Operation data analysis and prediction system based on offshore wind turbine generator

Abstract

The invention relates to the technical field of wind turbine generator data analysis, and discloses an offshore wind turbine generator operation data analysis and prediction system. The system comprises a marine environment data integration module for collecting data to generate a multi-source time-space synchronization data set; the multi-modal feature fusion module is used for extracting cross-modal correlation features to generate a high-dimensional fusion feature tensor; the dynamic fault prediction module is used for constructing a two-way gating circulation network model to predict the degradation probability and the residual life of key components of the equipment; and the self-adaptive optimization control module is used for constructing a multi-target dynamic programming model to optimize a fan operation strategy. In addition, the system is further provided with a feedback correction module for correcting prediction model parameters, and a virtual sensor module based on a physical information neural network is used for monitoring tower stress and diagnosing sensor faults. According to the system, comprehensive monitoring, accurate fault prediction and optimal control of the offshore wind turbine generator are realized, the operation efficiency, reliability and safety of the wind turbine generator are effectively improved, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine data analysis, and in particular to an analysis and prediction system based on offshore wind turbine operation data. Background Art

[0002] With the growing global demand for clean energy, offshore wind power, with its abundant resources and stable wind speeds, has become a key development direction in the renewable energy sector. Operating in a complex marine environment, offshore wind turbines face numerous challenges. This places extremely high demands on their operational stability, reliability, and safety, prompting continuous technological innovation.

[0003] The complex and ever-changing marine environment, with harsh conditions such as strong winds, huge waves, and salt spray posing significant challenges to the operation of offshore wind turbines. Strong winds can overload turbine blades, causing fatigue damage or even breakage. The impact of large waves can affect the stability of wind turbine towers, increasing the risk of structural damage. Salt spray is corrosive, accelerating the aging of equipment components and posing a serious threat to the safe operation of wind turbines.

[0004] Current wind turbine operational data monitoring systems face numerous shortcomings. Existing monitoring systems collect limited data types, making it difficult to fully reflect the operational status of wind turbines and the marine environment. Furthermore, data accuracy and synchronization are poor, and data collected by different sensors exhibit temporal and spatial deviations, making them inadequate for subsequent analysis. For example, the sampling frequencies of some wind speed sensors and vibration sensors are inconsistent, resulting in inaccurate data matching and hindering the assessment of equipment operational status.

[0005] Inadequate fault prediction capabilities are a key challenge facing offshore wind turbines. Traditional fault prediction methods are mostly based on single data sets or simple models, failing to fully exploit the potential correlations between multiple data sources. In complex marine environments, equipment failures are often the result of multiple factors, making it difficult for a single data point to accurately predict the timing and extent of a failure. This makes it difficult for operators to implement effective maintenance measures in a timely manner, resulting in extended equipment downtime and increased repair costs.

[0006] Wind turbine control strategies need to be optimized. Existing control strategies are often based on fixed rules or empirical evidence, failing to fully account for the dynamic nature of the marine environment. Under varying wind speeds and wave conditions, the turbine's power output, yaw angle, and pitch rate cannot be optimally adjusted, resulting in low power generation efficiency and increased mechanical fatigue losses. For example, when wind speed suddenly changes, the turbine cannot adjust the blade pitch angle in a timely manner, reducing wind energy utilization efficiency and subjecting the equipment to unnecessary impact.

[0007] Furthermore, offshore wind turbine sensor fault detection and handling mechanisms are imperfect. Sensors are prone to malfunction in harsh marine environments. Sensor failure can lead to data loss or errors, impacting the normal operation of the entire monitoring and control system. Existing fault detection methods are slow to respond and fail to detect and address sensor failures in a timely manner, reducing system reliability. Summary of the Invention

[0008] The purpose of the present invention is to provide an analysis and prediction system based on offshore wind turbine operation data to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a system for analyzing and predicting offshore wind turbine operation data, the system comprising:

[0010] The marine environment data integration module is used to collect wind turbine operating data and marine environment parameters to generate a multi-source spatiotemporal synchronized data set;

[0011] A multimodal feature fusion module is used to map the multi-source spatiotemporal synchronized dataset into a graph neural network architecture, extract cross-modal correlation features of equipment status, wave load, and salt spray concentration through a node embedding and edge weight dynamic allocation algorithm, and generate a high-dimensional fused feature tensor;

[0012] A dynamic fault prediction module is used to build a bidirectional gated recurrent network model based on a spatiotemporal attention mechanism, perform temporal dependency modeling on the high-dimensional fused feature tensor, and output the degradation probability and remaining life prediction value of key components of the equipment;

[0013] The adaptive optimization control module is used to build a multi-objective dynamic programming model based on the predicted values, and uses a non-dominated sorting genetic algorithm to optimize the wind turbine power output, yaw angle and pitch rate to generate an anti-disturbance operation strategy.

