Wind power tower group twinborn cooperative intelligent operation and maintenance method

Abstract

The invention discloses a twinborn cooperative intelligent operation and maintenance method for a wind power tower group, and relates to the technical field of wind power generation towers, and the method comprises the following steps: S1, sensor deployment and real-time data collection; the method comprises the following steps: acquiring operation state data and environmental parameters through a sensor network deployed in each wind power generation tower in a wind power plant; s2, digital twinborn body construction and early warning are carried out; constructing a corresponding digital twinborn body for each wind power generation tower, wherein the digital twinborn body integrates a physical mechanism model and a prediction module driven by real-time data; generating a fault early warning signal and a health degree evaluation result based on the deviation between the simulation output of the digital twin and the sensor data; s3, multi-agent collaborative decision making is carried out; step S4, performing closed-loop optimization; according to the twinborn cooperative intelligent operation and maintenance method for the wind power tower group, the fault early warning precision is improved, the service life prediction error is compressed, the maintenance efficiency is improved, and the wake flow loss is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of wind power tower technology, and in particular to a twin-collaborative intelligent operation and maintenance method for wind power tower groups. Background Technology

[0002] Current wind farm operation and maintenance technologies face several key bottlenecks: isolated single-tower operation and maintenance, relying on traditional methods such as periodic inspections and post-failure repairs, which fail to predict equipment degradation trends (e.g., fatigue cracks in gearbox bearings); existing predictive maintenance technologies (such as SCADA-based systems) have a false alarm rate as high as 18-25%, leading to unnecessary downtime losses; and a lack of cross-tower collaboration capabilities: maintenance team scheduling often suffers from response delays (average 4.2 hours) due to task conflicts. Furthermore, the wake effect of wind farms causes annual power generation losses of 6-8%, and existing yaw control only optimizes individual unit power, lacking a cluster collaboration mechanism.

[0003] Data silos and model lag: Data from each power generation tower is analyzed independently, and fault knowledge cannot be shared. Traditional LSTM life prediction models have errors exceeding 25% and lack online calibration mechanisms, making them unable to adapt to component aging processes; Decision-making and execution are disconnected: Operation and maintenance strategies are based on static rules (such as changing gearbox oil every six months), and the execution results are not fed back to the decision-making system, leading to a mismatch between strategies and actual operating conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a twin-collaborative intelligent operation and maintenance method for wind turbine tower groups, which improves fault early warning accuracy, reduces life prediction error, increases maintenance efficiency, and significantly reduces wake loss.

[0005] This invention provides a twin-collaborative intelligent operation and maintenance method for wind turbine tower groups, including the following steps: Step S1, sensor deployment and real-time data acquisition;

[0006] The system collects operational status data and environmental parameters through a sensor network deployed on each wind turbine in the wind farm.

[0007] Step S2: Digital Twin Construction and Early Warning;

[0008] A corresponding digital twin is constructed for each wind turbine tower. The digital twin integrates a physical mechanism model and a real-time data-driven prediction module. Based on the deviation between the simulation output of the digital twin and the sensor data, fault warning signals and health assessment results are generated.

[0009] Step S3: Multi-agent collaborative decision-making;

[0010] Deploy independent multi-agent reinforcement learning agents for each wind power tower; each agent generates initial operation and maintenance instructions based on the output of its corresponding digital twin and cross-tower information shared by the global coordination module; the global coordination module detects resource conflicts or task conflicts between the initial operation and maintenance instructions and corrects them according to the dynamic coordination strategy to generate a collaborative operation and maintenance plan.

[0011] Step S4: Closed-loop optimization;

[0012] The collaborative operation and maintenance scheme is executed, and the execution results are fed back to the digital twin to update the model parameters; the updated digital twin output data is used to optimize the reinforcement learning policy network of each agent online.

[0013] Preferably, in step S2, the prediction module of the digital twin includes an anomaly detection submodule and a lifetime prediction submodule. The anomaly detection submodule is constructed using an autoencoder or a single-class support vector machine; the lifetime prediction submodule regresses the remaining lifetime of key components using an LSTM neural network.

[0014] The running status data is shown in the following formula:

[0015]

[0016] Where a(t) is a three-dimensional acceleration vector; a x a y a z These are the acceleration components in the X, Y, and Z axes, respectively.

