An operation and maintenance scheduling method for offshore wind power generation equipment based on deep learning

Abstract

The present invention discloses an operation and maintenance scheduling method for offshore wind power generation equipment based on deep learning, which includes the following steps: S1, collecting the operation status data of the wind power generation equipment and performing preprocessing; S2, using a variational autoencoder to perform dimensionality reduction processing on the preprocessed data, extracting features, and calculating the health status score of the wind power generation equipment; S3, defining the state space of the Dyna-Q reinforcement learning model, constructing the action space and the reward function, and initializing the Q-value function; S4, using the Dyna-Q reinforcement learning model to optimize the operation and maintenance scheduling and updating the Q-value function; S5, selecting the optimal operation and maintenance scheduling plan and sending it to the remote monitoring system. The present invention combines deep learning with Dyna-Q reinforcement learning to achieve intelligent operation and maintenance scheduling of offshore wind power generation equipment, having the advantages of accurate prediction, optimized scheduling, high efficiency in convergence, and cost reduction.

Description

Technical Field

[0001] The present invention relates to the field of intelligent operation and maintenance of wind power generation equipment, and particularly to an operation and maintenance scheduling method for offshore wind power generation equipment based on deep learning. Background Art

[0002] As an important part of renewable energy, offshore wind power generation has been widely developed in recent years. Compared with onshore wind power, offshore wind farms have more stable wind energy resources and higher power generation efficiency. However, they also face higher operation and maintenance challenges. Since offshore wind power generation equipment is long-term exposed to complex marine environments and affected by various factors such as wind speed, wind direction, wave height, temperature and humidity, and salt spray corrosion, the operating state of the equipment is prone to fluctuations, and long-term operation may lead to wear and damage of key components such as gearboxes, generators, blades, and bearings. Therefore, how to accurately evaluate the health state of wind power generation equipment, predict potential faults, and optimize operation and maintenance scheduling strategies is an urgent problem to be solved in the offshore wind power industry.

[0003] Existing operation and maintenance scheduling methods for offshore wind power generation equipment mainly rely on regular maintenance strategies and empirical judgments, that is, operation and maintenance personnel inspect and maintain wind power generation equipment according to preset cycles. This method has great limitations. On the one hand, the regular maintenance strategy cannot be optimized according to the actual health state of wind power generation equipment, which may lead to unnecessary maintenance and increase operation and maintenance costs. On the other hand, empirical judgments often rely on the subjective judgments of operation and maintenance personnel, lack systematicness, and are difficult to meet the intelligent management requirements of large-scale wind farms. In addition, when sudden failures occur in wind power generation equipment, traditional passive maintenance methods are difficult to respond in a timely manner, which may lead to unplanned outages of wind power generation equipment and affect power generation benefits.

[0004] In recent years, with the rapid development of intelligent sensing technology and the Internet of Things, offshore wind farms have gradually realized remote monitoring of the operating state data of wind power generation equipment. However, how to effectively utilize these high-dimensional and strongly time-series massive data, accurately extract the health state characteristics of wind power generation equipment, and optimize maintenance strategies in combination with intelligent scheduling algorithms is still a bottleneck in current technology. Some existing data-driven methods, such as using machine learning models such as long short-term memory networks or random forests for fault prediction of wind power generation equipment, can improve the fault prediction ability to a certain extent. However, these methods usually only rely on the historical data of a single wind power generation equipment and do not fully consider the correlation between wind power generation equipment. In offshore wind farms, wind power generation equipment usually operates in a cluster form, and the spatial correlation and mutual influence of the operating states between wind power generation equipment within the wind farm cannot be ignored. Existing methods lack the ability to model the group state of wind power generation equipment, resulting in limited prediction accuracy.

[0005] In terms of the operation and maintenance scheduling optimization of wind power generation equipment, current methods mainly include those based on heuristic algorithms and reinforcement learning. Heuristic algorithms such as genetic algorithms and ant colony optimization can search for relatively optimal scheduling solutions, but their computational complexity is high, making it difficult to meet the optimization requirements in the dynamic environment of wind farms. When traditional reinforcement learning methods are used to solve operation and maintenance scheduling problems, the following problems exist: slow convergence speed, especially in the case of large state spaces and action spaces, traditional Q-learning requires a large amount of interaction data to achieve convergence; low exploration efficiency, making it difficult to quickly find the optimal scheduling strategy in the complex operation and maintenance environment of wind power generation equipment; insufficient combination of simulation data, resulting in the reinforcement learning model relying only on real interaction data for training. However, the real maintenance operation cost of offshore wind power equipment is high and the trial-and-error space is limited. Therefore, how to design a reinforcement learning method that can converge efficiently and combine real interaction data with simulated experience data has become the key to optimizing the operation and maintenance scheduling of wind power generation equipment. Summary of the Invention

[0006] An object of the present invention is to provide an operation and maintenance scheduling method for offshore wind power generation equipment based on deep learning. The present invention comprehensively utilizes a variational autoencoder and a graph attention network to extract the health state characteristics of wind power generation equipment, and optimizes the scheduling scheme based on Dyna-Q reinforcement learning combined with an entropy regularization strategy, fully considering the health state of the wind power generation equipment group, the allocation of maintenance resources, and weather conditions, realizing the optimization of operation and maintenance scheduling, improving the healthy operation level of wind power generation equipment, and reducing the maintenance cost.