[0014] Preferably, the multi-source spatiotemporal synchronization data set is subjected to denoising and alignment processing based on wavelet transform and adaptive filtering algorithm, specifically including:

[0015] Extract vibration signals, wind speed and sea temperature series from multi-source spatiotemporal synchronized data, and generate high-frequency noise components and low-frequency trend components through multi-resolution wavelet decomposition;

[0016] Adaptive filter is designed using the minimum mean square error criterion to dynamically suppress high frequency noise.

[0017] The timestamp alignment algorithm is used to interpolate and synchronize the denoised low-frequency components with the device status data.

[0018] Preferably, generating a high-dimensional fusion feature tensor includes:

[0019] Define device nodes, environment nodes, and load nodes as graph structure vertices and construct the device-environment-load association matrix;

[0020] A graph attention mechanism is used to aggregate features of adjacent nodes and generate node-level spatiotemporal embedding vectors.

[0021] The association strength between nodes is optimized through a dynamic edge weight update strategy, and a feature tensor containing nonlinear coupling relationships is output.

[0022] Preferably, the construction of a bidirectional gated recurrent network model based on a spatiotemporal attention mechanism includes:

[0023] In the time dimension, a bidirectional GRU network is used to capture the forward and backward dependencies of device states;

[0024] A spatial attention layer is introduced to calculate the weight distribution of each sensor data and focus on abnormal signals of key components;

[0025] The feature segments in the historical sequence that are strongly correlated with the current fault are filtered through the temporal attention layer to generate a degradation probability time series curve.

[0026] Preferably, the construction of the multi-objective dynamic programming model includes:

[0027] The objective function is defined as a weighted combination of maximizing power generation efficiency, minimizing mechanical fatigue loss, and minimizing the sum of squared yaw errors;

[0028] The constraints are set as the maximum allowable pitch angle, rated power threshold, and safety threshold range of wave impact force;

[0029] The elite retention strategy and the congestion comparison operator are used to optimize the Pareto solution set and screen the anti-disturbance operation strategy.

[0030] Preferably, the system further comprises:

[0031] The feedback correction module is used in an incremental learning framework based on a sliding time window. It combines Kalman filtering and Bayesian inference algorithms to perform online corrections on prediction model parameters and generate an adaptively updated fault prediction model. Specifically, it includes:

[0032] Divide the real-time data stream based on the sliding window and extract the prediction error sequence within the window as the feedback signal;

[0033] Design the Kalman gain matrix to recursively update the network weights to reduce the dependence on historical data;

[0034] Bayesian posterior probability distribution is used to adjust model hyperparameters to achieve adaptive calibration of the prediction model.

[0035] Preferably, the multi-source spatiotemporal synchronized dataset includes:

[0036] Integrate unit electrical parameters collected by the SCADA system, fiber optic sensor vibration data, and wave spectrum data provided by weather buoys;

[0037] A tensor completion algorithm is used to jointly interpolate missing segments of heterogeneous data to construct a complete spatiotemporal data cube;

[0038] Dimensional differences were eliminated by combining Z-score standardization and range normalization.

[0039] Preferably, the system further comprises:

[0040] Construct a virtual sensor module based on physical information neural network to estimate the tower stress distribution that cannot be directly measured in real time;

[0041] The estimated results are compared with the measured data to generate a residual signal, which triggers sensor fault diagnosis and data source switching instructions.

[0042] Preferably, the present invention further includes an electronic device, comprising:

[0043] at least one processor; and a memory communicatively coupled to the at least one processor;

[0044] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the operations of the above-mentioned offshore wind turbine operation data analysis and prediction system.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] In terms of data processing, the marine environmental data integration module integrates wind turbine operating data with marine environmental parameters to generate a multi-source, spatiotemporally synchronized dataset. A tensor completion algorithm is used to jointly interpolate missing segments of heterogeneous data to construct a complete spatiotemporal data cube, ensuring data integrity. A combined strategy of Z-score standardization and range normalization eliminates dimensional discrepancies, improving data quality and laying a solid foundation for subsequent precise analysis. This fusion of multi-source data provides a more comprehensive and accurate description of wind turbine operating status, avoiding analytical biases caused by missing or inconsistent data.

