A sea-land wind power double-layer operation and maintenance method driven by a neural network for predicting fan health

Abstract

This application belongs to the field of power system optimization and dispatching, specifically disclosing a neural network-driven two-layer operation and maintenance method for onshore and offshore wind power based on wind turbine health prediction. The method includes the following steps: obtaining the operating parameters of the onshore-offshore interconnected system; establishing a wind turbine health prediction model and a two-layer optimization model based on a physical guided neural network; optimizing power generation scheduling and maintenance decisions in the upper layer of the two-layer optimization model, and optimizing UAV inspection paths in the lower layer; iteratively solving the two-layer optimization model based on logical Benders decomposition; inputting the scheduling plan solved in the upper layer into the neural network to extract physical gradients and generate linear tangent plane constraints to be added to the upper layer model; passing the maintenance decisions output by the upper layer to the lower layer solution and feeding back the Benders cutting plane; repeating the iteration until the iteration gap between the upper and lower layer objective functions meets the convergence condition, and outputting the optimal scheduling plan and UAV path. This application can effectively achieve accurate prediction of wind turbine health status and collaborative optimization of operation and maintenance scheduling in onshore-offshore interconnected systems.

Description

Technical Field

[0001] This application belongs to the field of power system optimization and dispatching, and more specifically, relates to a two-layer operation and maintenance method for offshore and onshore wind power with a neural network-driven wind turbine health prediction. Background Technology

[0002] By constructing an interconnected sea-land system, offshore wind power clusters can be centrally connected to nearby onshore power transmission networks, greatly improving energy allocation efficiency and absorption capacity.

[0003] However, offshore wind turbines operate for extended periods in harsh marine environments characterized by high humidity, high salinity, and strong winds, while also enduring complex alternating loads, significantly accelerating the aging process of their structures and critical components. This accelerated degradation not only shortens the equipment's lifespan but also introduces uncertainty into its health status. Furthermore, due to the high logistical costs, complex sea conditions, and unavoidable response delays involved in offshore maintenance, the difficulty and cost of handling any malfunctions increase dramatically. Traditional passive maintenance strategies or periodic manual inspections are no longer adequate to meet the practical needs of today's large-scale offshore wind power operations.

[0004] Therefore, how to achieve accurate prediction of wind turbine health status and coordinated optimization of operation and maintenance scheduling in the sea-land interconnection system is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a neural network-driven method for predicting the health of wind turbines in a two-tiered operation and maintenance system for offshore and onshore wind power, which can achieve accurate prediction of the health status of wind turbines and collaborative optimization of operation and maintenance scheduling in an interconnected offshore and onshore system.

[0006] To achieve the above objectives, in a first aspect, this application provides a neural network-driven method for predicting the health of wind turbines in a two-tiered operation and maintenance system for onshore and offshore wind power, comprising the following steps: S10, acquire the operating parameters of the land-sea interconnection system, and establish a wind turbine health prediction model based on a physical guidance neural network, a maintenance path planning subsystem model based on UAVs, and a two-layer optimization model for power generation scheduling and UAV inspection path optimization; the upper layer model of the two-layer optimization model optimizes power generation scheduling and maintenance decisions with the goal of minimizing the overall system operating cost, and the lower layer model optimizes the inspection path based on the UAV-based maintenance path planning subsystem model with the goal of minimizing the UAV flight cost. S20, solve the two-layer optimization model iteratively based on logical Benders decomposition: In each iteration, input the scheduling plan obtained by the upper-layer model into the physical guidance neural network, extract the physical gradient and generate linear tangent plane constraints to be added to the upper-layer model, pass the maintenance decision output by the upper-layer model to the lower-layer model for solution, and feed back the Benders cutting plane to the upper-layer model according to the solution state of the lower-layer model. S30, repeat step S20 until the iteration gap between the upper and lower objective functions meets the convergence condition, and output the optimal scheduling plan and UAV path.

