A Smart Early Warning Method for Wind Farm Operation and Maintenance Based on Big Data

Abstract

This invention discloses a smart early warning method for wind farm operation and maintenance based on big data, belonging to the field of smart wind farm operation and maintenance technology. The invention generates virtual sensor data in real time by constructing a high-fidelity digital twin of the target wind turbine, calculates the residual sequence between measured and virtual values ​​to eliminate common fluctuations in operating conditions, learns the dynamic evolution of the residuals through a temporal convolutional network composed of four stacked dilated causal convolutional residual blocks, and outputs the predicted mean and variance. It quantifies cognitive uncertainty and random uncertainty by applying Monte Carlo dropout to the network's convolutional layers and performing 100 random forward propagations. By introducing an asymmetric correction rule and a dynamic threshold formula that couples the cognitive uncertainty penalty term with the turbulence sensitivity term, a real-time adaptive threshold band is constructed, achieving high-sensitivity detection and low false alarm rate early warning of early faults in wind turbines under varying operating conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for wind farms, specifically to an intelligent early warning method for wind farm operation and maintenance based on big data. Background Technology

[0002] Wind farm operation and maintenance management relies on real-time monitoring of operating parameters such as temperature, vibration, speed, and power collected by the unit's SCADA system. When a parameter deviates from a preset fixed threshold, an alarm is triggered, and maintenance personnel arrange on-site repairs based on the alarm information. This model has been used in wind farm operation and maintenance practice for many years, but fixed threshold alarms face an irreconcilable contradiction: if the threshold is set too narrowly, during seasonal wind speed changes, gusts, or drastic load adjustments caused by grid dispatch, a large number of instantaneous parameter fluctuations under normal operating conditions will frequently trigger alarms, leading to alarm fatigue among maintenance personnel. Even alarm signals when real faults occur are drowned out by a large number of false alarms. If the threshold is set too wide, during degradation processes such as early pitting corrosion of the unit's gearbox, gradual deterioration of generator bearing grease, and slow corrosion of the blade surface, parameter changes may remain below the wide threshold for months and cannot be detected until the degradation accumulates to a critical point, causing sudden failure. At this point, the repair window is lost, and only passive replacement of major components is possible.

[0003] To address the limitations of fixed-threshold alarms, wind farms have begun to explore data-driven anomaly detection methods. These methods utilize historical data from turbine health operations to train models such as autoencoders, isolated forests, or single-class support vector machines, establishing data representations of normal behavior. During online detection, the deviation of new data from these normal representations is calculated to trigger warnings. While these purely data-driven methods perform well within the operating conditions covered by the training data, they struggle when turbines operate in wind speed ranges, yaw angle ranges, or grid voltage fluctuation ranges that are absent or extremely infrequently observed in the training data. The models lack recognition of these unseen conditions, easily misinterpreting normal but rare operating states as anomalies, leading to false alarms due to cognitive biases. Furthermore, the anomaly scores output by purely data-driven methods lack reliability metrics, making it difficult for maintenance personnel to distinguish between high-scoring anomalies that represent genuine early faults or incorrect extrapolations from unencountered conditions, resulting in a lack of informed decision-making.

[0004] In summary, the early warning methods urgently needed in wind farm operation and maintenance practice should possess the following capabilities: First, they should be able to distinguish between common environmental fluctuations and the unit's own state deviations under changing operating conditions, preventing a batch of false alarms due to changes in wind conditions; second, they should be able to proactively inform the level of unreliability when the model enters the cognitive blind zone, so that operation and maintenance personnel know the confidence level of the current alarm; third, they should be able to maintain a high sensitivity to detect early and subtle degradations within the cognitively reliable range, discovering faults at the stage of planned maintenance rather than at the stage of forced shutdown. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a smart early warning method for wind farm operation and maintenance based on big data. It can generate virtual sensor data in real time by constructing a high-fidelity digital twin of the target wind turbine, calculate the residual sequence between measured and virtual values ​​to eliminate common fluctuations in operating conditions, learn the dynamic evolution of residuals through a temporal convolutional network composed of four stacked dilated causal convolutional residual blocks, and output the predicted mean and variance. It quantifies cognitive uncertainty and random uncertainty by applying Monte Carlo dropout to the network convolutional layers and performing 100 random forward propagations. It constructs a real-time adaptive threshold band by introducing an asymmetric correction rule and a dynamic threshold formula that couples the cognitive uncertainty penalty term with the turbulence sensitivity term. It distinguishes between high-confidence warnings and low-confidence alerts through a sliding window voting and cognitive uncertainty joint discrimination rule, achieving high-sensitivity detection and low false alarm rate early warning of wind turbine faults under varying operating conditions.

[0006] To solve the above-mentioned technical problems, this invention provides the following technical solution: a smart early warning method for wind farm operation and maintenance based on big data, which includes the following components:

[0007] S1. Construct a high-fidelity digital twin of the target wind turbine, synchronize the unit's operating conditions and environmental parameters in real time, and generate physically consistent virtual sensor data;

[0008] S2. Synchronously collect actual sensor data, calculate the residual between actual data and virtual data, and form a multivariate residual time series.

[0009] S3. Construct a temporal convolutional network model consisting of four stacked dilated causal convolutional residual blocks. Each residual block contains a gated dilated convolutional layer and a skip connection. The model takes the residual sequence and environmental parameters as input and outputs the predicted mean and predicted variance of the residuals.

[0010] S4. Apply Monte Carlo dropout to each convolutional layer of the temporal convolutional network model, and obtain the empirical distribution of residual prediction through 100 random forward propagations to quantify the random uncertainty and cognitive uncertainty of residual prediction.

[0011] S5. Based on the aforementioned random uncertainty and cognitive uncertainty, a dynamic adaptive threshold band is constructed by combining the environmental turbulence intensity.