[0047] The multimodal feature fusion module utilizes a graph neural network architecture to extract cross-modal correlation features between equipment status, wave loads, and salt spray concentration, generating a high-dimensional fused feature tensor. This process constructs a correlation matrix by defining different types of nodes, aggregates features using a graph attention mechanism, and optimizes correlation strength with a dynamic edge weight update strategy. This fully exploits the complex nonlinear coupling relationships between multi-source data. Compared to traditional single-feature analysis methods, this method captures more critical information hidden in the data, providing richer and more valuable data support for fault prediction and operational optimization.

[0048] The dynamic fault prediction module uses a bidirectional gated recurrent network model with a spatiotemporal attention mechanism to model the temporal dependencies of high-dimensional fused feature tensors. In the temporal dimension, the bidirectional GRU network captures the forward and backward dependencies of equipment states, enabling a more comprehensive understanding of the changing trends of equipment operating states over time. The spatial attention layer focuses on abnormal signals from key components, while the temporal attention layer selects feature fragments that are strongly correlated with the current fault, thereby accurately outputting the degradation probability and remaining life prediction values ​​of key equipment components. This prediction method greatly improves the accuracy and timeliness of fault prediction. Operations and maintenance personnel can formulate maintenance plans in advance based on the prediction results, rationally arrange maintenance resources, effectively reduce downtime caused by sudden equipment failures, reduce maintenance costs, and improve the reliability and availability of wind turbines.

[0049] The adaptive optimization control module constructs a multi-objective dynamic programming model based on the predicted values ​​and uses a non-dominated sorting genetic algorithm to optimize the wind turbine's power output, yaw angle, and pitch rate to generate an anti-disturbance operation strategy. This strategy aims to maximize power generation efficiency, minimize mechanical fatigue loss, and minimize the sum of squared yaw errors, while taking into account constraints such as the maximum allowable pitch angle, rated power threshold, and wave impact force safety threshold. Through optimized control, wind turbines can better adapt to changes in the marine environment, while ensuring safe and stable operation of the equipment, improving power generation efficiency, reducing mechanical fatigue loss, and extending equipment service life. For example, under different wind speeds and wave conditions, the wind turbine can automatically adjust the pitch angle and yaw angle to achieve efficient capture of wind energy and reduce unnecessary wear and tear on the equipment.

[0050] The feedback correction module uses an incremental learning framework based on a sliding time window, combining Kalman filtering and Bayesian inference algorithms to perform online corrections to the prediction model parameters. The sliding window extracts the prediction error sequence as feedback, recursively updates the network weights using the Kalman gain matrix to reduce reliance on historical data, and uses the Bayesian posterior probability distribution to adjust model hyperparameters, achieving adaptive calibration of the prediction model. This enables the model to adjust in real time as wind turbine operating conditions and environmental conditions change, maintaining consistently high prediction accuracy and further enhancing the system's reliability and adaptability.

[0051] A virtual sensor module based on a physical information neural network provides real-time estimation of tower stress distribution, which cannot be directly measured. By comparing this with measured data, it generates a residual signal that triggers sensor fault diagnosis and data source switching. This module not only effectively monitors tower stress but also promptly detects sensor failures and initiates data source switching, ensuring the accuracy and continuity of system data. This improves the reliability and stability of the entire monitoring system and ensures the safe and stable operation of wind turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a working principle diagram of the offshore wind turbine operation data analysis and prediction system according to the present invention;

[0053] Figure 2 Flowchart for high-dimensional fused feature tensor generation;

[0054] Figure 3 A flowchart for a bidirectional gated recurrent network model based on spatiotemporal attention mechanism;

[0055] Figure 4 This is the workflow diagram of the virtual sensor module based on physical information neural network. DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] See also Figure 1-Figure 4 The present invention provides an offshore wind turbine operating data analysis and prediction system, which aims to achieve operational status monitoring, fault prediction, and optimized control of offshore wind turbines through the collaborative work of multiple modules. The overall implementation scheme is as follows:

[0058] The system first collects wind turbine operating data and marine environmental parameters through the marine environmental data integration module. The operating data includes turbine electrical parameters and vibration data, while the marine environmental parameters include wave spectrum data and meteorological data. These data are integrated to create a multi-source, spatiotemporally synchronized dataset. This dataset, integrating data from various sources and across different temporal and spatial dimensions, provides a comprehensive foundation for subsequent analysis.