[0007] An operation and maintenance scheduling method for offshore wind power generation equipment based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect the operation status data of wind power generation equipment, and perform real-time data transmission through a sensor network and an Internet of Things gateway to obtain the original data, and preprocess the original data;

[0009] S2. Use a variational autoencoder to perform dimensionality reduction processing on the preprocessed data, extract features, and combine a graph attention network to learn the correlation between wind power generation equipment, and calculate the health state score of wind power generation equipment;

[0010] S3. Define the state space of the Dyna-Q reinforcement learning model based on the health state score of wind power generation equipment, and construct an action space and a reward function, and initialize the Q-value function;

[0011] S4. Use the Dyna-Q reinforcement learning model to optimize the operation and maintenance scheduling. Based on the constructed state space, action space, and reward function, generate and execute a preliminary operation and maintenance scheduling plan to form real interaction data. Use the Monte Carlo sampling method to generate simulated experience data. Combine the real interaction data and the simulated experience data to update the Q-value function and generate an optimized operation and maintenance scheduling plan.

[0012] S5. Simulate the execution effects of different operation and maintenance scheduling plans, evaluate the impacts of the operation and maintenance scheduling plans on the operation efficiency and operation and maintenance costs of wind power generation equipment, select the optimal operation and maintenance scheduling plan, and send the optimal operation and maintenance scheduling plan to the remote monitoring system.

[0013] Optionally, the operating state data includes vibration signals, temperature and humidity, current, voltage, wind speed, wind direction, and wave height.

[0014] Optionally, the preprocessing includes data cleaning, outlier removal, time series alignment, and data normalization.

[0015] Optionally, S2 specifically includes:

[0016] S21. Use a variational autoencoder to perform dimensionality reduction on the preprocessed data, extract the latent features of the operating state of the wind power generation equipment, and define the preprocessed data as input data. The encoder network of the variational autoencoder maps the input data to the latent variable. The latent variable follows a normal distribution:

[0017] ;

[0018] Among them, represents the latent variable, describing the latent features of the operating state of the wind power generation equipment. represents the weight vector of the encoder network. represents the bias term of the encoder network. represents the input data. represents the standard deviation vector output by the encoder. represents the random noise of the standard normal distribution.

[0019] S22. Define the comprehensive loss function of the variational autoencoder, train the variational autoencoder, and use the comprehensive loss function for optimization. The comprehensive loss function includes a reconstruction error term, a KL divergence term, and a regularization term:

[0020] ;

[0021] Among them, represents the comprehensive loss function of the variational autoencoder. and Represents the data reconstructed by the decoder, Represents the mathematical expectation, Represents the fourth-order error weight, Represents the input data, Represents the weight factor of the KL divergence loss, Represents the dimension of the latent variable, Represents the L1 regularization coefficient, which controls sparsity, Represents the weights of the variational autoencoder, Represents the L1 norm regularization term, Represents the standard deviation vector of the encoder output, Represents the mean vector of the encoder output, Represents the dimension of the input data;

[0022] S23. Using the latent variable Calculate the low-dimensional feature vector of the wind power generation equipment:

[0023] ;

[0024] Among them, Represents the low-dimensional feature vector of the wind power generation equipment, Represents the rectified linear unit activation function, And Represents the feature transformation weight matrix, And Represents the bias term, Represents the weight factor, Represents the hyperbolic tangent activation function;

[0025] S24. Define that the adjacency relationship between wind power generation equipment is represented by an adjacency matrix, and the low-dimensional feature vector of each wind power generation equipment is , and use the graph attention mechanism to calculate the association weights between wind power generation equipment:

[0026] ;

[0027] Among them, Represents the wind power generation equipment And the wind power generation equipment The association weight between them, Represents the non-linear activation function, Represents the weight vector, Represents the transpose operation, Represents the transformation matrix for attention calculation, Represents the concatenation operation, Represents the distance conversion function, Represents the wind power generation equipment And the wind power generation equipment the geographical distance between represents the low-dimensional feature vector of the wind power generation equipment ;

[0028] S25. Calculate the adjusted health state feature vector of the wind power generation equipment by the graph attention network:

[0029] ;

[0030] wherein, represents the adjusted health state feature vector of the wind power generation equipment by the graph attention network, represents the wind power generation equipment 's neighbor set, represents the attention weight between wind power generation equipment, represents the transformation matrix for attention calculation, represents the wind power generation equipment 's low-dimensional feature vector;

[0031] S26. Calculate the health state score of the wind power generation equipment based on the adjusted health state feature vector of the wind power generation equipment:

[0032] ;

[0033] wherein, represents the health state score of the wind power generation equipment, represents the weight coefficient of the characteristics of the wind power generation equipment, represents the influence factor of the health state of the neighboring wind power generation equipment, represents the wind power generation equipment and the wind power generation equipment 's association weight, represents the wind power generation equipment and the wind power generation equipment 's feature correlation, obtained by cosine similarity.

[0034] Optionally, the state space includes the health state score of the wind power generation equipment, maintenance resources, weather conditions and operation and maintenance task load, the action space includes maintenance task allocation, maintenance team dispatch and maintenance path optimization, and the reward function is optimized based on operation and maintenance cost, equipment availability and maintenance response time.