[0012] S6. During online monitoring, calculate the real-time residual sequence and input it into the trained model to perform Bayesian inference. Set a sliding window with a width of 5 sampling points. When at least 3 points in the window have residuals that exceed the dynamic adaptive threshold band and the cognitive uncertainty of all exceeding points is lower than the 80th percentile of historical cognitive uncertainty, generate a high confidence warning. When at least 3 points in the window exceed the threshold, but the cognitive uncertainty of any exceeding point exceeds the historical 80th percentile, generate a low confidence prompt.

[0013] Furthermore, in step S1, the construction of the high-fidelity digital twin adopts a hybrid architecture of mechanistic model and data-driven compensation. The mechanistic model is established based on the unit's aerodynamics, multibody dynamics, and electromagnetic transient equations, and the finite element method is used to perform modal reduction on the blade, main shaft, and nacelle structures to adapt to real-time calculations. The data-driven compensation uses a least-squares support vector machine to correct the uncertainties in the friction damping coefficient and thermal convection coefficient that the mechanistic model cannot accurately describe online. The inputs to the correction model are the unit's active power, speed, and ambient temperature, and the outputs are virtual sensor values ​​of generator front bearing temperature, generator winding temperature, gearbox oil temperature, main shaft X-direction vibration, and nacelle axial vibration. In the initial stage of unit commissioning, the parameters of the mechanism model were initially calibrated using factory test and standard IEC61400-12 power characteristic test data. During the first 12 months after commissioning, a steady-state healthy data segment with wind speed of 6m / s-12m / s and turbulence intensity below 0.15 was selected. With the goal of minimizing the root mean square error between virtual power and actual power, the kernel width and regularization coefficient of the Gaussian kernel in the data-driven compensation model were determined by Bayesian optimization to complete the adaptive calibration of the digital twin. After calibration, the deviation between the virtual and measured values ​​of key temperature measurement points under steady-state conditions can be controlled within 1.2K, and the deviation of vibration measurement points can be controlled within 0.08mm / s.

[0014] Furthermore, the formation of the residual sequence in step S2 includes: aligning the signals from eight measurement points collected by the SCADA system—generator front bearing temperature, generator winding temperature, gearbox oil temperature, main shaft X-direction vibration, main shaft Y-direction vibration, nacelle axial vibration, active power, and generator speed—with the corresponding virtual signals synchronously output by the digital twin at the microsecond timestamp level. In actual engineering, the SCADA data sampling rate is at the second level. To meet the time alignment accuracy, linear interpolation is used to uniformly interpolate the actual and virtual signals onto a millisecond-level time grid, achieving equivalent microsecond-level alignment accuracy. The aligned original data is then processed using a base... Dynamic threshold filtering based on median absolute deviation is used to eliminate instantaneous spike interference from the sensor. Then, ensemble empirical mode decomposition is applied to decompose each signal into a finite number of intrinsic mode functions. The two highest frequency intrinsic mode functions are eliminated to remove random measurement noise. After recombination, a denoised signal is obtained. The difference between the denoised actual signal and the virtual signal is calculated to form an 8-dimensional residual vector. The residual vector is averaged at fixed intervals of 1 minute. For individual missing moments in the mean sequence, weighted local linear interpolation is used to supplement them. The weights are determined according to the time difference with the missing moment. Finally, a regularly sampled residual time series is formed and input into the subsequent model.

[0015] Furthermore, in step S3, the forward propagation of the temporal convolutional network model follows the formula in each residual block: ,in For the first The input feature map of each residual block, Here is the filter weight matrix. For the gated weight matrix, This represents the dilated causal convolution operation. This indicates element-wise multiplication. The sigmoid function maps the gated signal to the 0-1 interval. When the gated response approaches 0, the residual block output is mainly composed of jump connections. The dominant behavior is to preserve the slight fluctuations in the original residuals as the gated response approaches 1. Branches fully participate in feature extraction to capture anomalous patterns, the weight matrix and The initialization uses a Ho's normal distribution with a scale of . ,in The size of the one-dimensional convolution kernel. The number of input channels for the current residual block is given. During network training, the healthy historical residual sequence is used as input. Maximum likelihood estimation is used to maximize the prediction probability as the optimization objective. Each weight matrix is ​​updated through mini-batch gradient descent. In each iteration, the weight update amount is the product of the negative gradient and the learning rate of 0.001. Training stops when the negative log-likelihood loss on the validation set does not decrease for 50 consecutive epochs, and the final weight matrix is ​​determined.

[0016] Furthermore, the training of the temporal convolutional network model employs a joint loss function that combines negative log-likelihood loss with a cognitive uncertainty regularization term. The negative log-likelihood loss term is calculated by dividing the square of the difference between the predicted mean and the true residual by twice the predicted variance, plus half the natural logarithm of the predicted variance. The cognitive uncertainty regularization term penalizes samples where the cognitive uncertainty is higher than the random uncertainty during training. The penalty strength is the portion of the ratio of cognitive uncertainty to random uncertainty exceeding 1 multiplied by the regularization strength coefficient of 0.05. During training, the entire historical health residual dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The AdamW optimizer is used with an initial learning rate of 0.001 and a weight decay coefficient of 0.0001. Each batch size is 256 sequences. During training, the validation set loss is monitored. If the validation loss does not decrease for 20 consecutive epochs, the learning rate is decayed back to 0.5. The model parameters with the minimum validation loss are saved as the final deployment model.

[0017] Furthermore, the random uncertainty in step S4 Defined as the output variance in 100 predictions The square root of the arithmetic mean, cognitive uncertainty is defined as: ,in For the first The mean of the predictions. To predict the standard deviation of the mean, the coefficient is... Let t be the quantile value of the t distribution at 99 degrees of freedom and 99% two-sided confidence level. Dividing by 10 converts the half-width of the confidence interval into an uncertainty metric usable in engineering. Finally, the total uncertainty is obtained by geometrically merging random uncertainty and cognitive uncertainty. The total uncertainty is input into the subsequent threshold construction steps.

