inputs: ,in P The power sample can be either a real power sample or a generated power sample. The discriminator consists of a three-layer fully connected neural network. The input layer has a 129-dimensional dimension, the first hidden layer contains 256 neurons, the second hidden layer contains 128 neurons, and the output layer is a single neuron. The first two layers use a leak-corrected linear unit activation function with a negative slope coefficient of 0.2; the output layer uses the sigmoid function, and the output value... This represents the probability that the input sample is judged as real. The discriminator and the condition generator constitute a conditional generative adversarial network, whose adversarial loss function is defined as: ; in, G Represents a condition generator; D Indicates the discriminator; EIndicates the expected value; This represents the true power distribution; It follows a standard normal distribution.

[0034] In some specific implementations, in step S5, the physical consistency constraint monitors the physical rationality of the generated results in real time during the training process. First, the generation power is calculated. Corresponding wind speed The power coefficient between (hub height, wind speed modulus, and length from the original input) : ; in, ρ The air density is calculated from the ambient temperature, atmospheric pressure, and relative humidity using the ideal gas law and the water vapor pressure formula. A The swept area of ​​the wind turbine is determined by the blade length.

[0035] Theoretical Bates Limit The maximum value is 16 / 27 ≈ 0.593. This invention sets a dynamic threshold. ,in, This is the design power factor for this fan model under rated operating conditions. If > Then a quadratic penalty term is introduced into the total loss function. , This is the regularization coefficient, initially set to 10.0.

[0036] In addition, physical consistency constraints also monitor the power time derivative. Let the sampling period be Δ. t The power change rate between adjacent time steps is... The inertia of the wind turbine determines the maximum permissible power slope. Its value is determined by the fan's moment of inertia, the damping of the transmission system, and the converter's response capability, with a typical value of 0.15 pu / s (based on rated power).

[0037] like Then a penalty term is introduced. , The initial value is set to 5.0. The two penalties together constitute the physical consistency loss: ; The total training loss is the sum of the adversarial loss and the physical consistency loss: ;in β The dynamic weighting coefficient is set to 0.1 at the beginning of training and increases linearly to 1.0 with each training round to achieve a smooth transition from prioritizing generation quality to prioritizing physical plausibility.

[0038] Furthermore, the missing pattern awareness mechanism is initiated during the data preprocessing stage. The system first scans the power field in the input data stream, marking all missing points. It then identifies consecutive missing segments through connected component analysis and records the start time index of each segment. End Time Index and length .like L If the value is ≤20, it is considered a short missing segment; if L If the value is greater than 20, it is considered a long missing segment. For short missing segments, the context feature extraction module extracts local temporal features using a sliding window (360 points in length) that includes the missing segment. The context encoder extracts features at each time step within the missing segment. t Output the corresponding Condition generator with and fixed z =0 generated For long missing segments, the system only... and Calculate the context embedding vector at the location and Then, linear interpolation is performed on the two in a 128-dimensional Euclidean space: ; Generated intermediate context sequence A point-by-point driven condition generator is used to reconstruct long missing segments. This strategy avoids state drift caused by the accumulation of recursive errors in the context encoder under long periods of no observation.

[0039] In some specific implementations, in step S6, the post-processing verification performs a triple verification before generating the power sequence output. The first verification is based on the envelope of the theoretical power curve. This envelope is obtained by fitting historical normal data using a five-parameter logistic function, and its form is: ; in, P max , k , v 0 , a , b To fit the parameters, the upper and lower envelopes are obtained by offsetting them at the 95th and 5th quantiles of the residual distribution, respectively. If any Exceeding [ v t If the interval corresponding to the upper and lower envelopes is specified, then the check is marked as failed.

[0040] Furthermore, the second verification is based on spectral characteristics. A 100-point window is formed by concatenating the generated sequence and its 50 nearest normal data points before and after it, and a Fast Fourier Transform is performed to extract the energy percentage of the dominant frequency component. If this percentage is lower than the 10th percentile of historical data under similar operating conditions, the spectrum is considered abnormal. The third verification is based on smoothness. A third-order spline interpolation curve is constructed using the 10 normal points before and after the missing segment as anchor points, and the mean square residual between the generated sequence and this curve is calculated. If the residual exceeds a threshold (set as 1% of the rated power), a sudden change point is identified. Any verification failure triggers a reconstruction iteration: resampling. z (Training phase) or adjust the context interpolation strategy (inference phase) until all validations are passed.