[0059] The multimodal feature fusion module receives multi-source, spatiotemporally synchronized datasets and maps them into a graph neural network architecture. Within this architecture, a node embedding and edge weighting dynamic allocation algorithm extracts cross-modal correlation features between different modal data, such as equipment status, wave loads, and salt spray concentration. These features are processed to generate a high-dimensional fused feature tensor, effectively integrating the feature information from multiple sources and providing richer data support for subsequent predictions.

[0060] The dynamic fault prediction module builds a bidirectional gated recurrent network model based on a spatiotemporal attention mechanism. This model models the temporal dependencies of high-dimensional fused feature tensors, fully accounting for data correlations in both temporal and spatial dimensions. It outputs degradation probabilities and predicted remaining lifespans for key equipment components. This approach enables early prediction of failures in key wind turbine components, providing a crucial basis for maintenance decisions.

[0061] The adaptive optimization control module constructs a multi-objective dynamic programming model based on the predicted values ​​output by the dynamic fault prediction module. A non-dominated sorting genetic algorithm is used to optimize the wind turbine's power output, yaw angle, and pitch rate, generating an anti-disturbance operation strategy. This strategy enables the wind turbine to maintain stable operation in complex marine environments, improve power generation efficiency, and reduce equipment losses.

[0062] The implementation of the present invention will be further described below with reference to Examples 1 to 6.

[0063] Example 1:

[0064] During system operation, data quality is crucial to the accuracy of subsequent analysis and prediction. Therefore, denoising and alignment are required for multi-source spatiotemporally synchronized datasets, while ensuring the integrity and standardization of the datasets.

[0065] Multi-source spatiotemporally synchronized data includes vibration signals, wind speed, and seawater temperature series. Leveraging the multi-resolution characteristics of the wavelet transform, these series undergo multi-resolution wavelet decomposition. During the wavelet decomposition process, the data is broken down into high-frequency noise components and low-frequency trend components. The high-frequency noise component primarily contains noise information within the data, while the low-frequency trend component retains the data's primary trend characteristics. For example, after wavelet decomposition of a vibration signal, the high-frequency component may contain subtle interference noise generated by equipment vibration, while the low-frequency component reflects the overall vibration trend.

[0066] To dynamically suppress high-frequency noise, an adaptive filter is designed using the minimum mean square error (MMSE) criterion. The principle behind this criterion is to minimize the mean square error (MSE) between the filter output and the desired output by continuously adjusting the filter parameters. In practical applications, the desired output can be determined by analyzing the original signal or by referencing other relevant data. The adaptive filter dynamically adjusts its parameters based on changes in the input signal, effectively suppressing noise in high-frequency components and thus improving data quality.

[0067] After denoising, the low-frequency components and device status data must be interpolated and synchronized. Using a timestamp alignment algorithm, the denoised low-frequency components and device status data are processed based on the data's timestamp information. Interpolation is used to supplement and synchronize missing or unsynchronized data points in the time series. For example, if wind speed data is missing at a certain moment, the wind speed value at that moment can be estimated using the interpolation algorithm based on the wind speed data from the preceding and subsequent moments. This ensures that all data are synchronized in time, providing an accurate data foundation for subsequent analysis and processing.

[0068] The multi-source spatiotemporal synchronized dataset integrates a variety of heterogeneous data, including unit electrical parameters collected by the SCADA system, vibration data from fiber-optic sensors, and wave spectrum data from meteorological buoys. Because various factors can lead to missing data during the data collection process, a tensor completion algorithm is used to jointly interpolate missing segments of heterogeneous data. Based on the low-rank nature of tensors, the tensor completion algorithm estimates the value of missing data by analyzing and learning from known data. For example, in a three-dimensional spatiotemporal data cube, if a certain type of data is missing at a certain time or location, the tensor completion algorithm can use relevant data from other times and locations to algorithmically calculate an estimate of the missing data, thereby constructing a complete spatiotemporal data cube.