[0007] As a further preferred embodiment, in step S10, the wind turbine health prediction model based on the physical guided neural network is used to calculate the health index and health degradation increment of each wind turbine, so as to quantify the degree of wind turbine degradation and trigger maintenance decisions; the health index and health degradation increment are calculated by the following formula:

[0008]

[0009] in, HI t,n express t Time Fan n The health index is 1 for healthy and 0 for needing to be shut down. IHI t,n Indicates the increase in health deterioration; FNN( ) This represents a physically guided neural network function; V t,n eff , P w,t,n , t t , Tem and Hum These represent the equivalent inflow wind speed, the actual dispatch power of the fan, the cumulative operating time, the ambient temperature, and the ambient humidity, respectively.

[0010] As a further preferred embodiment, the health index and health degradation increment are used to establish maintenance logic and state transition constraints to determine the maintenance decision binary variables output by the upper-level model. Z t,n The constraint includes:

[0011]

[0012] in, HI init,n This indicates the initial health value of the fan after it has fully recovered. HImin Indicates the minimum safe and healthy threshold; u w,t,n and Z t,n These represent the operating status and maintenance decision-making variables of the wind turbine, respectively. K uav Indicates the total capacity of the drone fleet; M big and M HI All of these are large positive constants used for Big-M relaxation.

[0013] As a further preferred embodiment, in step S10, the upper-level model aims to minimize the overall system operating cost, and the objective function is:

[0014] in, C gen , C cess , C pen , C wind and These represent the costs of thermal power generation, energy storage cycle aging, wind curtailment and load shedding penalties, wind power operation, and drone fixed maintenance, respectively. or t A Benders proxy variable representing the drone routing cost passed down from the lower layer; This represents the set of all decision variables in the upper-level model.

[0015] As a further preferred embodiment, step S20, which involves extracting the physical gradient and generating the linear tangent plane constraint, specifically includes: The marginal physical gradient of wind turbine health on power adjustment is calculated using the chain rule. ; Transforming it into the following linear inequality constraint and incorporating it into the upper-level model, the calculation formula is as follows:

[0016]

[0017] in, IHI t,n LP This represents the amount of degradation after linearization; and Indicates the reference power and the optimal pre-scheduled power; IHI true,t,n This represents the actual physical degradation value obtained from the forward propagation of the neural network.

[0018] As a further preferred embodiment, in step S20, the objective function of the lower-level model and the feedback Benders cutting plane satisfy the following relationship: Lower-level model objective function: ; in, J sp,t: This represents the flight cost of the drone in the lower-level model; C deone_km This indicates the unit cost of flight per kilometer for a drone; d ij Represents a node i To the node j Flight distance; y k,i,j Indicates drone k From node i Fly to the node j A binary variable; The optimality cut or feasibility cut fed back from the lower-level model to the upper-level model is represented as:

[0019]

[0020] in, Indicates the current maintenance decision variable Interpreting History Hamming distance between them; M benders It is a large positive number.

[0021] As a further preferred embodiment, in step S10, the sea-land interconnection system operation and wake effect subsystem model is used to calculate the equivalent inflow wind speed and the maximum available power of the wind turbine, so as to constrain the actual wind power output. The actual output of the wind power It constitutes part of the scheduling plan described in step S20; The model includes the following formulas: Wake radius: ; Single speed loss factor: ; Equivalent inflow velocity: ; The maximum available power is determined by the power curve of the wind turbine generator set, using the following formula: ; in, R wake , R 0、 D , k wake and CT These represent the wake radius, rotor radius, rotor diameter, wake attenuation constant, and thrust coefficient, respectively. x ij This is the downstream axial distance; d ij This is the speed reduction factor; V t free Free-flow wind speed; P w,t,n avail , P curt,t,n These represent the maximum available power of the wind turbine and the power of the wind curtailed; v in , 、v out , v rate For cut-in, cut-out, and rated wind speed.