[0018] Furthermore, in the context of the aforementioned Before inputting the threshold formula in step S5, for Applying asymmetric correction when the cognitive uncertainty of 10 consecutive sampling points shows a monotonically increasing trend will... Multiply by an amplification factor of 1.3 to broaden the threshold band and prevent false alarms in the cognitive blind spot. When the cognitive uncertainty remains below the historical median level for 15 consecutive sampling points and the random uncertainty fluctuation is less than 10%, Multiply by a shrinkage factor of 0.85 to tighten the threshold band within a stable high confidence interval, thereby improving the sensitivity to capturing subtle anomalies.

[0019] Furthermore, the upper threshold U(t) of the dynamic adaptive threshold band in step S5 is determined by the formula: ,in This is the baseline of the residual prediction mean obtained through sliding window regression of historical health data under current operating conditions. The total uncertainty after correction. and These are real-time cognitive uncertainty and stochastic uncertainty, respectively. The environmental turbulence intensity is calculated using a 30-minute window; its value is the ratio of the standard deviation of wind speed to the mean wind speed. The base threshold coefficient is set to 2.0. The penalty coefficient for cognitive uncertainty is 0.8. This is the turbulence sensitivity coefficient, with a value of 1.5. , , The permissible deviation ranges are ±0.2, ±0.1, and ±0.2, respectively, and the lower threshold is... Depend on Subtracting the product of the same coefficients yields a threshold band whose width can be adjusted in real time based on cognitive uncertainty and turbulence intensity.

[0020] Furthermore, in step S6, when the sliding window determines a high-confidence warning, the anomaly probability is calculated as the ratio of the cumulative area of ​​the residual exceeding the threshold band to the allowable area outside the threshold band within the sliding window. When the ratio is greater than 0.6, a high-probability warning is output. Based on the relative magnitude of the deviation of each measuring point in the residual vector, the top 3 measuring points with the largest contribution are determined as the main contributing variables of the fault. The warning message is encapsulated using the Modbus TCP protocol and mapped to data objects according to the IEC61400-25 standard. The message payload includes the unit number, anomaly timestamp, anomaly probability value, and a list of main contributing variables of the fault. A one-byte confidence flag distinguishes between high-confidence warnings and low-confidence prompts. Maintenance personnel decide whether to immediately arrange on-site investigation based on the confidence flag.

[0021] Furthermore, in step S6, the overall operating environment for online monitoring is deployed on the wind farm edge computing server. The digital twin runs as a containerized microservice and receives real-time data from the wind turbine controller via a message queue telemetry transmission protocol. SCADA data is temporarily stored in a ring buffer and then written to the time-series database. The residual calculation module consumes new data from the time-series database, performs time alignment and noise reduction, and then forwards it to the inference engine. The inference engine loads the trained temporal convolutional network model and calls the graphics processing unit to complete 100 random forward propagations and uncertainty calculations. The dynamic threshold generation module updates the historical health statistics database hourly. Based on the 80th percentile baseline parameters, the early warning decision module performs sliding window judgment and generates alarms according to the real-time threshold band. A single edge server simultaneously processes the early warning tasks of all units in the wind farm. At 2:00 AM every day, the edge server automatically pulls the latest model parameter file from the central training platform and completes the hot loading of the model without interrupting the early warning service.

[0022] Compared with existing technologies, this big data-based intelligent early warning method for wind farm operation and maintenance has the following beneficial effects:

[0023] I. This invention maps the health status monitoring of wind turbine units from the original sensor signal space to the residual space by constructing a residual sequence between a high-fidelity digital twin and the measured data. The residual space eliminates common fluctuations in external operating conditions such as wind speed, ambient temperature, and grid dispatch, so that the residual abnormality directly reflects the state deviation of the unit's own structure or components. This effectively solves the problem of frequent threshold alarms triggered by drastic changes in operating conditions when directly monitoring SCADA data in the traditional way. Simultaneously, a temporal convolutional network is introduced to learn the dynamic evolution of residuals under varying operating conditions. Combined with Bayesian inference, the uncertainty of residual prediction is decomposed into two components: random uncertainty and cognitive uncertainty. When cognitive uncertainty is high, it indicates that the model has entered a sparse region of the training data. At this time, the threshold band is automatically widened to prevent false alarms. When cognitive uncertainty is low and random uncertainty is stable, the threshold band is automatically tightened to improve the detection sensitivity of early minor anomalies. This method achieves adaptive adjustment of the threshold under operating conditions and adaptive adjustment of model confidence using asymmetric correction rules and dynamic threshold formulas. Validation on historical health datasets shows that the false alarm rate is reduced by about 60% compared to the fixed threshold method and by about 35% compared to the dynamic threshold method that does not consider cognitive uncertainty. At the same time, the detection delay is shortened to within 5 minutes after the anomaly occurs.

[0024] Second, this invention constructs a sliding window voting mechanism and a joint discrimination rule for cognitive uncertainty in the early warning decision-making process. A high-confidence early warning is triggered only when the number of out-of-limit points in the window reaches a preset proportion and the cognitive uncertainty of all out-of-limit points is lower than the historical 80th percentile. Otherwise, only a low-confidence prompt is generated without triggering a shutdown alarm. This rule avoids false alarms caused by unreliable predictions in the cognitive blind zone, allowing maintenance personnel to distinguish between fault symptoms that require immediate action and abnormal fluctuations that require continuous observation based on the confidence level. This reduces the number of unplanned shutdowns caused by false alarms. At the same time, the early warning message uses the Modbus TCP protocol to map data objects according to the IEC61400-25 standard, which can be directly connected to the existing monitoring system of the wind farm without the need for additional middleware deployment, thus reducing system integration costs.

