Wind power plant intelligent operation and maintenance method and system based on cloud platform

Abstract

The invention relates to the technical field of energy supply, in particular to a wind power plant intelligent operation and maintenance method and system based on a cloud platform, and the method comprises the following steps: collecting current, voltage and temperature, calculating the discharge depth, the charging duration and the variable pitch frequency to construct a working condition feature vector, generating a virtual attenuation data set through the generative adversarial network in combination with physical consistency judgment, and a health state prediction model is obtained by optimizing the gradient lifting decision tree, the model is called to calculate future residual capacity, a threshold value is compared to determine an early warning state, and a structured operation and maintenance instruction is constructed. According to the method, the multi-dimensional operation parameters are collected to construct the working condition feature vectors, battery attenuation is simulated to generate the virtual data set, the prediction model is trained, the health state of the wind turbine generator equipment is accurately evaluated, early warning is judged according to the prediction result, the operation and maintenance instruction is constructed, manual inspection requirements are reduced, resource waste and operation interruption are avoided, and the operation efficiency is improved. Potential fault risks are recognized prospectively, maintenance strategies are optimized, operation stability is improved, and the service life of equipment is prolonged.

Description

Technical Field

[0001] This invention relates to the field of energy supply technology, and in particular to a cloud-based intelligent operation and maintenance method and system for wind farms. Background Technology

[0002] The energy supply technology field involves systems for managing the production, scheduling, and distribution of resources such as electricity, gas, or water. Core aspects of this field include monitoring energy facilities using data processing methods, optimizing supply networks, and supporting operation and maintenance processes. Traditional intelligent operation and maintenance (O&M) methods for wind farms refer to technologies that monitor the operating status of wind turbine generators and schedule maintenance. These methods typically rely on regular on-site inspections by technicians, combined with analysis of operating parameters reported by local data acquisition and monitoring systems. Maintenance decisions are primarily made based on fixed time cycles or after equipment malfunctions are reported.

[0003] Current technologies rely on regular on-site inspections by technicians, consuming significant manpower and resources. This makes it difficult to achieve continuous, real-time monitoring and refined management of wind power equipment. Operational parameter analysis is limited to the instantaneous state of the equipment, failing to delve into potential fault patterns. Maintenance decisions are primarily based on fixed time cycles or reactive execution after equipment malfunctions trigger alarms. This approach fails to fully utilize equipment operational data, potentially leading to premature maintenance and resource waste, or measures being taken only after a fault has occurred, increasing downtime and maintenance costs. Ultimately, this impacts the overall operational efficiency and reliability of the wind farm, demonstrating a lack of proactive maintenance capabilities. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a cloud-based intelligent operation and maintenance method and system for wind farms.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a cloud-based intelligent operation and maintenance method for wind farms, comprising the following steps: S1: Collect the pitch motor drive current sequence, battery pack terminal voltage sequence and battery ambient temperature, calculate the depth of discharge, charging time and pitch frequency, and combine the depth of discharge, the charging time, the pitch frequency and the battery ambient temperature to construct the operating condition feature vector; S2: Based on the operating condition feature vector, the discriminant network of the generative adversarial network calls the battery equivalent circuit differential calculation module to determine whether the internal equivalent series resistance change trajectory is monotonically non-decreasing, performs physical consistency judgment, and generates a virtual battery degradation dataset. S3: Using the operating condition feature vector in the virtual battery degradation dataset as input and the predicted capacity value as the output label, optimize the gradient boosting decision tree model to obtain the health status prediction model. S4: Collect real-time signals to construct a real-time operating condition feature vector, call the health status prediction model to calculate the future remaining capacity percentage value, compare the future remaining capacity percentage value with the threshold to determine the early warning status signal, and construct structured operation and maintenance instructions.

[0006] As a further aspect of the present invention, the virtual battery degradation dataset includes the operating condition feature vector and the output label, and the structured operation and maintenance instructions include the future remaining capacity percentage value, the early warning status signal, and the threshold.

[0007] As a further aspect of the present invention, the calculation process for the depth of discharge, the charging duration, and the pitch frequency specifically includes: Based on the pitch motor drive current sequence and the battery pack terminal voltage sequence, the cumulative discharge ampere-hours of the current discharge cycle are calculated using the ampere-hour integration method, and the depth of discharge is determined by the ratio of the cumulative discharge ampere-hours to the nominal battery capacity. By monitoring the voltage sequence at the battery pack terminals and determining the voltage slope and current direction, the charging start point and charging end point are accurately identified, and the time interval from the charging start point to the charging end point is calculated to obtain the charging duration. The fluctuation characteristics of the pitch motor drive current sequence are analyzed, and the time-frequency spectrum of the pitch motor drive current sequence is extracted by short-time Fourier transform. The region with concentrated high-frequency energy is identified from the time-frequency spectrum to count the number of pitch movements per unit time, and the pitch frequency is calculated.

[0008] As a further aspect of the present invention, step S2 specifically comprises: S21: Initialize the generator network of the generative adversarial network. The generator network receives a random noise vector and combines it with the working condition feature vector as a conditional input to generate an initial change trajectory. S22: Input the initial change trajectory and the working condition feature vector together into the discrimination network; S23: The discrimination network calls the battery equivalent circuit differential calculation module to perform the physical consistency judgment on the initial change trajectory; S24: Based on the result of the physical consistency judgment, update the parameters of the generating network and the discriminating network until the internal equivalent series resistance change trajectory output by the generating network satisfies the physical consistency judgment, thus forming the virtual battery degradation dataset.