[0069] In order to eliminate the dimensional differences between different data, a combination of Z-score standardization and range normalization strategy is adopted. The formula for Z-score standardization is: Where x is the original data, μ is the mean of the data, and σ is the standard deviation of the data. Through Z-score normalization, the data is converted into standard normal distribution data with a mean of 0 and a standard deviation of 1. The formula for range normalization is: Where x is the original data, x min is the minimum value in the data, x max is the maximum value in the data. After range normalization, the data is mapped to the interval [0, 1]. The combination of these two methods can effectively eliminate dimensional differences between different data, making different types of data comparable and facilitating subsequent analysis and model training.

[0070] Example 2:

[0071] In the multimodal feature fusion module, to construct a graph structure reflecting the relationships between data, device nodes, environment nodes, and load nodes are defined as graph vertices. For example, device nodes can represent various components of a wind turbine, such as blades, gearboxes, and generators; environment nodes can represent various factors in the marine environment, such as wind speed, wave height, and salt spray concentration; and load nodes are used to describe the various loads acting on the wind turbine, such as wave loads and wind loads. Based on the actual relationships between these nodes, a device-environment-load association matrix is ​​constructed. The elements in the association matrix represent the connection relationship and association strength between nodes. If there is a direct association between two nodes, the corresponding element value is non-zero, and the size of the value reflects the strength of the association.

[0072] A graph attention mechanism is used to aggregate features of adjacent nodes. The core idea of ​​the graph attention mechanism is to determine the importance of each node in the feature aggregation process by calculating the attention weights between nodes. For each node, a corresponding attention weight is calculated based on the feature information of its adjacent nodes. For example, for a device node, it will pay attention to the features of its adjacent environmental nodes and load nodes. By calculating the attention weights, the node features that have a greater impact on it will be more effectively aggregated. During the specific calculation process, some learnable parameters are used to adjust the calculation method of the attention weights, thereby generating more representative node-level spatiotemporal embedding vectors. These embedding vectors fuse the feature information of the node itself and its adjacent nodes, and can better reflect the characteristics of the node in the entire graph structure.

[0073] To further optimize the strength of the associations between nodes, a dynamic edge weight update strategy is used to adjust the edge weights in the association matrix. During wind turbine operation, the associations between equipment status, environmental factors, and loads are not fixed, but are dynamically adjusted over time and with changes in operating conditions. The dynamic edge weight update strategy continuously adjusts the edge weights in the association matrix based on real-time data and model feedback. For example, when wind speed changes significantly, the blade load and equipment status of the wind turbine may be significantly affected. In this case, it is necessary to increase the edge weights between the wind speed node and related equipment nodes and load nodes to highlight the changes in this association. After the dynamic edge weight update, the output feature tensor contains more accurate nonlinear coupling relationships, which can provide more valuable information for subsequent fault prediction and analysis.

[0074] Example 3:

[0075] In the dynamic fault prediction module, building a bidirectional gated recurrent network model based on the spatiotemporal attention mechanism is the core of achieving accurate fault prediction.

[0076] In the time dimension, a bidirectional GRU (Gated Recurrent Unit) network is used to capture the forward and backward dependencies of the equipment status. The bidirectional GRU network consists of a forward GRU and a backward GRU. The forward GRU starts from the beginning of the data sequence and processes the data at each moment in turn to learn the forward time dependency of the data; the backward GRU starts from the end of the data sequence and processes the data in reverse to learn the reverse time dependency of the data. For example, when monitoring the vibration data of wind turbine blades, the forward GRU can learn the trend of blade vibration gradually changing over time, while the backward GRU can capture the impact of abnormal changes that may occur in the vibration data in the later stage. By splicing the outputs of the forward GRU and the backward GRU, the forward and backward dependencies of the equipment status in the time dimension can be fully captured, providing richer time series information for subsequent fault analysis.

[0077] A spatial attention layer is introduced to calculate the weight distribution of each sensor data. In a wind turbine, multiple sensors are installed to monitor information at different locations and parameters. However, the importance of different sensor data for fault prediction varies. The spatial attention layer calculates the weight of each sensor data in fault prediction based on the characteristics of the sensor data. For example, for a sensor monitoring gearbox temperature and a sensor monitoring tower vibration, the weight of the gearbox temperature sensor data may be relatively high when determining a gearbox fault. In this way, the spatial attention layer can focus on abnormal signals from key components, highlight sensor data that has a significant impact on fault prediction, and improve the accuracy of fault prediction.