[0025] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0027] Figure 1 The main flowchart of a smart early warning method for wind farm operation and maintenance based on big data;

[0028] Figure 2 A flowchart for the construction and calibration of a digital twin for a big data-based intelligent early warning method for wind farm operation and maintenance;

[0029] Figure 3 This is a flowchart illustrating the dynamic threshold generation and sliding window early warning decision-making process of a big data-based intelligent early warning method for wind farm operation and maintenance. Detailed Implementation

[0030] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0031] Example 1

[0032] This embodiment applies to an onshore wind farm in northern my country. The wind farm has an installed capacity of 200MW, consisting of 80 doubly-fed asynchronous wind turbine generators, each with a capacity of 2.5MW, and an annual equivalent utilization of approximately 2200 hours. Located on the southern edge of the Mongolian Plateau, the wind farm experiences an average annual wind speed of approximately 7.5m / s. However, due to the combined influence of monsoons and topography, wind speed fluctuates dramatically, and turbulence intensity often exceeds 0.2 during the winter heating season, posing significant challenges to the health monitoring of critical components. The wind farm's existing early warning system uses a fixed threshold alarm method, resulting in a high false alarm rate during periods of large wind speed fluctuations. Maintenance personnel are overwhelmed by ineffective alarms, and early signs of faults that truly require attention are easily overlooked. This embodiment uses turbine WT-017 in this wind farm as an example to illustrate the complete implementation process of a big data-based intelligent early warning method for wind farm operation and maintenance.

[0033] Step S1: Construct a high-fidelity digital twin of the target wind turbine.

[0034] For the WT-017 unit, a hybrid architecture digital twin based on mechanistic models and data-driven compensation was first established. The mechanistic model encompasses the unit's aerodynamic model, multibody dynamics model, and electromagnetic transient model. The aerodynamic model uses blade element momentum theory to calculate the aerodynamic torque captured by the rotor, with input parameters including real-time wind speed, blade pitch angle, tip speed ratio, and air density. The multibody dynamics model treats the blades, main shaft, gearbox, and generator rotor as a multi-rigid-body system, establishing the torsional vibration differential equation of the transmission chain using Lagrange's equations. The electromagnetic transient model describes the electromagnetic relationships of stator and rotor voltage, current, and flux linkage based on the fifth-order transient equation of the doubly-fed induction generator. To meet real-time calculation requirements, the finite element method was used to perform modal reduction on the blade, main shaft, and nacelle structures. The blades were discretized into 15 beam elements, and the main shaft into 6 beam elements, retaining the first 6 modes. This reduced the original finite element model with tens of thousands of degrees of freedom to a reduced-order model with approximately 200 degrees of freedom, keeping the single-step calculation time within 10 milliseconds.

[0035] The data-driven compensation section employs a least squares support vector machine (LSVM) to correct uncertainties in the mechanistic model, such as the friction damping coefficient and thermal convection coefficient, which are difficult to model precisely. The LSSVM's input feature vector includes four dimensions: current speed, load torque, ambient temperature, and lubricating grease temperature. The output is the correction amount for the damping coefficient and thermal convection coefficient.

[0036] In the initial stage of unit commissioning, key parameters of the mechanistic model were initially calibrated using factory bench test data and standard IEC61400-12 power characteristic test data. The calibration included 23 parameters such as blade aerodynamic parameters, gearbox transmission efficiency, and generator electromagnetic parameters. For the first 12 months after commissioning, a steady-state healthy data range was selected—with wind speeds between 6 m / s and 12 m / s and turbulence intensity below 0.15. This operating condition range covered approximately 65% ​​of the unit's operating time, providing sufficient data and ensuring stable unit operation. With the optimization objective of minimizing the root mean square error between virtual and actual power, a Bayesian optimization method was used to determine the kernel width and regularization coefficient of the Gaussian kernel in the LSSVM model. The surrogate model used for Bayesian optimization is a Gaussian process, and the acquisition function employs an expectation improvement strategy. The search interval for the kernel width is set to [0.1, 10], and the search interval for the regularization coefficient is set to [0.001, 100]. After 40 iterations of evaluation, the model converged, and the kernel width was finally determined to be 2.3 and the regularization coefficient to be 8.7. After the above calibration is completed, the digital twin can receive the unit's operating conditions and environmental parameters in real time and synchronously output 8 virtual sensor signals consistent with the behavior of the physical unit.

[0037] Step S2: Synchronously collect actual sensor data and form a multivariate residual time series.

[0038] The SCADA system of the WT-017 unit collects real-time sensor signals from eight measuring points, including: generator front bearing temperature, generator winding temperature, gearbox oil temperature, main shaft X-direction vibration, main shaft Y-direction vibration, nacelle axial vibration, active power, and generator speed. The digital twin synchronously outputs eight corresponding virtual signals, and the real and virtual signals are aligned at the microsecond level via GPS time synchronization.

[0039] The aligned raw data is then filtered using a dynamic threshold filtering method based on median absolute deviation (MAD) to remove instantaneous spike interference from the sensor. For each measurement point's continuous time series, a window of 11 points (5 points before and 5 points after the current point) is taken as the center. The median absolute deviation within the window is calculated. If the current point deviates from the median within the window by more than 5 times the MAD, it is identified as a spike and removed. Subsequently, ensemble empirical mode decomposition (EMD) is used to decompose each signal into a finite number of intrinsic mode functions (IMFs). By analyzing the power spectral density of each IMF component, the two highest-frequency IMF components are identified as random measurement noise and removed. The resulting recombined signal is the denoised signal. The difference between the denoised actual signal and the virtual signal is calculated to form an 8-dimensional residual vector.

[0040] The 8-dimensional residual vector is averaged at fixed intervals of 1 minute, that is, the arithmetic mean of the residual values ​​of 60 sampling points is calculated every minute to obtain one residual sample point per minute. For individual missing moments in the mean sequence, weighted local linear interpolation is used to supplement them: three valid points before and after the missing moment are taken, and linear interpolation is performed with the inverse distance weighted, finally forming a regularly sampled residual time series.