[0009] As a further aspect of the present invention, the specific process of the physical consistency judgment is as follows: The battery equivalent circuit differential calculation module sets the initial parameters of the battery second-order RC equivalent circuit model based on the battery ambient temperature and the depth of discharge in the operating condition feature vector. The initial parameters include ohmic internal resistance, first polarization internal resistance, second polarization internal resistance, first polarization capacitance, and second polarization capacitance. The variable pitch motor drive current sequence in the operating condition feature vector is used as the excitation input of the second-order RC equivalent circuit model. By solving the state differential equation of the second-order RC equivalent circuit model in real time, a theoretical trajectory of the change of internal equivalent series resistance is calculated. Compare the trajectory of the internal equivalent series resistance variation generated by the generative adversarial network with the theoretical trajectory of the internal equivalent series resistance variation. Determine whether the trajectory of the internal equivalent series resistance variation generated by the generative adversarial network is monotonically non-decreasing throughout the entire process, and whether its slope is within the preset fluctuation range of the theoretical trajectory of the internal equivalent series resistance variation. If so, it is determined that the physical consistency judgment has been passed.

[0010] As a further aspect of the present invention, step S3 specifically comprises: S31: Use the operating condition feature vector in the virtual battery degradation dataset as the training input, and use the predicted capacity value corresponding to the virtual battery degradation dataset as the output label; S32: The Bayesian optimization algorithm is used to automatically search for the optimal combination of hyperparameters of the gradient boosting decision tree model. The optimal combination of hyperparameters includes the maximum depth of the decision tree, the learning rate, and the subsampling ratio. S33: Configure the gradient boosting decision tree model using the optimal hyperparameter combination, and iteratively train the gradient boosting decision tree model using the virtual battery decay dataset until the root mean square error of the gradient boosting decision tree model on the validation set is lower than the preset convergence threshold. S34: Save the gradient boosting decision tree model after training to obtain the health status prediction model.

[0011] As a further aspect of the present invention, step S4 specifically comprises: S41: Real-time acquisition of the pitch motor drive current sequence, the battery pack terminal voltage sequence, and the battery ambient temperature, and construction of the real-time operating condition feature vector by calculating the real-time depth of discharge, the charging time, and the pitch frequency; S42: Input the real-time operating condition feature vector into the health status prediction model, and the health status prediction model outputs the percentage of the future remaining capacity within a specified future time period. S43: Set a capacity decay threshold of 80%, compare the future remaining capacity percentage with the threshold, and if the future remaining capacity percentage is lower than the threshold, generate a high-level warning signal; if the future remaining capacity percentage is not lower than the threshold, generate a normal-level warning signal. S44: Combine the future remaining capacity percentage, the early warning status signal, and the corresponding battery number and wind turbine number to construct a structured operation and maintenance instruction and push it to the cloud platform.

[0012] As a further aspect of the present invention, the battery equivalent circuit differential calculation module solves the battery pack terminal voltage sequence based on the second-order RC equivalent circuit model. Differential equations: ; in, Let be the value of the battery pack terminal voltage sequence at time t. Let be the value of the pitch motor drive current sequence at time t. The state of charge of the battery. The open-circuit voltage corresponding to the state of charge. The ohmic internal resistance is given. The polarization voltage of the first RC network. The polarization voltage of the second RC network; The trajectory of the change in the internal equivalent series resistance From the above and the The above The equivalent first polarization internal resistance and the second polarization internal resistance are jointly determined during the charging and discharging process. The discrimination network verifies the... time derivative Whether it holds true at all points in time is used to perform the physical consistency judgment on whether the trajectory of the change of the internal equivalent series resistance is monotonically non-decreasing.

[0013] As a further aspect of the present invention, the optimization process of the gradient boosting decision tree model further includes: During the iterative training of the gradient boosting decision tree model, a dimensionally normalized regularized loss function is introduced. ; and ; in, Let i be the predicted capacity value for the i-th sample. For the gradient boosting decision tree model, the i-th feature vector of the working condition is... The predicted output, The reference capacity value used for dimensional normalization. For the k-th decision tree, For the regularization term of the k-th decision tree, The regularization coefficient is . Let be the number of leaf nodes in the k-th decision tree. Let J be the weight of the j-th leaf node of the k-th decision tree. The penalty coefficient is the number of leaf nodes. The penalty coefficient for the weight of the leaf node is... For sample index, For decision tree indexing, Index for leaf nodes; The Bayesian optimization algorithm is used to adjust the... The above The above and the learning rate, to minimize the regularization loss function. With the goal of improving the gradient boosting decision tree model, the generalization ability and prediction accuracy of the health status prediction model are optimized.

[0014] A cloud-based intelligent operation and maintenance system for wind farms, the system being used to implement the aforementioned cloud-based intelligent operation and maintenance method for wind farms, the system comprising: The acquisition module is used to acquire the pitch motor drive current sequence, the battery pack terminal voltage sequence, and the battery ambient temperature; The operating condition feature construction module is used to calculate the depth of discharge, the charging time and the pitch frequency, and combine the depth of discharge, the charging time, the pitch frequency and the battery ambient temperature to construct the operating condition feature vector. The virtual data generation module is used to, based on the operating condition feature vector, call the battery equivalent circuit differential calculation module through the discriminant network of the generative adversarial network to determine whether the trajectory of the change of the internal equivalent series resistance is monotonically non-decreasing, perform physical consistency judgment, and generate the virtual battery degradation dataset. A state prediction modeling module is used to take the operating condition feature vector in the virtual battery degradation dataset as input, the predicted capacity value as output label, optimize the gradient boosting decision tree model, and obtain a health state prediction model. The intelligent operation and maintenance instruction module is used to collect real-time signals to construct the real-time operating condition feature vector, call the health status prediction model to calculate the future remaining capacity percentage value, compare the future remaining capacity percentage value with a threshold to determine the early warning status signal, and construct structured operation and maintenance instructions.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention constructs a feature vector of operating conditions by collecting multi-dimensional operating parameters, simulates the battery degradation process to generate a virtual dataset, trains a prediction model, and accurately assesses the health status of the equipment. Based on the prediction results, it determines the early warning status and constructs structured operation and maintenance instructions, thereby achieving predictive maintenance management of wind turbine units. This significantly reduces the need for manual inspections and avoids resource waste and operational interruptions caused by reactive maintenance. The solution can proactively identify potential fault risks, optimize maintenance strategies, significantly improve the operational stability of wind farms, reduce operation and maintenance costs, and extend equipment lifespan. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the main steps of the present invention; Figure 2 This is a detailed flowchart of step S1 of the present invention; Figure 3 This is a detailed flowchart of step S2 of the present invention; Figure 4 This is a detailed flowchart of step S3 of the present invention; Figure 5 This is a detailed flowchart of step S4 of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a cloud-based intelligent operation and maintenance method for wind farms, comprising the following steps: S1: Collect the pitch motor drive current sequence, battery pack terminal voltage sequence and battery ambient temperature, calculate the depth of discharge, charging time and pitch frequency, and combine the depth of discharge, charging time, pitch frequency and battery ambient temperature to construct the operating condition feature vector. The calculation process for depth of discharge, charging time, and pitch frequency specifically includes: Based on the pitch motor drive current sequence and the battery pack terminal voltage sequence, the cumulative discharge ampere-hours of the current discharge cycle are calculated by the ampere-hour integration method, and the depth of discharge is determined by the ratio of the cumulative discharge ampere-hours to the nominal battery capacity. By monitoring the battery pack terminal voltage sequence and judging the voltage slope and current direction, the charging start point and charging end point are accurately identified, and the time interval from the charging start point to the charging end point is counted to obtain the charging time. The fluctuation characteristics of the pitch motor drive current sequence were analyzed. The time-frequency spectrum of the pitch motor drive current sequence was extracted by short-time Fourier transform. The region with high-frequency energy concentration was identified from the time-frequency spectrum to count the number of pitch movements per unit time and calculate the pitch frequency.