[0035] Optionally, the S4 specifically includes:

[0036] S41. Based on the Dyna-Q reinforcement learning model, use the health state score of the wind power generation equipment, maintenance resources, weather conditions and operation and maintenance task load as state inputs, and combine with the Q-value function to adjust the operation and maintenance scheduling plan:

[0037] ;

[0038] Among them, represents the Q-value function, describing the expected reward value after executing the action in the state ; represents the mathematical expectation symbol, describing the expectation calculation of the reward value, represents the discount factor, controlling the attenuation degree of future rewards, with a value between 0 and 1, represents the index of the future time step, represents the time at which the reward function value is measured to evaluate the benefit after executing the action ;

[0039] S42. Optimize the operation and maintenance scheduling plan using the entropy regularization strategy to generate a preliminary operation and maintenance scheduling plan:

[0040] ;

[0041] Among them, represents the preliminary operation and maintenance scheduling plan, represents the variable value when taking the maximum of the function, represents the Q-value function, describing the expected reward value after executing the action in the state ; represents the entropy regularization coefficient. The larger the value, the more uniform the policy distribution, represents the policy function, describing the probability distribution of selecting the action in the current policy in the state ;

[0042] S43. Apply the selected preliminary operation and maintenance scheduling plan to the real environment, execute the preliminary operation and maintenance scheduling plan and monitor the operating status of wind power generation equipment, maintenance resource consumption, and task completion, and construct a wind power generation equipment state transition model:

[0043] ;

[0044] Among them, represents the health status score of the wind power generation equipment at the time ; represents the health status score of the wind power generation equipment at the time ; represents the preliminary operation and maintenance scheduling plan, represents the amount of maintenance resources invested, represents the weather condition, represents the environmental disturbance term, indicating the impact of unpredictable factors on the state of the wind power generation equipment. , and represent the weight coefficients of state changes;

[0045] S44. Record the key data after implementing the preliminary operation and maintenance scheduling plan to form real interaction data:

[0046] ;

[0047] Among them, represents the real interaction data, represents the operating state of the wind power generation equipment at time , represents the operating state of the wind power generation equipment at time , represents the preliminary operation and maintenance scheduling plan, represents the reward function value obtained after executing in the real environment;

[0048] S45. Based on the wind power generation equipment state transition model, evaluate the impact of the preliminary operation and maintenance scheduling plan, and use the Monte Carlo sampling method to generate simulated experience data:

[0049] ;

[0050] ;

[0051] Among them, represents the state of the wind power generation equipment generated by the - th sampling, represents the noise, simulating state uncertainty, , and represent the weight coefficients of state changes, represents the simulated experience data, represents the reward function value obtained by Monte Carlo sampling;

[0052] S46. Based on the Dyna - Q reinforcement learning method, combine the real interaction data and the simulated experience data to update the Q - value function and dynamically adjust the Q - value update strategy:

[0053] ;

[0054] ;

[0055] Among them, represents the Q - value calculated based on the real interaction data, represents the Q - value calculated based on the simulated experience data, represents the real data update learning rate, represents the simulated data update learning rate, represents the reward function value in the real environment, represents the discount factor, which controls the impact of future rewards on current decisions, represents at state and state the optimal action selected;

[0056] S47. Based on the finally trained Q-value function, when a new task arrives, select the optimal operation and maintenance scheduling plan according to the optimal Q-value, and generate an optimized operation and maintenance scheduling plan:

[0057] ;

[0058] wherein, represents the finally optimized operation and maintenance scheduling plan, represents the weighting coefficient, which balances the impact of real data and simulated data on decisions.

[0059] The beneficial effects of the present invention are as follows:

[0060] First of all, based on the combination of deep learning and reinforcement learning, the present invention realizes the intelligent operation and maintenance scheduling optimization of offshore wind power equipment. Compared with traditional regular maintenance and experience-based scheduling methods, it has higher prediction accuracy, better scheduling efficiency and stronger adaptability.

[0061] Secondly, the present invention uses a variational autoencoder to perform dimensionality reduction processing on the operation state data of wind power equipment, extracts key features from high-dimensional and complex sensing data, and combines a graph attention network to model the correlation between wind power equipment, calculates the health state score of wind power equipment, and more accurately evaluates the operation state and potential failure risks of wind power equipment, overcoming the limitation of existing methods that independently model individual wind power equipment, and improving the prediction accuracy.

[0062] In addition, in terms of operation and maintenance scheduling optimization, the present invention constructs a Dyna-Q reinforcement learning model, defines the state space based on the health status score of wind power generation equipment, maintenance resources, weather conditions, and operation and maintenance task loads, and optimizes the scheduling scheme by combining the entropy regularization strategy, effectively improving the intelligent level of operation and maintenance scheduling. Compared with traditional reinforcement learning methods, the present invention adopts the Dyna-Q mechanism, which can generate simulated experience data for training by using Monte Carlo sampling while combining real interaction data, thereby reducing the dependence on the real environment, accelerating the model convergence speed, and improving the robustness of decision-making. The optimization strategy based on entropy regularization ensures the diversity of the scheduling scheme, enabling the model to balance the exploration of unknown scheduling strategies and the utilization of known optimal scheduling schemes, and avoiding falling into local optima.

[0063] Finally, the operation and maintenance scheduling optimization method of the present invention can dynamically adjust the maintenance plan under different meteorological conditions, the health status of wind power generation equipment, and operation and maintenance resource constraints, reduce the unplanned downtime, improve the equipment availability rate, optimize the allocation of maintenance tasks, and improve the work efficiency of the operation and maintenance team. Through application in large-scale wind farms, the present invention can effectively reduce the operation and maintenance costs, improve the overall operation efficiency of the wind power generation equipment group, and enhance the economic benefits of the wind farm. At the same time, combined with the integration of the remote monitoring system, it can achieve the remote intelligent scheduling and maintenance of wind power generation equipment, reduce the work risks of operation and maintenance personnel in the harsh marine environment, and enhance the safety and reliability of operation and maintenance work.