[0041] Step S3: Construct a temporal convolutional network model.

[0042] A temporal convolutional network model is constructed, consisting of four stacked dilated causal convolutional residual blocks. Each residual block contains a gated dilated convolutional layer and skip connections. The dilation rate of the residual blocks is set to 1, 2, 4, and 8 respectively from the first to the fourth block. The kernel size is uniformly 3, and the number of filters per layer is 64. The model input is the residual sequence of the previous 120 time steps and the corresponding environmental parameters. The output is the predicted mean and predicted variance of the residual at the current time step.

[0043] The forward propagation of each residual block follows the following formula: ,in For the first The input feature map of each residual block, Here is the filter weight matrix. For the gated weight matrix, This represents the dilated causal convolution operation. This indicates element-wise multiplication. The sigmoid function maps the gated signal to the 0-1 interval, and the dilated causal convolution ensures that the model only uses historical information and does not introduce future information when making predictions, which meets the causal requirements of time series prediction. The skip connections in the residual block directly superimpose the input onto the output of the gated convolution, which can effectively alleviate the gradient vanishing problem in deep networks.

[0044] The model training employs a joint loss function that combines the negative log-likelihood loss with a cognitive uncertainty regularization term. The negative log-likelihood loss term is: ,in For true residuals, To predict the mean for the model, To predict variance, a cognitive uncertainty regularization term is applied to samples where cognitive uncertainty is higher than random uncertainty during training. The penalty strength is the portion of the ratio exceeding 1 multiplied by a regularization strength coefficient of 0.05.

[0045] The historical health residual dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The AdamW optimizer was used for training, with an initial learning rate of 0.001, a weight decay coefficient of 0.01, a batch size of 256, and a maximum of 200 training epochs. An early stopping strategy was implemented, terminating training when the validation set loss did not decrease for 20 consecutive epochs. The validation set loss was calculated after each epoch during training, and the model parameters were saved at the epoch with the minimum validation set loss.

[0046] Step S4: Quantify the stochastic and cognitive uncertainties of residual prediction.

[0047] Monte Carlo dropout is applied to each convolutional layer of the trained TCN model, with a dropout rate of 0.1. During the inference phase, the dropout mechanism remains active, and 100 random forward propagations are performed on the same input sample to obtain 100 sets of prediction mean values. and prediction variance .

[0048] Random uncertainty Defined as the output variance in 100 predictions The square root of the arithmetic mean: Cognitive uncertainty is defined as: in The standard deviation of the mean of 100 predictions is given. The coefficient 2.364 represents the quantile of the t-distribution at 99 degrees of freedom and a two-sided 99% confidence level. Dividing by 10 converts the half-width of the confidence interval into an engineering-usable uncertainty metric. The total uncertainty is obtained by merging random uncertainty and cognitive uncertainty using a geometric mean. In Before the input threshold construction step, for Applying asymmetric correction: When the cognitive uncertainty of 10 consecutive sampling points shows a monotonically increasing trend, Multiply by an amplification factor of 1.3 to broaden the threshold band and prevent false alarms in the cognitive blind spot. The Spearman rank correlation coefficient of the most recent 10 points is used for judgment; if the correlation coefficient is greater than 0.6 and the p-value is less than 0.05, it is considered monotonically increasing. When cognitive uncertainty remains below the historical median level for 15 consecutive sampling points, and the random uncertainty fluctuation is less than 10%, [the following is missing from the original text]. Multiply by a contraction factor of 0.85 to tighten the threshold band and improve early warning sensitivity within a stable high confidence interval.

[0049] Step S5: Construct a dynamic adaptive threshold band.

[0050] Upper threshold of dynamic adaptive threshold band Determined by the following formula: The baseline value is the mean of the residual prediction obtained by sliding window regression of health history data under the current operating conditions. The specific calculation method is as follows: taking the current moment as the center, take the 200 samples in the 30-day health history data that are closest to the current wind speed and rotation speed, and perform local weighted regression on their mean residual prediction values ​​to obtain the baseline value. This represents the total uncertainty after correction in step S4; and These are real-time cognitive uncertainty and stochastic uncertainty, respectively. The environmental turbulence intensity is calculated using a 30-minute window. The calculation formula is as follows: ,in The standard deviation of wind speed over 30 minutes. The average wind speed over a 30-minute period. The base threshold coefficient is set to 3.0. The penalty coefficient for cognitive uncertainty is set to 2.0. The turbulence sensitivity coefficient is set to 5.0; lower threshold. With respect to the upper threshold Symmetry, that is .

[0051] The principle behind the above threshold construction method is as follows: when cognitive uncertainty is significantly higher than random uncertainty, it indicates that the current sample is located in a low-density region of the model training data distribution, and the model's prediction confidence for the sample is low. Therefore, the threshold needs to be appropriately relaxed to avoid false alarms. At the same time, increased turbulence intensity means that the wind conditions are more complex and changeable, and the natural fluctuation of the residuals will also increase. Therefore, the threshold should be increased synchronously with the turbulence intensity.

[0052] Step S6: Online monitoring and early warning decision-making.

[0053] During online monitoring, the real-time SCADA data of the WT-017 unit is temporarily stored in a ring buffer and then written to the time-series database. The residual calculation module consumes the new data from the time-series database, performs time alignment and noise reduction, and then forwards it to the inference engine. The inference engine loads the trained TCN model and performs 100 random forward propagations for each new sample to complete uncertainty calculation and threshold construction.

[0054] A sliding window with a width of 5 sampling points is set, and the window slides forward minute by minute as new data arrives. A high-confidence warning is generated when at least 3 points within the window have residuals exceeding the dynamic adaptive threshold band and the cognitive uncertainty of all exceeding points is below the 80th percentile of historical cognitive uncertainty. A low-confidence warning is generated when at least 3 points within the window exceed the threshold, but the cognitive uncertainty of any exceeding point exceeds the historical 80th percentile.