[0017] The environmental parameters are shown in the following formula:

[0018] V env (t)=[v ω (t),θ ω (t),T a (t)];

[0019] Among them, v ω For real-time wind speed; θ ω The wind direction angle is 0, where 0 represents true north; T a The ambient temperature.

[0020] Preferably, in step S2, the physical model is executed as shown in the following equation:

[0021]

[0022] Where M is the mass matrix; C is the damping matrix; K is the stiffness matrix; x is the displacement vector; F aero This is the aerodynamic load vector;

[0023] The anomaly detection submodule is shown in the following formula:

[0024] ∈(t)=‖‖a(t)-f d (f e (a(t);W))‖;

[0025] ∈(t)>τ th =u ∈ +3σ ∈ ;

[0026] Where ∈(t) represents the reconstruction error of the autoencoder; f e f is the encoder function; d Here, W is the decoder function; W represents the neural network weight parameters; u ∈ σ is the historical mean error. ∈ The historical error standard deviation;

[0027] When the real-time error exceeds the threshold τ th Trigger warning when the time is right; output warning level A. i .

[0028] Preferably, in step S3, when a maintenance team scheduling conflict is detected, the task sequence is rearranged according to the priority of power generation loss; when the wake effect causes power fluctuations in adjacent power towers, the yaw angle and power output are adjusted synchronously.

[0029] In step S3, the global coordination module integrates the anomaly detection models of each digital twin through a federated learning mechanism to generate a fault knowledge graph shared across towers.

[0030] In step S3, the multi-agent reinforcement learning adopts a centralized training and distributed execution architecture. The loss function of the policy network includes a single tower operation and maintenance cost penalty, a total power generation gain reward, and a conflict resolution success reward.

[0031] The input for the intelligent agent is constructed as shown in the following equation:

[0032]

[0033] Among them, H i Health index; A i It is at the warning level; D is the wake influence coefficient. team Distance to the maintenance team; P grid Priority is given to the power grid.

[0034] The collision detection mechanism is as follows:

[0035] Conflict={(i,j)|T start,i <T end,j <T start,i +Δt};

[0036] Among them, Tstart T is the planned start time. end Δt is the planned end time; Δt is the minimum conflict threshold.

[0037] Dynamic priority calculation is shown in the following formula:

[0038]

[0039] in, E is the rated power of the unit. loss,i D represents power generation losses due to shutdown. team Distance for the repair team to reach;

[0040] Wake-flow cooperative control is shown in the following equation:

[0041] Δφ d =-φmax·tanh(β∑k u→d (P u -P d ));

[0042] Where, Δφ d For yaw adjustment feet; φ max β is the maximum allowable yaw angle; β is the cooperative gain coefficient; P u -P d This is due to the power difference between upstream and downstream units.

[0043] Preferably, in step S4, federated knowledge sharing is performed as follows:

[0044]

[0045] Where w is the weight of the anomaly detection model; η is the learning rate; n k The data volume of the k-th generator unit is N; the total data volume is N. For local model gradients;

[0046] Reinforcement learning rewards are shown in the following formula:

[0047] R t = -0.5C cost +0.8ΔP POWER +0.4R c ;

[0048] Among them, C cost For operation and maintenance costs; ΔP POWER For power generation gain; R c Reward efficiency in conflict resolution;

[0049] Lifetime prediction calibration is shown in the following formula:

[0050]

[0051] Among them, L true L represents the actual service life of the component. pred Predict lifetime for the model; For lifetime prediction gradient;

[0052] The reward for successful conflict resolution is shown in the following formula:

[0053]

[0054] Where, τ k τ is the time for handling a single conflict. max This is the maximum tolerance time.

[0055] Preferably, in step S4, the online optimization uses the asynchronous advantage actor-critic algorithm to update the strategy network, and the network weights are updated once every K operation and maintenance decisions.

[0056] Preferably, in step S4, the feedback data includes the deviation between the actual maintenance time and the estimated time, false alarm / missed alarm flags, and corrections for the component performance degradation rate.