[0020] In the wind farm turbine monitoring system on the cloud platform, the data acquisition cycle is set to 1 second, continuously collecting data from the battery pack of the pitch system of turbine WT01. A 10-minute (600-second) data segment is acquired, including the pitch motor drive current sequence. (Unit: A) Battery pack terminal voltage sequence (Unit: V) and battery ambient temperature (unit: °C). In At 10:00, the ambient temperature of the battery was 15°C.

[0021] First, the depth of discharge is calculated using the ampere-hour integration method. and The system detected that... At what time, the current The change from 0A to 50A indicates the start of a discharge cycle; At what time, the current The discharge cycle ends when the current drops to 0A. This process continues for 100 seconds. to During the discharge cycle, for Sequence (e.g.: A) Perform time integration. Cumulative discharge ampere-hours = ,in The duration is 1 second. Assuming the average discharge current over this 100 seconds is 36A, the cumulative discharge capacity is... Convert the cumulative discharge ampere-hours to ampere-hours (Ah): The nominal capacity of the battery pack is set to 100Ah. The depth of discharge (DoD) is calculated as follows: The depth of discharge was determined to be 1%. Next, the charging time was calculated: monitoring continued. and .exist At the second, it was detected The current changes from 0A to -20A (a negative value indicates charging), and at the same time... The voltage begins to rise, and the slope is greater than the preset charging judgment slope of 0.01V / s. The system then determines... The charging start point is in seconds. Continuous monitoring. ,exist seconds, The float charge voltage reaches 28.8V, at which point the charging current... It has dropped to -1.5A. At what time, the current The current drops to -0.9A, which is less than the charging cutoff current threshold of 1.0A. The system then determines... The charging stop time is specified in seconds. Charging time is then recorded. The charging time was 1810 seconds. Next, the pitch frequency was calculated: Analysis. The fluctuation characteristics of the sequence over the entire 600-second data segment were analyzed. Short-Time Fourier Transform (STFT) processing was employed: the STFT window function was set to a Hamming window, the window length to 60 seconds, and the overlap length to 50 seconds (i.e., a step size of 10 seconds). For the 600-second data segment... The sequence is subjected to STFT, resulting in a time-frequency spectrum matrix. Rows represent frequencies (e.g., 0Hz to 0.5Hz, resolution 0.01Hz), and columns represent times (60 time windows). High-frequency energy concentration regions are identified from the time-frequency spectrum: these regions are defined as those with frequencies in the range of 0.1Hz to 0.5Hz and energy amplitudes greater than a preset energy threshold (e.g., -20dB). The number of pitch maneuvers per unit time (i.e., 600 seconds) is counted: by scanning the time-frequency spectrum, the number of pitch maneuvers within a given time frame is determined. Second, Second, Second, Seconds and Within each time window corresponding to a second, an energy concentration region meeting the above conditions appeared. The number of pitch maneuvers was counted as 5.

[0022] Calculate the pitch frequency: Finally, the combined operating condition feature vector is derived from the combination of discharge depth (1%), charging time (1810 seconds), pitch frequency (0.00833Hz), and battery ambient temperature (15°C). The constructed operating condition feature vector is as follows: .