[0055] When the sliding window determines a high-confidence warning, the anomaly probability is further calculated: the anomaly probability is defined as the ratio of the cumulative area of ​​residuals exceeding the threshold band to the allowable area outside the threshold band within the sliding window. Specifically, for each point exceeding the threshold band within the window, the distance between the absolute value of its residual and the boundary of the threshold band is calculated, accumulated, and divided by the product of the window length and the threshold band width. When the ratio is greater than 0.6, a high-probability warning is output. Based on the relative magnitude of the deviation of each measuring point in the 8-dimensional residual vector, the top 3 measuring points with the largest contribution are identified as the main contributing variables of the fault.

[0056] The early warning message is encapsulated using the Modbus TCP protocol and mapped to data objects according to the IEC61400-25 standard. The message payload includes the unit number, anomaly timestamp, anomaly probability value, and a list of major contributing variables to the fault, with a one-byte confidence flag distinguishing between high-confidence warnings and low-confidence alerts. Maintenance personnel decide whether to immediately arrange on-site investigation based on the confidence flag—after a high-confidence warning is triggered, the station duty personnel must respond within 30 minutes and arrange maintenance personnel to go to the unit location for inspection; low-confidence alerts are only logged and added to the watchlist by the condition-based maintenance system.

[0057] This embodiment uses the WT-017 turbine unit of an onshore wind farm in northern my country as an example, and fully presents the implementation process from digital twin construction, residual sequence generation, TCN model training and inference, uncertainty quantification, dynamic threshold construction to online early warning decision-making. This embodiment demonstrates that by combining a mechanism-data hybrid driven digital twin with a Monte Carlo dropout-based Bayesian deep learning framework, accurate and reliable early fault warnings for wind turbine units can be achieved in onshore wind farm environments with drastic wind speed fluctuations.

[0058] Example 2

[0059] This embodiment applies to an offshore wind farm in southeastern coastal my country, located approximately 25 kilometers offshore. The wind farm has a total installed capacity of 400MW, consisting of 40 permanent magnet direct-drive wind turbine generators, each with a capacity of 10MW. Compared to onshore wind farms, offshore wind farms face more severe operating environment challenges: high salt spray, high humidity, frequent typhoons, and complex wave-foundation coupled vibrations. The wind turbine generators are mounted on jacket foundations, with the bottom of the tower bearing the combined effects of wave and tidal loads, resulting in significantly different vibration characteristics compared to onshore turbines. Furthermore, offshore wind farms suffer from poor transportation access, and the window for onboard maintenance after a fault is severely constrained by sea conditions, placing higher demands on the accuracy and reliability of early warning systems compared to onshore wind farms. This embodiment focuses on turbine WT-209 in this offshore wind farm to specifically illustrate the implementation process of this method.

[0060] Step S1: Construct a high-fidelity digital twin of the target wind turbine.

[0061] For the WT-209 permanent magnet direct-drive turbine, the mechanistic model architecture of the digital twin is basically the same as in Example 1, but adjustments are needed to suit the structural characteristics of the direct-drive turbine: the gearbox multibody dynamics model is cancelled and replaced with a rigid-flexible coupled model of the wind turbine directly driving the permanent magnet synchronous generator; a tower-foundation coupled vibration model is added, treating the jacket foundation as an elastic support and the tower as a variable cross-section cantilever beam, considering the combined excitation of wave loads and wind loads. The electromagnetic transient model adopts the dq-axis mathematical model of the permanent magnet synchronous generator.

[0062] In terms of modal reduction, the blades are discretized into 20 beam elements, the tower into 30 beam elements, and the main shaft into 8 beam elements. The first 8 modes are retained, and the model has approximately 350 degrees of freedom after mode reduction. The single-step solution time is controlled within 15 milliseconds.

[0063] The input features of the LSSVM model in the data-driven compensation section are expanded to 6 dimensions, adding two marine environmental parameters: wave height and tidal current velocity. The output remains the correction values ​​for damping coefficient and thermal convection coefficient. Initial calibration utilized factory test data and also referenced measured data from nearby offshore wind farms for the same model of turbine. The adaptive calibration phase selected a steady-state healthy data segment with wind speeds between 6 m / s and 12 m / s, turbulence intensity below 0.15, and significant wave height below 1.5 m—the time window for offshore wind farms to meet these conditions is relatively limited, accounting for approximately 40% of the total operating time. With the objective of minimizing the root mean square error between virtual and actual power, Bayesian optimization determined the kernel width to be 1.8 and the regularization coefficient to be 12.3.

[0064] Step S2: Synchronously collect actual sensor data and form a multivariate residual time series.

[0065] The SCADA system of the WT-209 unit collects signals from eight measurement points that differ slightly from those in Example 1: the generator front bearing temperature is replaced by the permanent magnet generator rotor bearing temperature; the generator winding temperature remains unchanged; the gearbox oil temperature is removed and replaced by the converter IGBT module temperature; and the main shaft X-direction vibration, main shaft Y-direction vibration, nacelle axial vibration, active power, and generator speed remain unchanged. The synchronization alignment of actual and virtual signals, spike filtering, EEMD noise reduction, and residual vector calculation are consistent with Example 1 and will not be repeated here.

[0066] Step S3: Construct a temporal convolutional network model.

[0067] The architecture of the TCN model is exactly the same as in Example 1: four stacked dilated causal convolutional residual blocks with dilation rates of 1, 2, 4, and 8 respectively, a kernel size of 3, and 64 filters. The input consists of the residual sequence of the first 120 time steps and environmental parameters. The model training uses the same joint loss function and optimizer configuration. The historical health dataset consists of eight consecutive months of health operation data from the WT-209 unit, and is still divided into training, validation, and test sets in an 8:1:1 ratio.