[0057] Therefore, the present invention employs the aforementioned twin-synergistic intelligent operation and maintenance method for wind turbine tower groups, which improves fault early warning accuracy, reduces lifespan prediction errors, increases maintenance efficiency, and significantly reduces wake loss. The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the overall structure of a twin-collaborative intelligent operation and maintenance method for wind turbine tower groups according to the present invention;

[0059] Figure 2 This is a system block diagram of the prediction module of a wind turbine tower group twin collaborative intelligent operation and maintenance method according to the present invention. Detailed Implementation

[0060] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0062] The terms "first," "second," and similar words used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0063] Example 1

[0064] like Figures 1-2 As shown, the present invention provides a twin-collaborative intelligent operation and maintenance method for wind power tower groups, which includes the following steps: Step S1, sensor deployment and real-time data acquisition.

[0065] A sensor network is deployed at the base of each wind turbine tower (vibration monitoring point), the gearbox casing (temperature monitoring point), and the blade root (strain monitoring point).

[0066] The system collects operational status data and environmental parameters through a sensor network deployed on each wind turbine in the wind farm; the operational status data is shown in the following formula:

[0067]

[0068] Where a(t) is a three-dimensional acceleration vector; a x a y a z These are the acceleration components in the X, Y, and Z axes, respectively.

[0069] The environmental parameters are shown in the following formula:

[0070] V env (t)=[v ω (t),θ ω (t),T a (t)];

[0071] Among them, v ω For real-time wind speed; θ ω The wind direction angle is 0, where 0 represents true north; T a The ambient temperature.

[0072] Step S2: Digital Twin Construction and Early Warning;

[0073] A corresponding digital twin is constructed for each wind turbine tower. The digital twin integrates a physical mechanism model and a real-time data-driven prediction module. Based on the deviation between the simulation output of the digital twin and the sensor data, fault warning signals and health assessment results are generated.

[0074] In step S2, the prediction module of the digital twin includes an anomaly detection submodule and a lifetime prediction submodule. The anomaly detection submodule is constructed using an autoencoder or a one-class support vector machine. The lifetime prediction submodule regresses the remaining lifetime of key components using an LSTM neural network.

[0075] The physical model is executed as shown in the following equation:

[0076]

[0077] Where M is the mass matrix; C is the damping matrix; K is the stiffness matrix; x is the displacement vector; F aero This is the aerodynamic load vector.

[0078] The anomaly detection submodule is shown in the following formula:

[0079] ∈(t)=‖‖a(t)-f d (f e (a(t);W))‖;

[0080] The warning trigger conditions are shown in the following formula:

[0081] ∈(t)>τ th =u ∈ +3σ ∈ ;

[0082] Where ∈(t) represents the reconstruction error of the autoencoder; f e f is the encoder function; d Here, W is the decoder function; W represents the neural network weight parameters; u ∈ σ is the historical mean error. ∈ This represents the historical standard deviation. Warning Level A. i The error is classified into three levels, from 0 to 3, based on the percentage exceeding the acceptable error.

[0083] When the real-time error exceeds the threshold τ th Trigger warning when the time is right; output warning level A. i ;

[0084] In the lifetime prediction submodule, the LSTM network predicts the remaining lifetime of critical components, as shown in the following formula:

[0085] L pred =LSTM(a(t),V env (t); W L );

[0086] The input is a sequence of vibration and environmental data for 200 consecutive time steps.

[0087] Step S3: Multi-agent collaborative decision-making;

[0088] Each wind turbine tower is equipped with an independent multi-agent reinforcement learning agent. Each agent generates initial operation and maintenance instructions based on the output of its corresponding digital twin and cross-tower information shared by the global coordination module. The global coordination module detects resource conflicts or task conflicts between the initial operation and maintenance instructions and makes corrections according to a dynamic coordination strategy to generate a collaborative operation and maintenance plan.

[0089] In step S3, when a maintenance team scheduling conflict is detected, the task sequence is rearranged according to the priority of power generation loss; when the wake effect causes power fluctuations in adjacent power towers, the yaw angle and power output are adjusted synchronously. When a maintenance team scheduling conflict is detected, the task sequence is rearranged according to the priority of power generation loss P. i Reorder the mission sequence; when wake effects cause power fluctuations, synchronously adjust the yaw angle Δφ. d and power output.

[0090] In step S3, the global coordination module integrates the anomaly detection models of each digital twin through a federated learning mechanism to generate a fault knowledge graph shared across towers.