[0023] Please see Figure 1 and Figure 3S2: Based on the operating condition feature vector, the discriminant network of the generative adversarial network calls the battery equivalent circuit differential calculation module to determine whether the internal equivalent series resistance change trajectory is monotonically non-decreasing, performs physical consistency judgment, and generates a virtual battery degradation dataset. The virtual battery degradation dataset includes operating condition feature vectors and output labels; The specific steps of S2 are as follows: S21: Initialize the generator network of the generative adversarial network. The generator network receives a random noise vector and combines it with the working condition feature vector as a conditional input to generate an initial change trajectory. S22: Input the initial trajectory change and the working condition feature vector together into the discriminant network; S23: The network calls the battery equivalent circuit differential calculation module to perform a physical consistency judgment on the initial change trajectory; S24: Based on the results of the physical consistency judgment, update the parameters of the generator network and the discriminator network until the internal equivalent series resistance change trajectory output by the generator network satisfies the physical consistency judgment, thus forming a virtual battery degradation dataset. The specific process for determining physical consistency is as follows: The battery equivalent circuit differential calculation module sets the initial parameters of the battery second-order RC equivalent circuit model based on the battery ambient temperature and discharge depth in the operating condition feature vector. The initial parameters include ohmic internal resistance, first polarization internal resistance, second polarization internal resistance, first polarization capacitance, and second polarization capacitance. The variable pitch motor drive current sequence in the operating condition feature vector is used as the excitation input of the second-order RC equivalent circuit model. By solving the state differential equation of the second-order RC equivalent circuit model in real time, a theoretical internal equivalent series resistance change trajectory is calculated. Compare the trajectory of the internal equivalent series resistance generated by the generative adversarial network with the theoretical trajectory of the internal equivalent series resistance. Determine whether the trajectory of the internal equivalent series resistance generated by the generative adversarial network is monotonically non-decreasing throughout the entire process, and whether its slope is within the preset fluctuation range of the theoretical trajectory of the internal equivalent series resistance. If so, it is judged as passing the physical consistency judgment. The battery equivalent circuit differential calculation module solves the battery pack terminal voltage sequence based on a second-order RC equivalent circuit model. Differential equations: ; in, Let be the value of the battery pack terminal voltage sequence at time t. Let be the value of the pitch motor drive current sequence at time t. This refers to the battery's state of charge. This is the open-circuit voltage corresponding to the state of charge. For ohmic internal resistance, The polarization voltage of the first RC network. The polarization voltage of the second RC network; Trajectory of internal equivalent series resistance variation Depend on as well as , The equivalent first polarization internal resistance and the second polarization internal resistance are jointly determined during the charging and discharging process, and the discrimination network is verified. time derivative Whether it holds true at all points in time is used to determine the physical consistency of whether the trajectory of the internal equivalent series resistance change is monotonically non-decreasing.

[0024] S21: Initialize the generator network of the Generative Adversarial Network (GAN). The generator network receives a 100-dimensional random noise vector. (For example, vectors generated by the standard normal distribution) Simultaneously, the working condition feature vector constructed in S1 will be... As conditional inputs, these are fed into the generator network. The generator network (e.g., a multilayer perceptron consisting of three fully connected layers, with hidden layer dimensions of 256 and 512, respectively) outputs an initial trajectory of the internal equivalent series resistance variation. The trajectory is a time series, for example, over 100 time steps: (Unit: Ohm)

[0025] S22: Will Sequence and operating condition feature vectors Together, they serve as input to the discrimination network.

[0026] S23: Determine if the network calls the battery equivalent circuit differential calculation module. Perform a physical consistency check. The specific process of the physical consistency check is as follows: The battery equivalent circuit differential calculation module first queries and sets the initial parameters of the battery's second-order RC equivalent circuit model from the preset parameter library (calibrated through offline experiments) based on the battery ambient temperature (15°C) and depth of discharge (1%) in the operating condition feature vector. The initial parameters are set as follows: ohmic internal resistance. First polarization internal resistance First polarization capacitor Second polarization internal resistance Second polarization capacitor The module retrieves the corresponding pitch motor drive current sequence from S1. ( A) As the excitation input for the second-order RC equivalent circuit model. By solving the state differential equations of the second-order RC equivalent circuit model in real time (e.g., using the fourth-order Runge-Kutta method with a time step of 0.1s), the excitation input at time step is calculated. Under excitation, polarization voltage and The state that changes over time. Based on , , With minute changes in aging and operating conditions (also described by the model's internal equations), a theoretical trajectory of the internal equivalent series resistance is calculated. .Compare and First judgment (monotonicity): Check Sequence. In hour, ;exist hour, .because ,Right now The sequence is in The time interval has begun to decrease. The physical consistency judgment is monotonically non-decreasing. Failure. Second judgment (slope range): Assuming In another section Its slope (rate of change) is At the same time, the module calculates The slope of the corresponding segment is Preset fluctuation range setting: This range is set to the theoretical slope. . Calculation example: ,Right now . slope Not here Within the range. The physical consistency check for the slope range failed.

[0027] S24: Due to If the physical consistency check fails (monotonicity failure), the discriminator network outputs a "pseudo" signal (e.g., 0). This signal is used to calculate the GAN's loss function and to update the parameters (e.g., weights and biases) of the generator and discriminator networks via backpropagation. This process (S21-S24) iterates until the generator network outputs a valid signal. Sequence (e.g.) It can reliably pass the physical consistency test of the discriminant network (i.e., monotonically non-decreasing throughout, and the slope is within the fluctuation range of the theoretical trajectory). At this point, the... Sequence and its corresponding operating condition feature vector It is accepted and becomes a record in the virtual battery degradation dataset.

[0028] In the physical consistency judgment in step S23, the battery equivalent circuit differential calculation module solves the battery pack terminal voltage sequence based on the second-order RC equivalent circuit model. Differential equations: The detailed explanations of each parameter in the formula are as follows: The value (V) of the battery pack terminal voltage sequence at time t; The open-circuit voltage (V) represents the battery's state of charge. A function of (%); The value (A) of the pitch motor drive current sequence at time t, with positive for discharging and negative for charging; The internal resistance is ohms (Ohm). The polarization voltage (V) of the first RC network; Let V be the polarization voltage of the second RC network. The operational logic of this equation lies in: the actual output voltage of the battery. It is equal to its ideal open-circuit voltage. Subtract the current Flow through instantaneous ohmic resistance The resulting pressure drop ( ), and then subtract the polarization voltage generated by the short-time polarization effect (described by the first RC network). Finally, subtract the polarization voltage generated by the long-term polarization effect (described by the second RC network). For the parameters in this formula The calculation example is assigned at each second: . From S1 Obtain from sequence The value at time s, . Calculated using the ampere-hour integration method. Assume the battery is initially... 90%, in arrive During period s, current integral (Collected via S1) The sequence is obtained by exact integration (this is an example value), with a nominal capacity of 100Ah. . By checking the battery Correspondence table (quantification of non-numerical data). When it is 89.861%, the corresponding open-circuit voltage It is 26.5V. The initial parameter values ​​set in S23 are used. . and As a state variable, its value is determined by... arrive The differential equation of s is obtained by solving (numerical integration). Assume... hour After 10 seconds of 50A current excitation, the model solution is obtained. At time s, , Substitute the above parameter values ​​into the formula to perform the calculation: The advantage of this equation is that it establishes an externally measurable current. ,Voltage With the internal state of the battery ( , , The instantaneous quantitative relationship between these two parameters forms the basis for calculating the theoretical trajectory in the physical consistency judgment. This calculation result... for The theoretical terminal voltage at time s. This result will be compared with the voltage collected in S1. Sequence in The actual measured value of s (e.g., 25.3V) is compared to verify the accuracy of the parameters of the second-order RC equivalent circuit model. Then by , , (and its aging model) jointly determine that the discriminant network passes validation. time derivative If the condition is always met, then the final physical consistency judgment will be executed.