[0064] In summary, through the organic combination of deep learning and reinforcement learning, the present invention has significant advantages in accurately predicting the health status of wind power generation equipment, optimizing the operation and maintenance scheduling strategy, and improving the overall operation efficiency of the wind farm, and can provide an innovative technical solution for the intelligent operation and maintenance management of offshore wind power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0066] Figure 1 is a flowchart of a method for operation and maintenance scheduling of offshore wind power generation equipment based on deep learning proposed by the present invention;

[0067] Figure 2 is a schematic flowchart of optimizing operation and maintenance scheduling based on the Dyna-Q reinforcement learning model for a method for operation and maintenance scheduling of offshore wind power generation equipment based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only schematically showing the basic structure of the present invention, so they only show the components related to the present invention.

[0069] Reference Figure 1 and Figure 2 , a maintenance and operation scheduling method for offshore wind power generation equipment based on deep learning, comprising the following steps:

[0070] S1. Collect the operation status data of the wind power generation equipment, and perform real-time data transmission through the sensor network and the Internet of Things gateway to obtain the original data, and preprocess the original data;

[0071] S2. Use the variational autoencoder to perform dimensionality reduction processing on the preprocessed data, extract features, and combine the graph attention network to learn the correlation between wind power generation equipment, and calculate the health status score of the wind power generation equipment;

[0072] S3. Define the state space of the Dyna-Q reinforcement learning model based on the health status score of the wind power generation equipment, and construct the action space and the reward function, and initialize the Q-value function;

[0073] S4. Use the Dyna-Q reinforcement learning model to optimize the maintenance and operation scheduling. Based on the constructed state space, action space and reward function, generate and execute the preliminary maintenance and operation scheduling plan to form real interaction data, use the Monte Carlo sampling method to generate simulated experience data, and combine the real interaction data and the simulated experience data to update the Q-value function and generate the optimized maintenance and operation scheduling plan;

[0074] S5. Simulate the execution effects of different maintenance and operation scheduling plans, evaluate the impact of the maintenance and operation scheduling plans on the operation efficiency and maintenance cost of the wind power generation equipment, select the optimal maintenance and operation scheduling plan, and send the optimal maintenance and operation scheduling plan to the remote monitoring system.

[0075] In this embodiment, the operation status data includes vibration signals, temperature and humidity, current, voltage, wind speed, wind direction and wave height.

[0076] In this embodiment, the preprocessing includes data cleaning, outlier removal, time series alignment and data normalization.

[0077] In this embodiment, the specific content of S2 is as follows:

[0078] S21. Use the variational autoencoder to perform dimensionality reduction processing on the preprocessed data, extract the latent features of the operation status of the wind power generation equipment, and define the preprocessed data as the input data , and the encoder network of the variational autoencoder maps the input data to the latent variable , and the latent variable follows a normal distribution:

[0079] ;

[0080] Among them, represents a latent variable, describing the potential characteristics of the operating state of the wind power generation equipment, represents the weight vector of the encoder network, represents the bias term of the encoder network, represents the input data, represents the standard deviation vector of the encoder output, represents the random noise of the standard normal distribution;

[0081] S22. Define the comprehensive loss function of the variational autoencoder, train the variational autoencoder, and optimize it using the comprehensive loss function. The comprehensive loss function includes a reconstruction error term, a KL divergence term, and a regularization term:

[0082] ;

[0083] Among them, represents the comprehensive loss function of the variational autoencoder, and represents the data reconstructed by the decoder, represents the mathematical expectation, represents the fourth-order error weight, represents the input data, represents the weight factor of the KL divergence loss, represents the dimension of the latent variable, represents the L1 regularization coefficient, controlling sparsity, represents the weight of the variational autoencoder, represents the L1 norm regularization term, represents the standard deviation vector of the encoder output, represents the mean vector of the encoder output, represents the dimension of the input data;

[0084] S23. Use the latent variable to calculate the low-dimensional feature vector of the wind power generation equipment:

[0085] ;

[0086] Among them, represents the low-dimensional feature vector of the wind power generation equipment, represents the rectified linear unit activation function, and represent the feature transformation weight matrix, and represent the bias term, Represents a weight factor, represents the hyperbolic tangent activation function;

[0087] S24. Define that the adjacency relationship between wind power generation devices is represented by an adjacency matrix, and the low-dimensional feature vector of each wind power generation device is , and use the graph attention mechanism to calculate the association weight between wind power generation devices:

[0088] ;

[0089] Among them, represents the wind power generation device and the wind power generation device the association weight between them, represents the non-linear activation function, represents the weight vector, represents the transpose operation, represents the transformation matrix for attention calculation, represents the concatenation operation, represents the distance conversion function, represents the wind power generation device and the wind power generation device the geographical distance between them, represents the wind power generation device the low-dimensional feature vector of;

[0090] S25. Calculate the adjusted health state feature vector of the wind power generation device by the graph attention network:

[0091] ;

[0092] Among them, represents the adjusted health state feature vector of the wind power generation device by the graph attention network, represents the neighbor set of the wind power generation device , represents the attention weight between wind power generation devices, represents the transformation matrix for attention calculation, represents the wind power generation device the low-dimensional feature vector of;

[0093] S26. Calculate the health state score of the wind power generation device based on the adjusted health state feature vector of the wind power generation device:

[0094] ;

[0095] Among them, represents the health state score of the wind power generation device, represents the weight coefficient of the characteristics of the wind power generation device, Indicates the influencing factor of the health status of neighboring wind power generation equipment, Indicates the wind power generation equipment and the wind power generation equipment The associated weight between them, Indicates the wind power generation equipment and the wind power generation equipment The feature correlation of, obtained by cosine similarity.