[0041] The entire system is deployed on an industrial-grade computing device equipped with a central processing unit, a graphics processing unit, and a dedicated tensor processing unit. Raw operational data streams are received in real-time via an industrial Ethernet interface using the Modbus TCP protocol. After protocol parsing and time alignment, the data is stored in a circular buffer. The context feature extraction module reads data from the buffer using a sliding window approach, with a window step size of one sampling period. Each neural network module runs on a TensorRT-optimized inference engine, sharing device memory to reduce data copying overhead. The reconstructed power data, after being formatted and encapsulated, is transmitted back to the wind farm monitoring system via the same Modbus TCP connection and written to the real-time database.

[0042] Furthermore, the training process employs a two-stage optimization strategy. The first stage freezes all parameters of the context encoder, training only the condition generator and discriminator end-to-end for 50 training epochs. The second stage unfreezes all parameters, performing end-to-end fine-tuning with a smaller learning rate, while introducing physical consistency loss, for 100 training epochs. The optimizer used is Adam, with an initial learning rate set to 2e. -4 β1=0.5, β2=0.999. The batch size was set to 64, and each batch contained mixed samples from different wind turbines and different seasons to enhance generalization ability.

[0043] In this implementation, the incremental learning mechanism is continuously activated during system operation. When the wind farm accumulates 1000 high-quality samples of new operational data—defined as samples with no missing data, no abnormal alarms, and located within the normal power generation range—the system automatically triggers a fine-tuning process. First, the standardized parameters of each variable are updated; second, an elastic weight solidification algorithm is used to calculate the importance weight Ω of each network parameter in historical tasks. This calculation is based on the second-order gradient approximation of the parameters using the loss function. ; in, For the first i One parameter, Dold The dataset is the historical training set. When fine-tuning with new data, a regularization term is added to the total loss function. , The old parameter values ​​are used to suppress overwriting of key historical knowledge. Fine-tuning is performed only for 5 training epochs to prevent overfitting to new data.

[0044] The reconstruction confidence metric is output synchronously with each generation. This metric is defined as: ; in, σ -1 It is the inverse function of the Sigmoid function, i.e., the logit transform. Because... D The output value is compressed to (0,1) by the Sigmoid function, and its inverse function can map the discrimination score back to the real number domain, and then linearly scaled to the [0,1] interval. γ t The closer the value is to 1, the closer the generated sample is to the true distribution. The system sets a threshold. γ th = 0.7. Periods with values ​​below this value are marked as low-confidence areas and reported to the operation and maintenance platform via message queue, prompting manual review or arranging sensor calibration.

[0045] To verify the effectiveness of the present invention, the following embodiments and comparative examples were designed.

[0046] In this embodiment, six consecutive months of operational data from thirty 2.5 MW doubly-fed asynchronous wind turbine generators at an onshore wind farm were selected, with a sampling frequency of 0.1 Hz. The dataset contains a total of 427 fully labeled power missing events, including 312 short missing segments and 115 long missing segments. Reconstruction was performed using the method proposed in this invention. The context feature extraction module employs a four-level temporal convolution, the context encoder is a bidirectional GRU, and the condition generator and discriminator structures are as described above. The training set contains data from the first five months, and the test set contains data from the sixth month. The physical consistency constraint threshold is set as described above, and all post-processing verifications are enabled.

[0047] As a comparative example, a traditional five-parameter logistic curve fitting method was used for reconstruction. This method uses only wind speed as input, and obtains the power-wind speed mapping function by fitting historical normal data through nonlinear least squares. Missing power values ​​are directly substituted into the wind speed calculation. This method does not consider any contextual information, nor does it have any generative diversity or physical constraint mechanisms.

[0048] Evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), and power coefficient violation rate (CPR). The sample size was >0.55 and the maximum power slope violation rate (SPR) was calculated. The test results are shown in Table 1.

[0049] Table 1:

[0050] As shown in Table 1, the proposed method significantly outperforms traditional methods in reconstruction accuracy while keeping the physical violation rate at an extremely low level. In particular, in scenarios with long missing segments, this invention effectively avoids the systematic bias caused by traditional methods neglecting the evolution of operating conditions by using context state transition modeling.