[0078] The temporal attention layer is used to filter feature segments in the historical sequence that are strongly correlated with the current fault. During equipment operation, not all segments in the historical data are equally correlated with the current fault. The temporal attention layer analyzes the historical data sequence based on the current data and the model's learning results, calculating the degree of correlation between each historical segment and the current fault, known as the attention weight. It then uses these weights to filter feature segments that are strongly correlated with the current fault. For example, when predicting generator faults, the temporal attention layer may find that vibration data following a sudden change in generator load within a certain period of time in the past is strongly correlated with the current generator fault, thereby filtering out these data segments. In this way, the generated degradation probability time series curve can more accurately reflect the actual degradation of key equipment components, providing a reliable basis for predicting the equipment's remaining life.

[0079] Embodiment 4:

[0080] In the adaptive optimization control module, the objective function is defined as a weighted combination of maximizing power generation efficiency, minimizing mechanical fatigue loss, and minimizing the sum of squared yaw errors. Maximizing power generation efficiency is intended to improve the wind turbine's power generation capacity and generate more electricity. Minimizing mechanical fatigue loss helps extend the equipment's service life and reduce maintenance costs. Minimizing the sum of squared yaw errors allows the wind turbine to better track wind direction and improve wind energy capture efficiency. The objective function can be expressed as: Where J is the objective function value, w1, w2, and w3 are the weight coefficients corresponding to power generation efficiency, mechanical fatigue loss, and yaw error square sum, respectively. actual is the actual generated power, P max is the maximum power generation, F fatigue is the mechanical fatigue loss value, F threshold is the fatigue loss threshold, e yaw is the yaw error, n is the number of sampling points. The weight coefficients w1, w2, and w3 are set according to actual needs and importance to balance the relationship between different objectives.

[0081] The constraints are set as the maximum allowable pitch angle, rated power threshold, and safety threshold range of wave impact force. The maximum allowable pitch angle limits the adjustment range of the blades to prevent damage to the equipment due to excessive blade adjustment. The rated power threshold ensures that the wind turbine operates within a safe and stable power range to avoid overload. The safety threshold range of wave impact force takes into account the impact of waves on wind turbines in the marine environment to ensure that the wind turbine does not exceed the safety limit when subjected to wave impact force. For example, the maximum allowable pitch angle may be set to [0,90°], and the rated power threshold is [0,P rated The safety threshold range of wave impact force is determined by the design of the wind turbine and the actual marine environment. In actual operation, the operating parameters of the wind turbine must meet these constraints, otherwise the corresponding protection mechanism or adjustment strategy will be triggered.

[0082] An elite retention strategy and a crowding comparison operator are used to optimize the Pareto solution set and screen for disturbance-resistant operation strategies. The Pareto solution set contains multiple non-inferior solutions that satisfy different trade-offs. The elite retention strategy involves retaining the best individuals from the current generation to the next generation during the iterative process of the genetic algorithm, thereby preventing the loss of excellent individuals during evolution. The crowding comparison operator is used to measure the degree of crowding between individuals in the Pareto solution set, selecting solutions that are more evenly distributed and farther away from other individuals. For example, when calculating the crowding of an individual, the distance between an individual and its neighbors in the target space can be used to measure this. By combining these two methods, it is possible to screen for optimal disturbance-resistant operation strategies within the Pareto solution set, enabling wind turbines to improve power generation efficiency while ensuring safe and stable operation in complex marine environments and reducing various losses.

[0083] Example 5:

[0084] In order to improve the accuracy and adaptability of the prediction model, the system sets up a feedback correction module, which is based on the incremental learning framework of the sliding time window, combined with the Kalman filter and Bayesian inference algorithm to perform online correction of the prediction model parameters.

[0085] The real-time data stream is segmented based on a sliding window. This refers to the various monitoring data continuously generated by wind turbines during operation, such as equipment status data and environmental data. The size of the sliding window is set based on actual conditions; for example, it can be set to include data from the past 10 minutes. Within each time window, a prediction error sequence is extracted as a feedback signal. The prediction error sequence is the difference between the model's predicted values ​​and the actual measured values. By analyzing these errors, we can understand the model's prediction deviations and provide a basis for correcting model parameters. For example, if the model's predicted generator temperature deviates significantly from the actual measured temperature within a certain window, this deviation is included in the prediction error sequence.