[0096] In this embodiment, the state space includes the health status score of the wind power generation equipment, maintenance resources, weather conditions, and operation and maintenance task load. The action space includes maintenance task allocation, maintenance team dispatch, and maintenance path optimization. The reward function is optimized based on operation and maintenance costs, equipment availability, and maintenance response time.

[0097] In this embodiment, the S4 specifically includes:

[0098] S41. Based on the Dyna-Q reinforcement learning model, using the health status score of the wind power generation equipment, maintenance resources, weather conditions, and operation and maintenance task load as state inputs, combined with the Q-value function Adjust the operation and maintenance scheduling plan:

[0099] ;

[0100] Among them, Represents the Q-value function, describing the expected reward value after executing the action in the state , Represents the mathematical expectation symbol, describing the expectation calculation of the reward value, Represents the discount factor, controlling the attenuation degree of future rewards, with a value between 0 and 1, Represents the index of the future time step, Represents the moment The reward function value at, measuring the benefit after executing the action ;

[0101] S42. Optimize the operation and maintenance scheduling plan using the entropy regularization strategy to generate a preliminary operation and maintenance scheduling plan:

[0102] ;

[0103] Among them, Represents the preliminary operation and maintenance scheduling plan, Represents the variable value when taking the maximum of the function, Represents the Q-value function, describing the expected reward value after executing the action in the state , denotes the entropy regularization coefficient. The larger the value, the more uniform the policy distribution. denotes the policy function, which describes the probability distribution of selecting an action in the current policy at state ;

[0104] S43. Apply the selected preliminary operation and maintenance scheduling plan to the real environment, execute the preliminary operation and maintenance scheduling plan, monitor the operating status of wind power equipment, maintenance resource consumption, and task completion, and construct a state transition model of wind power equipment:

[0105] ;

[0106] where denotes the health status score of the wind power equipment at time ; denotes the health status score of the wind power equipment at time ; denotes the preliminary operation and maintenance scheduling plan; denotes the input of maintenance resources; denotes the weather condition; denotes the environmental disturbance term, which represents the impact of unpredictable factors on the state of wind power equipment; , and denote the weight coefficients of state change;

[0107] S44. Record the key data after executing the preliminary operation and maintenance scheduling plan to form real interaction data:

[0108] ;

[0109] where denotes the real interaction data; denotes the operating status of the wind power equipment at time ; denotes the operating status of the wind power equipment at time ; denotes the preliminary operation and maintenance scheduling plan; denotes the reward function value obtained after executing in the real environment;

[0110] S45. Based on the state transition model of wind power equipment, evaluate the impact of the preliminary operation and maintenance scheduling plan, and use the Monte Carlo sampling method to generate simulated experience data:

[0111] ;

[0112] ;

[0113] Among them, represents the state of the wind power generation equipment generated by the th subsampling, represents noise, simulating state uncertainty, , and represent the weight coefficients of state changes, represents the simulated empirical data, represents the reward function value obtained by Monte Carlo sampling;

[0114] S46. Based on the Dyna-Q reinforcement learning method, combined with real interaction data and simulated empirical data, update the Q-value function and dynamically adjust the Q-value update strategy:

[0115] ;

[0116] ;

[0117] Among them, represents the Q-value calculated based on real interaction data, represents the Q-value calculated based on simulated empirical data, represents the real data update learning rate, represents the simulated data update learning rate, represents the reward function value in the real environment, represents the discount factor, controlling the influence of future rewards on the current decision, represents the optimal action selected in state and state ;

[0118] S47. Based on the finally trained Q-value function, when a new task arrives, select the optimal operation and maintenance scheduling plan according to the optimal Q-value to generate an optimized operation and maintenance scheduling plan:

[0119] ;

[0120] Among them, represents the finally optimized operation and maintenance scheduling plan, represents the weighting coefficient, balancing the influence of real data and simulated data on the decision.

[0121] Example 1:

[0122] To verify the feasibility of the present invention in implementation, the present invention is applied to a large-scale offshore wind farm along the coast. The wind farm is located in the offshore area near 121.45° east longitude and 32.12° north latitude, with a total installed capacity of 300 MW. A total of 100 wind power generation devices are arranged, and the impeller diameter of each wind power generation device is 150 meters, with a rated power of 3 MW. The wind farm is long-term affected by harsh marine environments such as typhoons, waves, and salt spray, and there is a large uncertainty in the operating state of the wind power generation devices. Traditional operation and maintenance methods have problems such as fixed maintenance cycles, inaccurate fault prediction, and unreasonable scheduling, resulting in frequent unplanned outages and affecting the economic benefits of the wind farm.

[0123] In this scenario, the intelligent operation and maintenance scheduling method of the present invention is used to optimize the maintenance plan of the wind power generation devices. First, the operation data of all wind power generation devices in the wind farm are collected through a sensor network and an Internet of Things gateway, including key parameters such as vibration signals, current, voltage, wind speed, wind direction, temperature, humidity, and wave height. The data collection frequency is 1 minute / time, and the duration is 6 months. The total amount of collected data exceeds 10 TB. Then, the original data is preprocessed, including data cleaning, outlier removal, time series alignment, and data normalization, to ensure the quality and consistency of the data.

[0124] I. Data collection description:

[0125] Collection frequency: once per minute; collection period: 180 consecutive days; total data volume: about 10 TB;

[0126] Types of sensors: vibration acceleration sensors (unit: g); current and voltage sensors (units: A, V); meteorological sensors (wind speed m / s, wind direction °, temperature °C, humidity %); sea condition sensors (wave height m).