[0051] In another embodiment, for a wind turbine in the same wind farm that has experienced blade icing, its power output exhibits atypical attenuation under low temperature and high humidity conditions. Traditional methods, relying on envelopes fitted from historical normal data, cannot reconstruct such anomalous but physically plausible operating conditions. This invention, however, utilizes contextual embedding vectors... c t The icing-related features, such as temperature, humidity, and propeller pitch angle, are explicitly encoded. The conditional generator has learned the power distribution under these contexts during the training phase, successfully generating inefficient power outputs that conform to aerodynamic principles, and the discriminator's confidence level is [high / high]. γ t A value above 0.75 indicates that although the generated results deviate from the normal range, they are still within a reasonable distribution range.

[0052] The above description is based on the preferred embodiments of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0053] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A context-aware wind turbine power curve reconstruction method based on a generative model, characterized in that, Includes the following steps: S1. Obtain the raw operating data stream of the wind turbine, which includes the three-dimensional wind speed vector at the hub height, active power output value, ambient temperature, atmospheric pressure, relative humidity, blade pitch angle, yaw angle, generator speed, pitch motor current, tower vibration acceleration and geographical coordinates. S2. Temporal features are extracted from the original running data stream through the context feature extraction module to obtain a high-dimensional temporal feature tensor. The context feature extraction module includes a multi-level temporal convolutional unit, and each level of temporal convolutional unit includes a one-dimensional causal convolutional layer, a batch normalization layer, and an activation function. S3. Input the high-dimensional temporal feature tensor into the context encoder to generate a context embedding vector representing the current running state; The context encoder adopts a bidirectional gated loop unit structure; S4. During the time period when power data is missing, a reconstructed power value is generated by a condition generator based on the context embedding vector and a preset noise vector; the condition generator is a multi-layer fully connected neural network. S5. Apply physical consistency constraints to the reconstructed power value, including constraining the power coefficient to not exceed the relevant threshold of the Bates limit and constraining the power change rate to not exceed the maximum allowable slope of the wind turbine. S6. Perform post-processing verification on the generated power sequence. After the verification passes, output the reconstructed power data.

2. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: In step S2, the context feature extraction module includes four cascaded temporal convolutional units. The number of input channels for each temporal convolutional unit is 64, 128, and 256 respectively, and the number of output channels is 128, 256, and 512 respectively. The kernel size is 7, the stride is 1, and causal padding is used.

3. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: In step S4, the condition generator comprises a four-layer fully connected neural network with the number of neurons in each layer being 512, 256, 128, and 1, respectively. Except for the last layer, all layers use exponential linear units as activation functions. The input to the condition generator is the concatenation vector of the context embedding vector and the 64-dimensional noise vector.

4. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: In step S4, if the length of the power missing segment does not exceed 20 sampling periods, the context embedding vector is directly used to drive the condition generator. If the length of the missing segment exceeds 20 sampling periods, the context embedding vectors are extracted from both ends of the missing segment, and linear interpolation is performed in its embedding space to obtain the intermediate context state sequence, which is then used to drive the condition generator point by point.

5. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: In step S5, the power coefficient threshold of the physical consistency constraint is dynamically set according to the rated power coefficient corresponding to the wind turbine model; the maximum allowable power slope is determined according to the wind turbine rotational inertia, transmission system damping and converter response capability.

6. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: In step S6, the post-processing verification includes: Verify whether the power value is between the upper and lower envelopes of the theoretical power curve, which are obtained by fitting historical normal data based on a five-parameter logistic function; Verify whether the spectral characteristics of the power sequence are consistent with the typical spectrum under similar operating conditions; Verify the smooth transition between the power sequence and adjacent normal data points.

7. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: Also includes: During the training phase, an adversarial training mechanism is adopted, in which a discriminator distinguishes between the generated power value and the actual power value, and joint optimization is performed by combining the physical consistency constraint loss. During the inference phase, the noise vector is set to zero to make the generated result deterministic.

8. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: It also includes an incremental learning step: when new high-quality operating data is added, the standardized parameters of the context feature extraction module are updated, and the condition generator is fine-tuned using an elastic weight solidification algorithm to adapt to the drift of wind turbine operating characteristics.

9. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: It also includes an output reconstruction confidence index, which is calculated based on the discriminant's discrimination score of the generated samples and is used to identify the reliability of the reconstruction result.

10. The context-aware wind turbine power curve reconstruction method based on a generative model according to claim 1, characterized in that: The original operating data stream is accessed in real time and transmitted via the industrial Ethernet protocol; the reconstructed power data is verified and then transmitted back to the wind farm monitoring system.

Images