[0022] As a further preferred embodiment, in step S10, the UAV-based maintenance path planning subsystem model is used to constrain the task allocation, flight distance, and inspection quantity of multiple UAVs in order to generate the path constraint conditions required for the lower-level model to solve. The model includes the following constraints:

[0023]

[0024]

[0025]

[0026]

[0027]

[0028] in, and Represent the set of all physical nodes and the set of target nodes maintained by the UAV, respectively; K sub Indicates the total number of available drones; y k,i,j Indicates drone k From node i Fly to the node j A binary variable; u k Indicates drone k The utilization status; d i,j Represents a node i To the node j Flight distance;D max and N max These represent the maximum flight range of the drone and the maximum number of wind turbines that can be inspected in a single operation, respectively. order k,i This is an auxiliary variable representing the node access order for eliminating sub-loops.

[0029] As a further preferred embodiment, the specific process of the iterative solution in step S20 includes: Initialization: Load system operating parameters, including grid parameters, wind turbine layout parameters, and pre-trained neural network weights; Solve the upper-level master problem: Solve the mixed-integer linear programming master problem under the current constraint pool, and extract the wind power dispatch plan and maintenance plan; Physical guidance interface verification: Input the wind power dispatch plan into the neural network gradient check module, and calculate the actual health degradation increment by combining wind speed, time, temperature and humidity characteristics; Determine if there is a physical default: Compare the degradation amount estimated by the upper linearization model with the actual degradation amount output by the neural network. If the error exceeds a preset threshold, dynamically generate a physical guiding tangent plane, add it to the constraint pool of the upper master problem, and resolve the upper master problem. If there is no default, pass the maintenance plan to the lower subproblem. Solve the sub-problem at the next level: Given a fixed maintenance plan, optimize the flight paths of multiple UAVs; Feedback Benders Cut Plane: Generate corresponding Benders optimality cuts or feasibility cuts based on the solution status of the lower-level subproblems and feed them back to the upper level, adding them to the constraint pool of the main problem; Convergence check: Calculate the iteration gap between the upper and lower objective functions. If the convergence condition is met, terminate the iteration and output the global optimal result, including the optimal joint scheduling plan, UAV flight trajectory and the cost of each part of the system. If the condition is not met, proceed to the next iteration and solve the upper master problem with the new cutting plane.

[0030] Secondly, this application provides a neural network-driven dual-layer operation and maintenance system for onshore and offshore wind power, used to implement the method described in any one of the above, comprising: The parameter acquisition module is used to acquire the operating parameters of the sea-land interconnection system; The model building module is used to build a wind turbine health prediction model based on a physical guided neural network, a maintenance path planning subsystem model based on UAVs, and a two-layer optimization model for power generation scheduling and UAV inspection path optimization. The upper layer model of the two-layer optimization model optimizes power generation scheduling and maintenance decisions with the goal of minimizing the overall system operating cost, while the lower layer model optimizes the inspection path based on the UAV-based maintenance path planning subsystem model with the goal of minimizing the UAV flight cost. The iterative solution module is used to iteratively solve the two-layer optimization model based on logical Benders decomposition: In each iteration, the scheduling plan obtained by the upper-layer model is input into the physical guidance neural network, the physical gradient is extracted and a linear tangent plane constraint is generated and added to the upper-layer model, the maintenance decision output by the upper-layer model is passed to the lower-layer model for solution, and the Benders cutting plane is fed back to the upper-layer model according to the solution state of the lower-layer model. The convergence determination and output module is used to iterate repeatedly until the iteration gap between the upper and lower objective functions meets the convergence condition, and output the optimal scheduling plan and UAV path.