[0029] Please see Figure 1 and Figure 4 S3: Using the operating condition feature vector in the virtual battery degradation dataset as input and the predicted capacity value as the output label, optimize the gradient boosting decision tree model to obtain the health status prediction model. The specific steps for S3 are as follows: S31: Use the operating condition feature vector in the virtual battery degradation dataset as the training input and the predicted capacity value corresponding to the virtual battery degradation dataset as the output label. S32: The Bayesian optimization algorithm is used to automatically search for the optimal combination of hyperparameters for the gradient boosting decision tree model. The optimal combination of hyperparameters includes the maximum depth of the decision tree, the learning rate, and the subsampling ratio. S33: Configure the gradient boosting decision tree model using the optimal hyperparameter combination, and iteratively train the gradient boosting decision tree model using a virtual battery decay dataset until the root mean square error of the gradient boosting decision tree model on the validation set is lower than the preset convergence threshold. S34: Save the trained gradient boosting decision tree model to obtain the health status prediction model; The optimization process of the gradient boosting decision tree model also includes: In the iterative training process of the gradient boosting decision tree model, a dimensionally normalized regularized loss function is introduced. ; and ; in, Let i be the predicted capacity value for the i-th sample. For the gradient boosting decision tree model, the feature vector of the i-th working condition is... The predicted output, The reference capacity value used for dimensional normalization. For the k-th decision tree, Let be the regularization term for the k-th decision tree. The regularization coefficient is . Let be the number of leaf nodes in the k-th decision tree. Let be the weight of the j-th leaf node in the k-th decision tree. The penalty coefficient for the number of leaf nodes. This is the penalty coefficient for the weight of the leaf nodes. For sample index, For decision tree indexing, Index for leaf nodes; Adjusted using Bayesian optimization algorithm , , And the learning rate, to minimize the regularized loss function. With the goal of improving the gradient boosting decision tree model, we aim to enhance the generalization ability and prediction accuracy of the health status prediction model.

[0030] S31: Divide the virtual battery degradation dataset generated in S2 (e.g., containing 100,000 records) into a training set (80,000 records), a validation set (10,000 records), and a test set (10,000 records). Then, extract the operating condition feature vectors from the dataset... ) as training input. The predicted capacity values ​​from the dataset ( For example, the percentage of battery capacity after a specific future time period (such as 98.5%) can be used as an output label.

[0031] S32: The Bayesian optimization algorithm is used to automatically search for hyperparameter combinations of the Gradient Boosting Decision Tree (GBDT) model. The set hyperparameter search space is: the maximum depth of the decision tree (integer, range [3,8]), the learning rate (continuous, range [0.01,0.2]), and the subsampling ratio (continuous, range [0.6,1.0]). Bayesian optimization process: 1. Randomly select 5 sets of hyperparameter combinations (e.g., [depth 3, learning rate 0.1, ratio 0.8], [depth 5, learning rate 0.05, ratio 0.7], [depth 6, learning rate 0.15, ratio 0.9], [depth 4, learning rate 0.02, ratio 1.0], [depth 7, learning rate 0.1, ratio 0.6]) for initial training. 2. Train the model on the training set using these 5 sets of parameters, and calculate the root mean square error (RMSE) on the validation set. 3. Construct a Gaussian process model to fit these 5 points (hyperparameter combination - RMSE relationship). 4. Based on the Gaussian process model and sampling function (e.g., ExpectedImprovement), select the next set of hyperparameter combinations most likely to improve performance (e.g., [depth 4, learning rate 0.08, subsampling ratio 0.75]). 5. Repeat steps 2-4 for a total of 50 iterations. 6. Finally, select the hyperparameter combination with the lowest RMSE on the validation set, for example: maximum depth of the decision tree = 6, learning rate = 0.07, subsampling ratio = 0.8. S33: Configure the gradient boosting decision tree model using the optimal hyperparameter combination determined in S32 (maximum depth = 6, learning rate = 0.07, subsampling ratio = 0.8). Iteratively train the model using the training set (80,000 trees). After each iteration (i.e., adding a new decision tree), calculate the root mean square error (RMSE) on the validation set (10,000 trees). Set a convergence threshold: To determine the "preset convergence threshold," a series of preliminary experiments were performed to compare model performance and training overhead at different thresholds.

[0032] Table 1. Experimental results of GBDT model convergence threshold and prediction accuracy. As shown in Table 1, the experimental data analysis shows that when the threshold decreases from 0.5% to 0.2%, the MAE significantly decreases (from 0.42% to 0.18%); however, when it decreases from 0.2% to 0.1%, the improvement in MAE diminishes (only 0.09%), but the training time increases significantly (from 65 min to 115 min). Based on the above analysis, 0.2% is selected as the preset convergence threshold for the root mean square error. The iterative training process continues, and training stops when the RMSE on the validation set is below 0.2%. S34: Save the trained gradient boosting decision tree model to obtain the health status prediction model.