[0127] Table 1 Example table of partial real collected data (10:00 on January 15, 2024)

[0128] ;

[0129] II. Data preprocessing process and description:

[0130] 1. Data cleaning: Deal with missing values. Records with more than 3 missing items are directly removed; for a small number of missing values, use the time linear interpolation method to fill them;

[0131] Example: The vibration data is missing at 10:03, and the interpolation is (0.21 + 0.26) / 2 = 0.235 g;

[0132] 2. Outlier removal: Use the Z-score method to detect and remove outliers;

[0133] Example: When the vibration value is 1.52 and the Z value is greater than 3, it is regarded as abnormal.

[0134] 3. Time series alignment: Unify all data to the whole hour and minute granularity and align it using the linear interpolation method.

[0135] 4. Data normalization: Normalize to the interval [0, 1] using the Min - Max method.

[0136] Example: When the wind speed is 12.5 m / s, it is normalized to (12.5 - 0.2) / (25.0 - 0.2) ≈ 0.503 in the interval [0.2, 25.0].

[0137] Table 2 Example table of data after pre - processing (normalization results)

[0138]

[0139] After data processing, use the variational auto - encoder to perform dimensionality reduction on the data, extract the potential health features of wind power generation equipment, and combine the graph attention network to learn the correlation between wind power generation equipment, calculate the health status score of wind power generation equipment. In the test stage, the health status score of wind power generation equipment is applied to the wind power generation equipment fault prediction system to analyze the fault trends of different wind power generation equipment in the next week.

[0140] Based on the health status score of wind power generation equipment, a Dyna - Q reinforcement learning model is constructed, and the state space, action space, and reward function are defined. During the optimization process, the state space includes the health status score of wind power generation equipment, maintenance resources, weather conditions, and operation and maintenance task load. The action space includes maintenance task assignment, maintenance team dispatch, and maintenance path optimization. The reward function comprehensively considers operation and maintenance costs, equipment availability, and maintenance response time.

[0141] To verify the feasibility and effectiveness of the present invention, combined with real experimental data, the construction and calculation process of the health status score of wind power generation equipment and the operation and maintenance scheduling reward function are described in detail:

[0142] In the present invention, the health status score of wind power generation equipment is calculated according to the following formula:

[0143] ;

[0144] In the experiment, the health status feature vector (8 - dimensional) of equipment W01 is:

[0145] ;

[0146] The corresponding feature weights are respectively:

[0147] ;

[0148] The self-evaluation item calculation result is:

[0149] ;

[0150] The neighbor devices of device W01 include W02 and W03, and the correlation calculation is as follows:

[0151] , the associated weight between wind power generation devices ;

[0152] , the associated weight between wind power generation devices ;

[0153] The neighbor influence item is:

[0154] ;

[0155] The final health status score of the wind power generation device is:

[0156] ;

[0157] The above score indicates that the device is in good operating condition, suitable for continuous operation and inclusion in the predictive maintenance queue.

[0158] In the present invention, the Dyna-Q reinforcement learning model adopts the following reward function:

[0159] ;

[0160] Among them, represents the actual operation and maintenance cost after scheduling in the present invention, which is 2.5 million yuan, represents the historical maximum operation and maintenance cost, which is 3.1 million yuan, represents the availability of the wind power generation device, which is 95.2% (recorded as 0.952), represents the average response time, which is 3.2 hours, represents the maximum response time, which is 6.5 hours;

[0161] The weights are set as: , , ;

[0162] Substitute into the calculation:

[0163] ;

[0164] ;

[0165] This reward value indicates that the current scheduling strategy has achieved a better balance in terms of cost control, availability, and response efficiency.

[0166] The preliminary scheduling plan is generated by the Dyna-Q reinforcement learning model and is sent to the operation and maintenance team for execution. During the execution process, the status changes of wind power generation equipment are monitored in real time, and the consumed resources and task completion times during operation and maintenance are collected to form real interaction data. At the same time, the Monte Carlo sampling method is used to generate simulated experience data, and the Q-value function is updated in combination with the real interaction data to dynamically optimize the scheduling plan.

[0167] To evaluate the effect of the present invention, during the implementation period, the traditional regular maintenance method is compared with the intelligent operation and maintenance scheduling method proposed by the present invention, and key indicators such as the availability of wind power generation equipment, operation and maintenance costs, and fault response time are mainly investigated.

[0168] Table 3 Experimental comparison data table

[0169] ;

[0170] In terms of the operating availability rate of wind power generation equipment, the availability rate after optimization of the present invention reaches 95.2%, which is 10.1% higher than that of the traditional maintenance method. This improvement benefits from the accurate modeling of the health status of wind power generation equipment by the present invention using deep learning and graph attention networks, enabling the maintenance plan to more accurately match the actual needs of wind power generation equipment, avoiding unnecessary shutdown maintenance, and at the same time reducing the impact of sudden failures on the operation of wind power generation equipment.

[0171] In terms of the number of unplanned outages, the traditional maintenance method has an average of 19 unplanned outages per month, while the number of unplanned outages after optimization of the present invention is reduced to 7 times, and the decline rate reaches 63.2%. This improvement is mainly attributed to the application of the Dyna-Q reinforcement learning model in scheduling optimization. By combining real interaction data and simulated experience data generated by Monte Carlo sampling, the present invention can effectively predict the health status of wind power generation equipment, arrange maintenance plans in advance, and reduce the incidence of sudden failures.