[0031] This application has the following beneficial effects: (1) A collaborative optimization framework for "Operation & Maintenance (O&M)" for land-sea interconnection systems was constructed, realizing the collaborative optimization of accurate prediction of wind turbine health status and operation and maintenance scheduling. By unifying the modeling and joint solution of wind power scheduling considering the Jensen wake effect and multi-UAV inspection path planning, the maintenance delay caused by the harsh marine environment can be alleviated, and a more flexible balance can be achieved between system operating costs and long-term equipment lifespan. This avoids the disconnect between prediction and scheduling caused by optimizing only a single objective.

[0032] (2) A dynamic linearization method with the introduction of Physically Guided Neural Network (PINN) is proposed, which significantly improves the prediction accuracy of wind turbine health status. The gradient of relevant physical quantities is calculated at the current operating point using the chain rule, which transforms the originally difficult black box mapping into a linear tangent plane with clear theoretical boundary constraints. This allows the optimization solver to follow the actual physical degradation law and safe operation boundary of the wind turbine during iteration, thus solving the problems of instability and poor prediction accuracy of pure data-driven methods in extrapolation scenarios.

[0033] (3) A two-layer solution model based on Logical Benders Decomposition (LBBD) was designed, which significantly improved the solution efficiency and stability of operation and maintenance scheduling. By adaptively generating a physically consistent tangent plane and constructing a combinatorial cut using Hamming distance, the continuous power grid scheduling problem and the discrete UAV path planning problem are structurally decoupled. Compared with directly solving the coupled nonlinear model, the computational load is greatly reduced, the curse of dimensionality of traditional methods is avoided, and the solution efficiency and stability are higher, thus ensuring the real-time performance and reliability of collaborative optimization. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of a two-layer optimized operation and maintenance mechanism for a land-sea interconnected system that considers neural network-driven wind turbine health prediction, provided in this application. Figure 2 This is a flowchart of the adaptive dynamic tangent plane strategy provided in this application; Figure 3This is a flowchart of the two-layer model solution provided in this application; Figure 4 This application provides a 7-day wind turbine health index fluctuation trajectory and drone scheduling diagram. Figure 5 These are wind power output dispatch diagrams for Scheme 2 and Scheme 3 provided in the embodiments of this application during a storm; Figure 6 This is a multi-UAV concurrent flight path diagram for the first hour of the fourth day provided in this application embodiment. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0036] This application provides a neural network-driven two-layer operation and maintenance method for onshore and offshore wind power based on wind turbine health prediction, including the following steps: like Figure 1 As shown, this application constructs a collaborative optimization framework for wind power operation and maintenance in an onshore-offshore interconnected system. The system consists of three parts: an offshore wind power and UAV inspection layer, a two-layer optimization and neural network coupled decision-making layer, and an onshore power grid operation layer, achieving unified coordination of wind power output optimization, equipment health management, and system scheduling. In terms of physical architecture, the offshore wind farm first collects the wind turbine output through a collection system, transmits it to the onshore via submarine cables, and finally connects to a six-node transmission network for centralized scheduling and consumption. The onshore power grid acts as the control center, coordinating and regulating the output of thermal power units, the charging and discharging of energy storage systems, and the load of each node to ensure power balance and stable operation of the system across multiple time scales. In the offshore portion, a fleet of UAVs is used for wind farm operation and maintenance. The UAVs are uniformly dispatched from a fixed offshore operation and maintenance base, conducting regular inspections and status monitoring of the wind turbines in a multi-drone parallel, closed-loop patrol manner, improving the real-time performance and coverage efficiency of equipment health status data collection.

[0037] At the intermediate decision-making layer, a dual-layer optimization and neural network coupled structure was designed: the upper layer aims to minimize the overall system operating cost, collaboratively optimizing thermal power output, energy storage regulation, and wind power consumption schemes; the lower layer plans UAV inspection routes based on maintenance needs, achieving multi-UAV collaborative path optimization. The upper and lower layers mutually feed back through maintenance decisions and path costs, forming a closed-loop optimization. Simultaneously, a neural network health prediction module is introduced, transforming various data such as wind speed, power generation, and operating time into equipment health changes, and feeding them back to the upper-layer scheduling model in the form of health constraints, allowing operational decisions to directly consider equipment degradation and lifespan loss. Relying on this multi-layer collaborative mechanism, the framework achieves deep integration and overall optimization of offshore wind power operation optimization, UAV maintenance scheduling, and onshore power grid economic scheduling.