[0033] During the iterative training of S33, a dimensionally normalized regularized loss function was introduced. . ;and The detailed explanations of each parameter in the formula are as follows: This is the total loss function; For sample index; The actual predicted capacity value (label) for the i-th sample; Let i be the feature vector of the i-th working condition; For the model pair The predicted output; This is the baseline capacity value used for dimensional normalization, which serves to make the loss function insensitive to the absolute scale of the capacity. This is the regularization coefficient, which controls the overall strength of the regularization term; For decision tree indexing; This is the k-th decision tree; Let be the regularization term for the k-th tree; This is a penalty coefficient for the number of leaf nodes, and its purpose is to control the complexity of the tree. The number of leaf nodes in the k-th tree; This is the penalty coefficient for the weight of the leaf nodes, and its function is to control the smoothness of the output values ​​of the leaf nodes; Index for leaf nodes; Let be the weight of the j-th leaf node in the k-th tree. Loss function. The operation logic: Part 1 It is the standard mean squared error (MSE) loss, but by dividing by Normalization was performed. Part Two It is a regularization term. It also includes two parts: The complexity of the tree is used to penalize it (the more leaves, the greater the penalty). It's L2 regularization, used to penalize excessively large leaf node weights. (also through) (Normalization). For the parameters in this formula in one iteration (e.g., calculating the ...) When the loss of a tree is considered, a calculation example is assigned. Assume there are 2 samples in the current batch. ), and there is already 1 tree in the model ( ): : Set the baseline capacity value. Because and All figures are percentages of capacity (e.g., 98.5%). Set to 100 (i.e., 100%). Sample 1 ( ): Actual value (%), Model Predictions (%). Sample 2 ( ): Actual value (%), Model Predictions (%). Regularization coefficient: determined through Bayesian optimization of S32. The first tree ( ): Parameters: , (Also determined by Bayesian optimization). Assume have Leaf nodes. Leaf node weights (j=1,2,3,4) are respectively: .

[0034] Substitute the above parameter values ​​into the formula to calculate: Calculate the loss component (MSE): ; Calculate the regularization term ( ): ; ; ; The advantage of this formula lies in the introduction of a basis... Dimensional normalization makes the regularization term (especially) and The physical meaning of ) is clearer, its penalty intensity is no longer affected by the capacity unit, and the hyperparameter ( The Bayesian optimization search space is more stable, improving the model's generalization ability and prediction accuracy. This calculation result... This represents the total loss under the current model state (1 tree, 2 samples). The next step in gradient boosting the decision tree (i.e., training the second tree) is... ) will be based on this loss The negative gradient is used as the fitting objective, and the Bayesian optimization process aims to minimize... (On the validation set) Adjust for the target .

[0035] Please see Figure 1 and Figure 5 S4: Collect real-time signals to construct real-time operating condition feature vectors, call the health status prediction model to calculate the future remaining capacity percentage value, compare the future remaining capacity percentage value with the threshold to determine the early warning status signal, and construct structured operation and maintenance instructions.

[0036] Structured operation and maintenance instructions include the percentage of remaining capacity in the future, early warning status signals, and thresholds; The specific steps for S4 are as follows: S41: Real-time acquisition of pitch motor drive current sequence, battery pack terminal voltage sequence and battery ambient temperature, and construction of real-time operating condition feature vector by calculating real-time depth of discharge, charging time and pitch frequency; S42: Input the real-time operating condition feature vector into the health status prediction model, and the health status prediction model outputs the percentage of future remaining capacity within a specified future time period. S43: Set a capacity decay threshold of 80%. Compare the future remaining capacity percentage with the threshold. If the future remaining capacity percentage is lower than the threshold, generate a high-level warning signal. If the future remaining capacity percentage is not lower than the threshold, generate a normal-level warning signal. S44: Combine the future remaining capacity percentage, early warning status signals, and corresponding battery and wind turbine numbers to construct structured operation and maintenance instructions and push them to the cloud platform.

[0037] S41: On the cloud platform operation and maintenance system, perform real-time monitoring of wind turbine pitch battery pack #1, numbered WT02. The system collects data segments from the 10 minutes prior to the current time (e.g., 04:00:00 on November 13, 2025): pitch motor drive current sequence. Battery pack terminal voltage sequence And the battery ambient temperature. The collected battery ambient temperature was -5°C (winter high-latitude wind field). The data was processed using the calculation procedure described in S1. and The calculated real-time depth of discharge (DOD) is 35% (due to the decrease in usable capacity caused by low temperature, the DoD percentage of a single pitch change is higher). The calculated real-time charging time is 3600s (low temperature reduces charging acceptance, extending the charging time). The calculated real-time pitch frequency is 0.01Hz (moderate wind conditions). The real-time operating condition feature vector is constructed as follows: .

[0038] S42: Transfer the real-time operating condition feature vector The input is fed into the health status prediction model (GBDT model) trained and saved in S3. The model (e.g., containing 850 decision trees, consistent with the convergence iteration number set in S3) performs forward computation on the vector, summing the weights of the leaf nodes of all decision trees, and outputting a scalar value. The output value for the percentage of remaining capacity in the future within a specified future time period (e.g., the next 30 days) is 79.5%.

[0039] S43: Set a capacity degradation threshold. To determine the rationality of a threshold of 80%, the following experimental verification process was performed: 50 retired wind farm pitch battery packs (of the same model) with complete EoL (End of Life) capacity records at the time of retirement were selected. The operation and maintenance records, failure rates, and replacement costs of these battery packs in the year prior to retirement were analyzed.