[0091] In step S3, the multi-agent reinforcement learning adopts a centralized training and distributed execution architecture. The loss function of the policy network includes a single-tower maintenance cost penalty term, a total power generation gain reward term, and a conflict resolution success reward term. The multi-agent reinforcement learning adopts a centralized training and distributed execution architecture, and the policy network loss function includes a single-tower maintenance cost penalty term C. cost Total power generation gain bonus ΔP power Rewards for successful conflict resolution (R) c .

[0092] The input for the intelligent agent is constructed as shown in the following equation:

[0093]

[0094] Among them, H i Health index; A i It is at the warning level; D is the wake influence coefficient. team Distance to the maintenance team; P grid Priority is given to the power grid.

[0095] The collision detection mechanism is as follows:

[0096]

[0097] Among them, T start T is the planned start time. end Δt is the planned end time; Δt is the minimum conflict threshold; Δt is two hours.

[0098] Dynamic priority calculation is shown in the following formula:

[0099]

[0100] in, E is the rated power of the unit. loss,i D represents power generation losses due to shutdown. team Distance for the maintenance team to reach.

[0101] Wake-flow cooperative control is shown in the following equation:

[0102] Δφ d =-φmax·tanh(β∑k u→d (P u -P d ));

[0103] Where, Δφ d For yaw adjustment feet; φ max β is the maximum allowable yaw angle; β is the cooperative gain coefficient; P u -P d This is due to the power difference between upstream and downstream units.

[0104] Step S4: Closed-loop optimization;

[0105] The collaborative operation and maintenance scheme is executed, and the execution results are fed back to the digital twin to update the model parameters; the updated digital twin output data is used to optimize the reinforcement learning policy network of each agent online.

[0106] In step S4, the online optimization uses the asynchronous advantage actor-critic algorithm to update the policy network, and the network weights are updated once every K operation and maintenance decisions.

[0107] In step S4, the feedback data includes the deviation between the actual repair time and the estimated time, false alarm / missed alarm flags, and corrections for component performance degradation rate.

[0108] Federated knowledge sharing, as shown in the following formula:

[0109]

[0110] Where w is the weight of the anomaly detection model; η is the learning rate; n k The data volume of the k-th generator unit is N; the total data volume is N. This represents the local model gradient.

[0111] Reinforcement learning rewards are shown in the following formula:

[0112] R t = -0.5C cost +0.8ΔP POWER +0.4R c ;

[0113] Among them, C cost For operation and maintenance costs; ΔP POWER For power generation gain; R c Reward efficiency in conflict resolution.

[0114] In step S4, the reward for successful conflict resolution is as follows:

[0115]

[0116] Lifetime prediction calibration is shown in the following formula:

[0117]

[0118] Among them, L true L represents the actual service life of the component. pred Predict lifetime for the model; This is the gradient for lifetime prediction.

[0119] Therefore, the present invention adopts the above-mentioned twin-collaborative intelligent operation and maintenance method for wind power tower groups, which improves the accuracy of fault early warning, reduces the error of life prediction, improves maintenance efficiency, and significantly reduces wake loss.

[0120] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A twin-synergistic intelligent operation and maintenance method for wind turbine tower groups, characterized in that, Includes the following steps: Step S1: Sensor deployment and real-time data acquisition; The system collects operational status data and environmental parameters through a sensor network deployed on each wind turbine in the wind farm. Step S2: Digital Twin Construction and Early Warning; A corresponding digital twin is constructed for each wind turbine tower. The digital twin integrates a physical mechanism model and a real-time data-driven prediction module. Based on the deviation between the simulation output of the digital twin and the sensor data, fault warning signals and health assessment results are generated. Step S3: Multi-agent collaborative decision-making; Deploy independent multi-agent reinforcement learning agents for each wind turbine tower; each agent generates initial operation and maintenance instructions based on the output of its corresponding digital twin and cross-tower information shared by the global coordination module. The global coordination module detects resource or task conflicts between initial operation and maintenance commands, corrects them according to dynamic coordination strategies, and generates a collaborative operation and maintenance plan. Step S4: Closed-loop optimization; The collaborative operation and maintenance scheme is executed, and the execution results are fed back to the digital twin to update the model parameters; the updated digital twin output data is used to optimize the reinforcement learning policy network of each agent online.