[0086] A Kalman gain matrix is ​​designed to recursively update the network weights. Kalman filtering is a commonly used state estimation method that optimizes the estimate of the system state by continuously fusing new measurement data with previous estimation results. In this system, the network weights are treated as the system state, and the Kalman gain matrix is ​​used to adjust the weight updates. The calculation of the Kalman gain matrix is ​​based on factors such as the current measurement noise and the system noise covariance. For example, when the measurement data is noisy, the Kalman gain matrix is ​​adjusted accordingly, making the weight updates more conservative and reducing the impact of noise on the weight updates. This recursive update method can reduce the dependence on historical data, enable the model to adapt to new data changes more quickly, and improve the real-time performance and accuracy of the model.

[0087] The Bayesian posterior probability distribution is used to adjust the model hyperparameters. Bayesian inference is a probability-based reasoning method that uses prior knowledge and new observations to update the probability estimates of unknown parameters. In this system, the model hyperparameters are regarded as unknown parameters and adjusted using the Bayesian posterior probability distribution. For example, for hyperparameters such as the learning rate and the number of hidden layer nodes in a neural network, a prior probability distribution is first set based on experience or previous experiments. Then, combined with new observations (such as the prediction error sequence), the posterior probability distribution is calculated using the Bayesian formula. Based on the posterior probability distribution, the optimal hyperparameter value is selected to achieve adaptive calibration of the prediction model. In this way, the model can automatically adjust the hyperparameters under different operating conditions, thereby improving the model's predictive performance and adaptability.

[0088] Example 6:

[0089] In the system, a virtual sensor module based on physical information neural network is constructed to perform real-time estimation of the tower stress distribution that cannot be directly measured, and realize sensor fault diagnosis and data source switching by comparing with the measured data.

[0090] A virtual sensor is constructed based on a physical information neural network. A physical information neural network is a model that combines physical principles with neural networks. When estimating the stress distribution of a tower, a corresponding physical model is first established based on the physical structure and mechanical principles of the tower. For example, based on the material properties, geometric shape, and load conditions of the tower, mechanical formulas are used to calculate the theoretical stress values ​​of the tower at different positions and working conditions. The information from these physical models is then integrated into the neural network. The neural network adjusts its own parameters by learning from a large amount of actual operating data, including environmental data related to tower stress (such as wind speed and direction) and equipment operating data (such as blade speed and generator power), to more accurately estimate the tower stress distribution. Compared with traditional neural networks, physical information neural networks utilize prior knowledge of physical models, which can improve the accuracy and generalization ability of the model when data is limited.

[0091] The tower stress distribution estimated by the virtual sensor is compared with the measured data to generate a residual signal. The measured data is obtained from sensors installed on the tower, such as strain gauges. The residual signal is the difference between the estimated and measured values. If the residual signal fluctuates within a certain range, it indicates that the virtual sensor's estimate is consistent with the actual situation and the sensor is functioning properly. If the residual signal exceeds the set threshold, it may indicate a sensor failure or other abnormality. For example, if the absolute value of the residual signal exceeds a set threshold for multiple consecutive moments, it can be determined that the sensor is faulty.

[0092] Once an abnormality in the residual signal is detected, the system triggers sensor fault diagnosis and data source switching instructions. During the fault diagnosis process, the residual signal's characteristics, change trends, and other related data are further analyzed to determine the type of fault and possible causes. For example, by analyzing the residual signal's change rate, periodicity, and other characteristics, it can be determined whether the fault is a hardware failure of the sensor itself or a data anomaly caused by external interference. If a sensor failure is determined, the system automatically switches the data source and obtains data from a backup sensor or other relevant data source to ensure normal system operation and continuous monitoring of the tower stress distribution. At the same time, the system records the fault information for subsequent maintenance and analysis. In this way, the reliability and stability of the system are improved, ensuring that sensor failures can be detected and handled in a timely manner during wind turbine operation.