[0040] Table 2 Battery EoL Capacity and Operation and Maintenance Cost-Benefit Analysis As shown in Table 2, data analysis indicates that: 1. Replacing the battery at 85% capacity still results in a relatively high remaining battery life (1500 cycles) and a low failure risk (5%), but premature replacement leads to a negative net benefit (-15,000 RMB), indicating premature investment in maintenance costs. 2. Replacing the battery at 80% capacity keeps the failure risk (20%) under control. Planned maintenance at this point (maintenance cost 10,000 RMB) can prevent subsequent frequent failures, resulting in the highest net benefit (5,000 RMB). 3. Replacing the battery at 75% or 70% significantly increases the failure risk (45%-80%), making unplanned downtime highly likely, leading to emergency maintenance costs (35,000-60,000 RMB) and power generation losses, resulting in a negative net benefit. Based on the above cost-benefit analysis, setting the threshold to 80% has sufficient economic and reliability justification. Compare the future remaining capacity percentage output by S42 (79.5%) with the threshold (80%). Judgment: If the comparison result is "yes", the system generates a high-level warning status signal (defined as "ALERT_H").

[0041] S44: The result of combining S42 and S43, and the corresponding device number. Future remaining capacity percentage = 79.5%. Warning status signal = "ALERT_H". Threshold = 80%. Battery number = 1. Turbine number = WT02. Construct a structured operation and maintenance instruction (e.g., a JSON object): {"timestamp":"2025-11-13T04:00:10Z","turbine_id":"WT02","battery_id":1,"predicted_soh_percent":79.5,"alert_level":"ALERT_H","threshold_percent":80.0} This instruction is pushed to the cloud platform, triggering the operation and maintenance work order system, notifying operation and maintenance personnel to arrange planned inspection and replacement of battery pack 1 of WT02.

[0042] A cloud-based intelligent operation and maintenance system for wind farms, used to execute the aforementioned cloud-based intelligent operation and maintenance method for wind farms, the system comprising: The data acquisition module is used to acquire the pitch motor drive current sequence, battery pack terminal voltage sequence, and battery ambient temperature. The operating condition feature construction module is used to calculate the depth of discharge, charging time and pitch frequency, and combine the depth of discharge, charging time, pitch frequency and battery ambient temperature to construct the operating condition feature vector. The virtual data generation module is used to call the battery equivalent circuit differential calculation module through the discriminant network of the generative adversarial network based on the working condition feature vector, to determine whether the trajectory of the change of the internal equivalent series resistance is monotonically non-decreasing, to perform physical consistency judgment, and to generate a virtual battery degradation dataset. The state prediction modeling module takes the operating condition feature vector in the virtual battery degradation dataset as input, the predicted capacity value as the output label, optimizes the gradient boosting decision tree model, and obtains the health state prediction model. The intelligent operation and maintenance instruction module is used to collect real-time signals to construct real-time operating condition feature vectors, call the health status prediction model to calculate the future remaining capacity percentage, compare the future remaining capacity percentage with the threshold to determine the early warning status signal, and construct structured operation and maintenance instructions.

[0043] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the scope of protection defined by the claims of the present invention.

Claims

1. A cloud platform-based intelligent operation and maintenance method for a wind farm, characterized in that, The method comprises the following steps: S1: Collecting a variable pitch motor driving current sequence, a battery pack terminal voltage sequence and a battery ambient temperature, calculating a discharge depth, a charging time length and a variable pitch frequency, and combining the discharge depth, the charging time length, the variable pitch frequency and the battery ambient temperature to construct a working condition feature vector; S2: Based on the working condition feature vector, calling a battery equivalent circuit differential calculation module through a discriminant network of a generative adversarial network to determine whether an internal equivalent series resistance change trajectory is monotonically non-decreasing, performing physical consistency judgment, and generating a virtual battery degradation data set; S3: Taking the working condition feature vector in the virtual battery degradation data set as input and a capacity value as output label, optimizing a gradient boosting decision tree model to obtain a health state prediction model; S4: Collecting real-time signals to construct real-time working condition feature vectors, calling the health state prediction model to calculate future residual capacity percentage values, comparing the future residual capacity percentage values with a threshold to determine a warning state signal, and constructing a structured operation and maintenance instruction.

2. The cloud platform-based intelligent operation and maintenance method for a wind farm according to claim 1, characterized in that, The virtual battery degradation data set comprises the working condition feature vector and the output label, and the structured operation and maintenance instruction comprises the future residual capacity percentage value, the warning state signal and the threshold.

3. The cloud platform-based intelligent operation and maintenance method for a wind farm according to claim 2, characterized in that, The calculation process of the discharge depth, the charging time length and the variable pitch frequency comprises: Based on the variable pitch motor driving current sequence and the battery pack terminal voltage sequence, the cumulative discharge ampere-hour number of the current discharge cycle is calculated by ampere-hour integration method, and the discharge depth is determined by using the ratio of the cumulative discharge ampere-hour number to the nominal capacity of the battery; The battery pack terminal voltage sequence is monitored, the charging starting point and the charging cutoff point are accurately identified through voltage slope and current direction judgment, the time interval from the charging starting point to the charging cutoff point is counted to obtain the charging time length; The fluctuation characteristics of the variable pitch motor driving current sequence are analyzed, the time-frequency spectrum of the variable pitch motor driving current sequence is extracted by using short-time Fourier transform, and the region with high-frequency energy concentration is identified from the time-frequency spectrum to count the number of variable pitch actions per unit time, and the variable pitch frequency is calculated.

4. The cloud platform-based intelligent operation and maintenance method for a wind farm according to claim 3, characterized in that, The step S2 specifically comprises: S21: initializing a generative network of the generative adversarial network, the generative network receiving a random noise vector and combining the working condition feature vector as a conditional input to generate an initial change trajectory; S22: inputting the initial change trajectory and the working condition feature vector into the discriminant network; S23: the discriminant network calling the battery equivalent circuit differential calculation module to perform the physical consistency judgment on the initial change trajectory; S24: updating the parameters of the generative network and the discriminant network according to the result of the physical consistency judgment until the internal equivalent series resistance change trajectories output by the generative network all meet the physical consistency judgment, forming the virtual battery degradation data set.