[0068] It should be noted that the reliability of sensors in offshore wind fields is lower than that in onshore wind fields due to salt spray corrosion and typhoon impacts, resulting in greater fluctuations in data quality. Therefore, a quality control step based on multi-sensor cross-validation was added to the data preprocessing stage: when the physical consistency of a reading at a certain measuring point with other related measuring points deviates beyond a preset threshold, the residual sample at that moment is marked as low-quality data and assigned a lower sample weight during training, with a weighting coefficient of [value missing]. ,in Standardized score for the degree of consistency deviation.

[0069] Steps S4 and S5: Quantify the uncertainty and construct a dynamic adaptive threshold band.

[0070] The uncertainty quantification method and the construction formula for the dynamic adaptive threshold band are exactly the same as in Example 1. It is worth noting, however, that the environmental turbulence intensity of the offshore wind field... The calculation window remains 30 minutes, but considering that wind speed changes are relatively gentler at sea than on land, Typical values ​​are usually between 0.08 and 0.15, lower than the 0.12 to 0.25 for onshore wind fields. Therefore, in actual operation, the variation range of the dynamic threshold bandwidth of offshore wind fields is smaller than that of onshore wind fields, and the dynamic adjustment of the threshold relies more on perceived uncertainty. The change is not a change in turbulence intensity.

[0071] Step S6: Online monitoring and early warning decision-making.

[0072] The overall online monitoring environment is deployed on an edge computing server within the offshore substation of the wind farm. The digital twin runs as a containerized microservice, receiving real-time data from the wind turbine controller via a message queue telemetry transmission protocol. SCADA data is temporarily stored in a ring buffer before being written to the InfluxDB time-series database. The residual calculation module consumes new data from the time-series database in a streaming manner, performs time alignment and noise reduction, and then forwards it to the inference engine. The inference engine loads the trained TCN model and calls the GPU to complete 100 random forward propagations and uncertainty calculations. Each inference takes approximately 120 milliseconds, meeting the real-time requirement of a 1-minute sampling interval. The dynamic threshold generation module updates the historical health statistics database hourly. And the 80th percentile baseline parameter of historical cognitive uncertainty.

[0073] The early warning decision module performs a 5-point sliding window judgment based on real-time threshold bands. For the WT-209 unit, the transmission of early warning messages uses a dedicated fiber optic communication network for offshore wind farms. Modbus TCP messages are aggregated through a gateway and then uploaded to the onshore control center. After a high-confidence early warning is triggered, the operation and maintenance response process is more complex than onshore due to the limitations of tides and sea states on offshore boarding operations: the early warning information is first confirmed by the duty engineer at the control center, and then the boarding operations of maintenance vessels or helicopters are arranged according to the sea state forecast window. The entire response cycle usually takes 12 to 48 hours. Therefore, the lead time and accuracy of early warnings are particularly important for offshore wind farms. This method effectively controls the false alarm rate through dynamic thresholds and uncertainty quantification mechanisms, providing reliable early warning information for offshore operation and maintenance decisions.

[0074] This embodiment uses the WT-209 turbine at an offshore wind farm in southeastern coastal my country as an example to demonstrate the specific implementation of the method of the present invention on a permanent magnet direct-drive offshore wind turbine. Compared with Embodiment 1, this embodiment focuses on illustrating the structural adjustments of the digital twin for direct-drive turbines and the marine environment, the special processing for marine data quality control, and the adaptation of early warning response procedures in marine operation and maintenance scenarios. This embodiment shows that the method of the present invention has good adaptability to different turbine types and environments, and can operate stably and provide reliable early warning services in harsh marine environments with high salt spray, high humidity, and frequent typhoons.

[0075] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart early warning method for wind farm operation and maintenance based on big data, characterized in that, The method comprises the following components: S1. Construct a high-fidelity digital twin of the target wind turbine, synchronize the unit's operating conditions and environmental parameters in real time, and generate physically consistent virtual sensor data; S2. Synchronously collect actual sensor data, calculate the residual between actual data and virtual data, and form a multivariate residual time series. S3. Construct a temporal convolutional network model consisting of four stacked dilated causal convolutional residual blocks. Each residual block contains a gated dilated convolutional layer and a skip connection. The residual sequence and environmental parameters are used as inputs, and the predicted mean and predicted variance of the residuals are output. S4. Apply Monte Carlo dropout to each convolutional layer of the temporal convolutional network model, and obtain the empirical distribution of residual prediction through 100 random forward propagations to quantify the random uncertainty and cognitive uncertainty of residual prediction. S5. Based on the aforementioned random uncertainty and cognitive uncertainty, a dynamic adaptive threshold band is constructed by combining the environmental turbulence intensity. S6. During online monitoring, calculate the real-time residual sequence and input it into the trained model to perform Bayesian inference. Set a sliding window with a width of 5 sampling points. When at least 3 points in the window have residuals that exceed the dynamic adaptive threshold band and the cognitive uncertainty of all exceeding points is lower than the 80th percentile of historical cognitive uncertainty, generate a high confidence warning. When at least 3 points in the window exceed the threshold, but the cognitive uncertainty of any exceeding point exceeds the historical 80th percentile, generate a low confidence prompt.

2. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 1, characterized in that, In step S1, the construction of a high-fidelity digital twin adopts a hybrid architecture of mechanistic model and data-driven compensation. The mechanistic model is established based on the unit's aerodynamics, multibody dynamics, and electromagnetic transient equations. The finite element method is used to perform modal reduction on the blade, main shaft, and nacelle structures to adapt to real-time calculation. The data-driven compensation uses least squares support vector machine to correct the uncertainties in friction damping coefficient and thermal convection coefficient that the mechanistic model cannot accurately describe. In the early stage of unit commissioning, the parameters of the mechanistic model are initially calibrated using factory test and standard IEC61400-12 power characteristic test data. During the first 12 months after commissioning, a steady-state healthy data segment with wind speed of 6m / s-12m / s and turbulence intensity below 0.15 is selected. With the goal of minimizing the root mean square error between virtual power and actual power, the kernel width and regularization coefficient of the Gaussian kernel in the data-driven compensation model are determined by Bayesian optimization to complete the adaptive calibration of the digital twin.

3. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 1, characterized in that, The formation of the residual sequence in step S2 includes: aligning the signals from eight measurement points (generator front bearing temperature, generator winding temperature, gearbox oil temperature, main shaft X-direction vibration, main shaft Y-direction vibration, nacelle axial vibration, active power, and generator speed) collected by the SCADA system with the corresponding virtual signals synchronously output by the digital twin at the microsecond timestamp level; using a dynamic threshold filtering method based on median absolute deviation to remove instantaneous spike interference from the sensors on the aligned raw data; then using ensemble empirical mode decomposition to decompose each signal into a finite number of intrinsic mode functions (IMFs); removing the two highest frequency IMFs to eliminate random measurement noise; and recombining them to obtain a denoised signal; calculating the difference between the denoised actual signal and the virtual signal to form an 8-dimensional residual vector; and taking the mean of the residual vector at fixed intervals of 1 minute; supplementing any missing moments in the mean sequence using weighted local linear interpolation; and finally, forming a regularly sampled residual time series for input into the subsequent model.

4. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 1, characterized in that, In step S3, the forward propagation of the temporal convolutional network model follows the formula in each residual block: ,in For the first The input feature map of each residual block, Here is the filter weight matrix. For the gated weight matrix, This represents the dilated causal convolution operation. This indicates element-wise multiplication. The sigmoid function maps the gated signal to the 0-1 interval.

5. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 4, characterized in that, The training of the temporal convolutional network model employs a joint loss function that combines negative log-likelihood loss with a cognitive uncertainty regularization term. The negative log-likelihood loss term is calculated by dividing the square of the difference between the predicted mean and the true residual by twice the predicted variance, plus half the natural logarithm of the predicted variance. The cognitive uncertainty regularization term penalizes samples where the cognitive uncertainty is higher than the random uncertainty during training. The penalty strength is the portion of the ratio of cognitive uncertainty to random uncertainty exceeding 1 multiplied by the regularization strength coefficient of 0.

05. During training, the entire historical health residual dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The AdamW optimizer is used, and the validation set loss is monitored during training.

6. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 1, characterized in that, Random uncertainty in step S4 Defined as the output variance in 100 predictions The square root of the arithmetic mean, cognitive uncertainty is defined as: ,in For the first The mean of the predictions. To predict the standard deviation of the mean, the coefficient is... Let t be the quantile value of the t distribution at 99 degrees of freedom and 99% two-sided confidence level. Dividing by 10 converts the half-width of the confidence interval into an uncertainty metric usable in engineering. Finally, the total uncertainty is obtained by geometrically merging random uncertainty and cognitive uncertainty. The total uncertainty is input into the subsequent threshold construction steps.

7. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 6, characterized in that, In the above Before inputting the threshold formula in step S5, for Applying asymmetric correction when the cognitive uncertainty of 10 consecutive sampling points shows a monotonically increasing trend will... Multiply by an amplification factor of 1.3 to broaden the threshold band and prevent false alarms in the cognitive blind spot. When the cognitive uncertainty remains below the historical median level for 15 consecutive sampling points and the random uncertainty fluctuation is less than 10%, Multiply by a contraction factor of 0.85 to tighten the threshold band within a stable high confidence interval.

8. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 1, characterized in that, In step S5, the upper threshold U(t) of the dynamically adaptive threshold band is determined by the formula: ,in This is the baseline of the residual prediction mean obtained through sliding window regression of historical health data under current operating conditions. The total uncertainty after correction. and These are real-time cognitive uncertainty and stochastic uncertainty, respectively. The environmental turbulence intensity is calculated with a 30-minute window. Based on the basic threshold coefficient, The penalty coefficient for cognitive uncertainty. This is the turbulence sensitivity coefficient.

9. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 1, characterized in that, In step S6, when the sliding window determines a high-confidence warning, the anomaly probability is calculated as the ratio of the cumulative area of ​​the residual exceeding the threshold band to the allowable area outside the threshold band within the sliding window. When the ratio is greater than 0.6, a high-probability warning is output. Based on the relative magnitude of the deviation of each measuring point in the residual vector, the top 3 measuring points with the largest contribution are determined as the main contributing variables of the fault. The warning message is encapsulated using the Modbus TCP protocol and mapped to data objects according to the IEC61400-25 standard. The message payload includes the unit number, anomaly timestamp, anomaly probability value, and a list of main contributing variables of the fault. A one-byte confidence flag distinguishes between high-confidence warnings and low-confidence prompts. Maintenance personnel decide whether to immediately arrange on-site investigation based on the confidence flag.

10. The intelligent early warning method for wind farm operation and maintenance based on big data as described in claim 1, characterized in that, In step S6, the overall operating environment for online monitoring is deployed on the wind farm edge computing server. The digital twin runs as a containerized microservice and receives real-time data from the wind turbine controller via a message queue telemetry transmission protocol. SCADA data is temporarily stored in a ring buffer and then written to the time-series database. The residual calculation module consumes new data from the time-series database, performs time alignment and noise reduction, and then forwards it to the inference engine. The inference engine loads the trained temporal convolutional network model and calls the graphics processing unit to complete 100 random forward propagations and uncertainty calculations. The dynamic threshold generation module updates the historical health statistics database hourly. Based on the baseline parameters of the 80th percentile, the early warning decision module performs sliding window judgment and generates alarms according to the real-time threshold band. A single edge server can simultaneously handle the early warning tasks of all units in the wind farm.

Images