[0093] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0094] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A system for analyzing and predicting offshore wind turbine operation data, characterized in that: The system comprises: The marine environment data integration module is used to collect wind turbine operating data and marine environment parameters to generate a multi-source spatiotemporal synchronized data set; A multimodal feature fusion module is used to map the multi-source spatiotemporal synchronized dataset into a graph neural network architecture, extract cross-modal correlation features of equipment status, wave load, and salt spray concentration through a node embedding and edge weight dynamic allocation algorithm, and generate a high-dimensional fused feature tensor; A dynamic fault prediction module is used to build a bidirectional gated recurrent network model based on a spatiotemporal attention mechanism, perform temporal dependency modeling on the high-dimensional fused feature tensor, and output the degradation probability and remaining life prediction value of key components of the equipment; The adaptive optimization control module is used to build a multi-objective dynamic programming model based on the predicted values, and uses a non-dominated sorting genetic algorithm to optimize the wind turbine power output, yaw angle and pitch rate to generate an anti-disturbance operation strategy.

2. The offshore wind turbine operation data analysis and prediction system according to claim 1, characterized in that: The multi-source spatiotemporal synchronization data set is subjected to denoising and alignment processing based on wavelet transform and adaptive filtering algorithm, specifically including: Extract vibration signals, wind speed and sea temperature series from multi-source spatiotemporal synchronized data, and generate high-frequency noise components and low-frequency trend components through multi-resolution wavelet decomposition; Adaptive filter is designed using the minimum mean square error criterion to dynamically suppress high frequency noise. The timestamp alignment algorithm is used to interpolate and synchronize the denoised low-frequency components with the device status data.

3. The offshore wind turbine operation data analysis and prediction system according to claim 1, characterized in that: Generating a high-dimensional fusion feature tensor includes: Define device nodes, environment nodes, and load nodes as graph structure vertices and construct the device-environment-load association matrix; A graph attention mechanism is used to aggregate features of adjacent nodes and generate node-level spatiotemporal embedding vectors. The association strength between nodes is optimized through a dynamic edge weight update strategy, and a feature tensor containing nonlinear coupling relationships is output.

4. The offshore wind turbine operation data analysis and prediction system according to claim 1, characterized in that: The construction of a bidirectional gated recurrent network model based on a spatiotemporal attention mechanism includes: In the time dimension, a bidirectional GRU network is used to capture the forward and backward dependencies of device states; A spatial attention layer is introduced to calculate the weight distribution of each sensor data and focus on abnormal signals of key components; The feature segments in the historical sequence that are strongly correlated with the current fault are filtered through the temporal attention layer to generate a degradation probability time series curve.

5. The offshore wind turbine operation data analysis and prediction system according to claim 1, characterized in that: The construction of the multi-objective dynamic programming model includes: The objective function is defined as a weighted combination of maximizing power generation efficiency, minimizing mechanical fatigue loss, and minimizing the sum of squared yaw errors; The constraints are set as the maximum allowable pitch angle, rated power threshold, and safety threshold range of wave impact force; The elite retention strategy and the congestion comparison operator are used to optimize the Pareto solution set and screen the anti-disturbance operation strategy.

6. The offshore wind turbine operation data analysis and prediction system according to claim 1, characterized in that: The system further comprises: The feedback correction module is used in an incremental learning framework based on a sliding time window. It combines Kalman filtering and Bayesian inference algorithms to perform online corrections on prediction model parameters and generate an adaptively updated fault prediction model. Specifically, it includes: Divide the real-time data stream based on the sliding window and extract the prediction error sequence within the window as the feedback signal; Design the Kalman gain matrix to recursively update the network weights to reduce the dependence on historical data; Bayesian posterior probability distribution is used to adjust model hyperparameters to achieve adaptive calibration of the prediction model.

7. The offshore wind turbine operation data analysis and prediction system according to claim 1, characterized in that: The multi-source spatiotemporal synchronized dataset includes: Integrate unit electrical parameters collected by the SCADA system, fiber optic sensor vibration data, and wave spectrum data provided by weather buoys; A tensor completion algorithm is used to jointly interpolate missing segments of heterogeneous data to construct a complete spatiotemporal data cube; Dimensional differences were eliminated by combining Z-score standardization and range normalization.

8. The offshore wind turbine operation data analysis and prediction system according to claim 1, characterized in that: The system further comprises: Construct a virtual sensor module based on physical information neural network to estimate the tower stress distribution that cannot be directly measured in real time; The estimated results are compared with the measured data to generate a residual signal, which triggers sensor fault diagnosis and data source switching instructions.

9. An electronic device, characterized in that: include: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the operations of the system as described in any one of claims 1 to 8.

Images