[0172] In terms of operation and maintenance costs, the maintenance cost after optimization of the present invention is reduced from 3.1 million yuan of the traditional maintenance method to 2.5 million yuan, a decrease of 19.8%. The reduction in cost is mainly due to accurate fault prediction and efficient operation and maintenance scheduling, enabling reasonable allocation of maintenance resources, avoiding unnecessary manual intervention and equipment replacement. At the same time, the present invention optimizes the dispatch strategy of the maintenance team, improving the execution efficiency of each maintenance task, thereby further reducing the overall operation and maintenance expenditure.

[0173] In terms of the fault response time, after optimization in the present invention, the fault response time is shortened from 6.5 hours in the traditional method to 3.2 hours, a reduction of 50.8%. This is mainly due to the optimization ability of Dyna-Q reinforcement learning in operation and maintenance scheduling. Compared with the traditional method that relies on a fixed time period or waits passively for a fault to occur before responding, the present invention monitors the health status of wind power generation equipment in real time and generates an optimal scheduling plan in combination with reinforcement learning, enabling the maintenance team to quickly locate problems and adopt the optimal repair strategy, thus significantly shortening the response time.

[0174] In terms of the fault prediction accuracy, the intelligent prediction model of the present invention reaches an accuracy of 99.9%, an increase of 27.8% compared with 72.1% of the traditional method. This breakthrough is mainly attributed to the present invention using a variational autoencoder for feature extraction of the health status of wind power generation equipment and combining a graph attention network to model the correlation between wind power generation equipment, thereby improving the prediction accuracy. The traditional method often relies on the historical data of a single wind power generation equipment, while the present invention utilizes global information and deep learning algorithms to improve the ability to identify the fault trend of wind power generation equipment.

[0175] In addition, in the case of bad weather (wind speed ≥ 20 m / s, wave height ≥ 3.5 m), the availability of the optimized wind power generation equipment in the present invention reaches 90.5%, an increase of 15.6% compared with 78.3% of the traditional method. This shows that the operation and maintenance scheduling method of the present invention can still maintain a high operation efficiency in a complex environment, while the traditional method is prone to be affected by weather factors due to a fixed scheduling strategy, resulting in a significant decrease in the availability of wind power generation equipment.

[0176] In summary, the present invention shows advantages in improving the availability of wind power generation equipment, reducing unplanned outages, lowering operation and maintenance costs, shortening the fault response time, and increasing the fault prediction accuracy. These experimental data prove that the present invention has high practical value in the optimization of operation and maintenance scheduling in offshore wind farms, effectively improving the operation efficiency of the wind farm, reducing maintenance costs, and enhancing the stability and reliability of the operation of wind power generation equipment.

[0177] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for operation and maintenance scheduling of offshore wind power generation equipment based on deep learning, characterized in that It includes the following steps: S1. Collect the operation status data of wind power generation equipment, and perform real-time data transmission through the sensor network and the Internet of Things gateway to obtain the original data, and preprocess the original data; S2. Use the variational autoencoder to perform dimensionality reduction on the preprocessed data, extract features, and combine the graph attention network to learn the correlation between wind power generation equipment, and calculate the health status score of wind power generation equipment; S3. Define the state space of the Dyna-Q reinforcement learning model based on the health status score of wind power generation equipment, and construct the action space and the reward function, and initialize the Q-value function; S4. Use the Dyna-Q reinforcement learning model to optimize the operation and maintenance scheduling. Based on the constructed state space, action space and reward function, generate and execute the preliminary operation and maintenance scheduling plan to form real interaction data, use the Monte Carlo sampling method to generate simulated experience data, and combine the real interaction data and the simulated experience data to update the Q-value function and generate the optimized operation and maintenance scheduling plan; S5. Simulate the execution effects of different operation and maintenance scheduling plans, evaluate the impacts of the operation and maintenance scheduling plans on the operation efficiency and operation and maintenance costs of wind power generation equipment, select the optimal operation and maintenance scheduling plan, and send the optimal operation and maintenance scheduling plan to the remote monitoring system; The specific content of S2 includes: S21. Use a variational autoencoder to perform dimensionality reduction on the preprocessed data, extract the latent features of the operating state of the wind power generation equipment, and define the preprocessed data as input data , and the encoder network of the variational autoencoder maps the input data to the latent variable , and the latent variable follows a normal distribution: ； Among them, represents a latent variable, describing the potential characteristics of the operating state of the wind power generation equipment, represents the weight vector of the encoder network, represents the bias term of the encoder network, represents the input data, represents the standard deviation vector of the encoder output, represents the random noise of the standard normal distribution; S22. Define the comprehensive loss function of the variational autoencoder, train the variational autoencoder, and optimize it using the comprehensive loss function. The comprehensive loss function includes a reconstruction error term, a KL divergence term, and a regularization term: ； Among them, represents the comprehensive loss function of the variational autoencoder, and represents the data reconstructed by the decoder, represents the mathematical expectation, represents the fourth-order error weight, represents the input data, represents the weight factor of the KL divergence loss, represents the dimension of the latent variable, represents the L1 regularization coefficient to control sparsity, represents the weights of the variational autoencoder, represents the L1 norm regularization term, represents the standard deviation vector of the encoder output, represents the mean vector of the encoder output, represents the dimension of the input data; S23. Using latent variables Calculate the low-dimensional feature vector of the wind power generation equipment: ； Among them, represents the low-dimensional feature vector of the wind power generation equipment, represents the rectified linear unit activation function, and represents the feature transformation weight matrix, and represents the bias term, represents the weight factor, represents the hyperbolic tangent activation function; S24. Define that the adjacency relationship between wind power generation devices is represented by an adjacency matrix, and the low-dimensional feature vector of each wind power generation device is , and use the graph attention mechanism to calculate the correlation weight between wind power generation devices: ； Among them, represents the wind power generation equipment and the wind power generation equipment the correlation weight between them, represents the non-linear activation function, represents the weight vector, represents the transpose operation, represents the transformation matrix for attention calculation, represents the concatenation operation, represents the distance conversion function, represents the wind power generation equipment and the wind power generation equipment the geographical distance between them, represents the wind power generation equipment the low-dimensional feature vector of; S25. Calculate the adjusted health status feature vector of wind power generation equipment by the graph attention network: ； Among them, represents the feature vector of the health state of the wind power generation equipment after adjustment by the graph attention network, represents the wind power generation equipment 's neighbor set, represents the attention weight between wind power generation equipment, represents the transformation matrix for attention calculation, represents the wind power generation equipment 's low-dimensional feature vector; S26. Calculate the health status score of wind power generation equipment based on the adjusted health status feature vector of wind power generation equipment: ； Among them, represents the health status score of the wind power generation equipment, represents the weight coefficient of the characteristics of the wind power generation equipment, represents the influence factor of the health status of the neighboring wind power generation equipment, represents the wind power generation equipment and the wind power generation equipment the correlation weight between them, represents the wind power generation equipment and the wind power generation equipment the feature correlation of, obtained by cosine similarity.