5. The cloud platform-based intelligent operation and maintenance method for a wind farm according to claim 4, characterized in that, The specific process of the physical consistency judgment is: The battery equivalent circuit differential calculation module sets initial parameters of a battery second-order RC equivalent circuit model based on the battery ambient temperature and the depth of discharge in the operating condition feature vector, the initial parameters including ohmic internal resistance, first polarization internal resistance, second polarization internal resistance, first polarization capacitance, and second polarization capacitance; The variable-pitch motor drive current sequence in the operating condition feature vector is taken as an excitation input of the second-order RC equivalent circuit model, and a theoretical internal equivalent series resistance variation trajectory is calculated by real-time solving of state differential equations of the second-order RC equivalent circuit model; The generated internal equivalent series resistance variation trajectory generated by the generative adversarial network is compared with the theoretical internal equivalent series resistance variation trajectory, to determine whether the internal equivalent series resistance variation trajectory generated by the generative adversarial network is monotonically non-decreasing throughout and whether the variation slope is within a preset fluctuation range of the theoretical internal equivalent series resistance variation trajectory, and if so, the physical consistency is determined to be passed.

6. The cloud platform-based intelligent operation and maintenance method for a wind farm according to claim 5, characterized in that, The steps of S3 are specifically: S31: Taking the operating condition feature vector in the virtual battery degradation data set as a training input and taking the predicted capacity value corresponding to the virtual battery degradation data set as the output label; S32: Automatically searching for an optimal hyperparameter combination of a gradient boosting decision tree model by using a Bayesian optimization algorithm, the optimal hyperparameter combination including the maximum depth of a decision tree, a learning rate, and a subsampling ratio; S33: Configuring the gradient boosting decision tree model by using the optimal hyperparameter combination, and iteratively training the gradient boosting decision tree model by using the virtual battery degradation data set until the root mean square error of the gradient boosting decision tree model on a validation set is lower than a preset convergence threshold; S34: Saving the trained gradient boosting decision tree model to obtain a health state prediction model.

7. The cloud platform-based intelligent operation and maintenance method for a wind farm according to claim 6, characterized in that, The steps of S4 are specifically: S41: Real-time collecting the variable-pitch motor drive current sequence, the battery pack terminal voltage sequence, and the battery ambient temperature, and constructing a real-time operating condition feature vector by calculating the real-time depth of discharge, the charging duration, and the variable-pitch frequency; S42: Inputting the real-time operating condition feature vector into the health state prediction model, and outputting a future residual capacity percentage value in a future specified time period by the health state prediction model; S43: Setting a capacity degradation threshold, the threshold being 80%, comparing the future residual capacity percentage value with the threshold, and if the future residual capacity percentage value is lower than the threshold, generating a high-level early warning state signal, and if the future residual capacity percentage value is not lower than the threshold, generating a regular-level early warning state signal; S44: Combining the future residual capacity percentage value, the early warning state signal, and corresponding battery numbers and wind turbine numbers, constructing a structured operation and maintenance instruction, and pushing the instruction to a cloud platform. 8.The cloud platform-based wind farm intelligent operation and maintenance method of claim 5, characterized in that, The battery equivalent circuit differential calculation module solves the battery pack terminal voltage sequence based on a differential equation of the second-order RC equivalent circuit model: ​ ； wherein, is a value of the battery pack terminal voltage sequence at time t, is a value of the pitch motor drive current sequence at time t, is a state of charge of the battery, is an open circuit voltage corresponding to the state of charge, is an ohmic internal resistance, is a polarization voltage of a first RC network, is a polarization voltage of a second RC network; The internal equivalent series resistance change trajectory The internal equivalent series resistance change trajectory The internal equivalent series resistance change trajectory The internal equivalent series resistance change trajectory The internal equivalent series resistance change trajectory The internal equivalent series resistance change trajectory The internal equivalent series resistance change trajectory 9.The cloud platform-based wind farm intelligent operation and maintenance method of claim 6, characterized in that, The optimization process of the gradient boosting decision tree model further includes: In the iterative training process of the gradient boosting decision tree model, a dimensionless regularization loss function is introduced ; and ; wherein, is the predicted capacity value for the i-th sample, is the predicted output of the gradient boosting decision tree model for the i-th working condition feature vector , is the reference capacity value for dimension normalization, is the k-th decision tree, is the regularization term of the k-th decision tree, is the regularization coefficient, is the number of leaf nodes of the k-th decision tree, is the weight of the j-th leaf node of the k-th decision tree, is the penalty coefficient of the number of leaf nodes, is the penalty coefficient of the leaf node weight, is the sample index, is the decision tree index, is the leaf node index. adjusting the Bayesian optimization algorithm to minimize the regularized loss function , the , the and the learning rate, to minimize the regularized loss function to optimize the gradient boosting decision tree model, improve the generalization ability and prediction accuracy of the health state prediction model.

10. A cloud platform-based intelligent operation and maintenance system for a wind farm, characterized in that, The system is used to implement the cloud platform-based intelligent operation and maintenance method of a wind farm according to any one of claims 1-9, and the system includes: The collection module is configured to collect a variable pitch motor driving current sequence, a battery pack terminal voltage sequence and the battery ambient temperature; The working condition feature construction module is configured to calculate a depth of discharge, a charging time length and the variable pitch frequency, and combine the depth of discharge, the charging time length, the variable pitch frequency and the battery ambient temperature to construct the working condition feature vector; The virtual data generation module is configured to call the battery equivalent circuit differential calculation module through a discriminant network of the generative adversarial network based on the working condition feature vector to determine whether an internal equivalent series resistance change trajectory is monotonically non-decreasing, perform physical consistency determination, and generate the virtual battery degradation data set; The state prediction modeling module is configured to input the working condition feature vector in the virtual battery degradation data set and output a predicted capacity value as a label to optimize a gradient boosting decision tree model and obtain a health state prediction model; The intelligent operation and maintenance instruction module is configured to collect real-time signals to construct the real-time working condition feature vector, call the health state prediction model to calculate a future residual capacity percentage value, compare the future residual capacity percentage value with a threshold to determine a warning state signal, and construct a structured operation and maintenance instruction.

Images