2. The wind turbine tower group twin collaborative intelligent operation and maintenance method according to claim 1, characterized in that, In step S2, the prediction module of the digital twin includes an anomaly detection submodule and a lifetime prediction submodule. The anomaly detection submodule is constructed using an autoencoder or a single-class support vector machine; the lifetime prediction submodule regresses the remaining lifetime of key components using an LSTM neural network. The running status data is shown in the following formula: Where a(t) is a three-dimensional acceleration vector; a x a y a z These are the acceleration components in the X, Y, and Z axes, respectively. The environmental parameters are shown in the following formula: V env (t)=[v ω (t),θ ω (t),T a (t)]; Among them, v ω For real-time wind speed; θ ω The wind direction angle is 0, where 0 represents true north; T a The ambient temperature.

3. The wind turbine tower group twin collaborative intelligent operation and maintenance method according to claim 1, characterized in that, In step S2, the physical model is executed, as shown in the following equation: Where M is the mass matrix; C is the damping matrix; K is the stiffness matrix; x is the displacement vector; F aero This is the aerodynamic load vector; The anomaly detection submodule is shown in the following formula: ∈(t)=‖‖a(t)-f d (f e (a(t);W))‖; ∈(t)>τ th =u ∈ +3s ∈ ; Where ∈(t) represents the reconstruction error of the autoencoder; f e f is the encoder function; d Here, W is the decoder function; W represents the neural network weight parameters; u ∈ σ is the historical mean error. ∈ The historical error standard deviation; When the real-time error exceeds the threshold τ th Trigger warning when the time is right; output warning level A. i .

4. The method for intelligent operation and maintenance of wind turbine tower groups using twin collaboration as described in claim 1, characterized in that, In step S3, when a maintenance team scheduling conflict is detected, the task sequence is rearranged according to the priority of power generation loss; when the wake effect causes power fluctuations in adjacent power towers, the yaw angle and power output are adjusted synchronously. In step S3, the global coordination module integrates the anomaly detection models of each digital twin through a federated learning mechanism to generate a fault knowledge graph shared across towers. In step S3, the multi-agent reinforcement learning adopts a centralized training and distributed execution architecture. The loss function of the policy network includes a single tower operation and maintenance cost penalty, a total power generation gain reward, and a conflict resolution success reward. The input for the intelligent agent is constructed as shown in the following equation: Among them, H i Health index; A i It is at the warning level; D is the wake effect coefficient. team Distance to the maintenance team; P grid Priority is given to the power grid. The collision detection mechanism is as follows: Conflict={(i,j)|T start,i <T end,j <T start,i +Δt}; Among them, T start T is the planned start time. end Δt is the planned end time; Δt is the minimum conflict threshold. Dynamic priority calculation is shown in the following formula: in, E is the rated power of the unit. loss,i D represents power generation losses due to shutdown. team Distance for the repair team to reach; Wake-flow cooperative control is shown in the following equation: Δφ d =-φmax·tanh(βΣk u→d (P u -P d )); Where, Δφ d For yaw adjustment feet; φ max β is the maximum allowable yaw angle; β is the cooperative gain coefficient; P u -P d This is due to the power difference between upstream and downstream units.

5. The method for intelligent operation and maintenance of wind turbine tower groups using twin collaboration according to claim 1, characterized in that, In step S4, federated knowledge sharing is performed as follows: Where w is the weight of the anomaly detection model; η is the learning rate; n k The data volume of the k-th generator unit is N; the total data volume is N. For local model gradients; Reinforcement learning rewards are shown in the following formula: R t =-0.5C cost +0.8ΔP POWER +0.4R c ; Among them, C cost For operation and maintenance costs; ΔP POWER For power generation gain; R c Reward efficiency in conflict resolution; Lifetime prediction calibration is shown in the following formula: Among them, L true L represents the actual service life of the component. pred Predict lifetime for the model; For lifetime prediction gradient; The reward for successful conflict resolution is shown in the following formula: Where, τ k τ is the time for handling a single conflict. max This is the maximum tolerance time.

6. The wind turbine tower group twin collaborative intelligent operation and maintenance method according to claim 1, characterized in that, In step S4, the online optimization uses the asynchronous advantage actor-critic algorithm to update the policy network, updating the network weights once every K operational decisions.

7. The method for intelligent operation and maintenance of wind turbine tower groups using twin collaboration according to claim 1, characterized in that, In step S4, the feedback data includes the deviation between the actual repair time and the estimated time, false alarm / missed alarm flags, and corrections for component performance degradation rate.

Images