2. The operation and maintenance scheduling method of an offshore wind power generation device based on deep learning according to claim 1, wherein The operation status data includes vibration signals, temperature and humidity, current, voltage, wind speed, wind direction, and wave height.

3. A method for operation and maintenance scheduling of offshore wind power generation equipment based on deep learning according to claim 1, characterized in that, The preprocessing includes data cleaning, outlier removal, time series alignment, and data normalization.

4. A method for operation and maintenance scheduling of offshore wind power generation equipment based on deep learning according to claim 1, characterized in that, The state space includes the health status score of wind power generation equipment, maintenance resources, weather conditions, and operation and maintenance task loads. The action space includes maintenance task allocation, maintenance team dispatch, and maintenance path optimization. The reward function is optimized based on operation and maintenance costs, equipment availability, and maintenance response time.

5. A method for operation and maintenance scheduling of offshore wind power generation equipment based on deep learning according to claim 1, characterized in that The specific content of S4 includes: S41. Based on the Dyna-Q reinforcement learning model, using the health status score of wind power generation equipment, maintenance resources, weather conditions, and operation and maintenance task load as state inputs, combined with the Q-value function to adjust the operation and maintenance scheduling plan: ； Among them, represents the Q-value function, which describes the state after performing the action and the expected reward value, where represents the mathematical expectation symbol, describing the expectation calculation of the reward value, represents the discount factor, which controls the attenuation degree of future rewards and takes values between 0 and 1, represents the index of future time steps, represents the time at which the reward function value is located, measuring the return after performing the action ; S42. Optimize the operation and maintenance scheduling plan using the entropy regularization strategy to generate the preliminary operation and maintenance scheduling plan: ； Among them, represents the preliminary operation and maintenance scheduling plan, represents the variable value when the function takes the maximum value, represents the Q-value function, which describes the state and the expected reward value after performing the action ; represents the entropy regularization coefficient. The larger the value, the more uniform the policy distribution, represents the policy function, which describes the probability distribution of selecting an action in the current policy at state ; S43. Apply the selected preliminary operation and maintenance scheduling plan to the real environment, execute the preliminary operation and maintenance scheduling plan, monitor the operating status of the wind power generation equipment, the consumption of maintenance resources, and the task completion situation, and construct a state transition model of the wind power generation equipment: ； Among them, represents the health status score of the wind power generation equipment at time . represents the health status score of the wind power generation equipment at time . represents the preliminary operation and maintenance scheduling plan, represents the input amount of maintenance resources, represents the weather condition, represents the environmental disturbance term, which represents the impact of unpredictable factors on the state of the wind power generation equipment, , and represent the weight coefficients of state changes; S44. Record the key data after executing the preliminary operation and maintenance scheduling plan to form real interaction data: ； Among them, represents the real interaction data, represents the operating state of the wind power generation equipment at time ; represents the operating state of the wind power generation equipment at time ; represents the preliminary operation and maintenance scheduling plan, represents the reward function value obtained after executing in the real environment; S45. Based on the state transition model of wind power generation equipment, evaluate the impact of the preliminary operation and maintenance scheduling plan, and use the Monte Carlo sampling method to generate simulated experience data: ； ； Among them, represents the state of the wind power generation equipment generated by the th sampling, represents noise, simulating state uncertainty, , and represent the weight coefficients of state changes, represents simulated empirical data, represents the reward function value obtained by Monte Carlo sampling; S46. Based on the Dyna-Q reinforcement learning method, combine the real interaction data and the simulated experience data to update the Q-value function and dynamically adjust the Q-value update strategy: ； ； Among them, represents the Q value calculated based on real interaction data, represents the Q value calculated based on simulated experience data, represents the real data update learning rate, represents the simulated data update learning rate, represents the reward function value in the real environment, represents the discount factor, which controls the impact of future rewards on the current decision, represents at state and state the optimal action selected; S47. Based on the finally trained Q-value function, when a new task arrives, select the optimal operation and maintenance scheduling plan according to the optimal Q value to generate the optimized operation and maintenance scheduling plan: ； Among them, represents the finally optimized operation and maintenance scheduling plan, represents the weighting coefficient to balance the influence of real data and simulation data on decision-making.

Images