[0038] Step 2: Establish models of each subsystem of the land-sea interconnection system. (1) Establish a maintenance path planning subsystem model based on UAVs Considering the marine environment, the UAV will not perform mid-mission battery swapping. A constraint model is established, including task allocation, flow conservation, physical operational limitations, and elimination of sub-loops, to limit the maximum range and the maximum number of wind turbines that can be inspected. (1a) (1b) (1c) (1d) (1e) (1f) in, and Represent the set of all physical nodes and the set of target nodes maintained by the UAV, respectively; K sub Indicates the total number of available drones; y k,i,j Indicates drone k From node i Fly to the node j A binary variable; u k Indicates drone k The utilization status; d i,j Represents a node i To the node j Flight distance; D max and N max These represent the maximum flight range of the drone and the maximum number of wind turbines that can be inspected in a single operation, respectively. order k,i The variable represents the node access order for eliminating sub-loops. Constraint (1a) ensures that each wind turbine node is visited exactly once by a drone. Constraint (1b) stipulates that the drone departs from the base and eventually returns to the base. Constraint (1e) mathematically eliminates illegal sub-loops that bypass the base by assigning a strictly increasing access number to each node visited by the drone.

[0039] (2) Establish a wind turbine health index (HI) prediction and maintenance logic subsystem model based on physical guided neural network (PINN). The Health Index (HI) is influenced by multiple factors and nonlinear coupling. A data-driven degradation function is constructed using PINN: (2a) (2b) in, HI t,n express t Time Fan n Health index (1 for healthy, 0 for need to shut down); IHI t,n Indicates the increase in health deterioration; FNN( ) This represents a physically guided neural network function; V t,n eff , P w,t,n , t t , Tem and Hum These represent the equivalent inflow wind speed, the actual dispatch power of the fan, the cumulative operating time, the ambient temperature, and the ambient humidity, respectively.

[0040] When the fan HI When the threshold value approaches the lower safety limit, an inspection is triggered. The maintenance logic and state transition constraints are as follows: (3a) (3b) in, HI init,n This indicates the initial health value of the fan after it has fully recovered. HI min Indicates the minimum safe and healthy threshold; u w,t,n and Z t,n These represent the operating status and maintenance decision-making variables of the wind turbine, respectively. K uav Indicates the total capacity of the drone fleet; M big and M HI All are large positive constants used for Big-M relaxation. Constraint (3a) establishes the mutually exclusive relationship between wind turbine operation and UAV inspection, while also imposing unit capacity limits and defining boundary conditions for forcibly triggering risk inspections when health indicators approach safety thresholds. Constraint (3b) introduces the "Big-M method" to linearize the conditional state transitions, thereby accurately characterizing the continuous accumulation process of risk in the uninterrupted state, as well as the state where the operational safety confidence is fully restored (i.e., reset to 1) after inspection confirmation.

[0041] (3) Establish a subsystem model for the operation of the land-sea interconnection system and the wake effect. 1) DC power flow and nodal power balance constraints: (4a) (4b) (4c) in, F t,l , B l , F l max They represent the lines respectively. l The actual active power flow, line susceptance, and upper limit of line power flow; I t,i and I t,j Indicates the phase angle of the node voltage; P g,t , P dis,t,e , P ch,t,e , Load t,b , Shed t,b These represent the output of thermal power units, energy storage discharge power, energy storage charging power, power load demand, and load shedding, respectively. Oh b Series symbols indicate connections to the busbar b The collection of various devices.

[0042] 2) Operational constraints of thermal power units and energy storage: (5a) (5b) (5c) (5d) (5e) in, u g,t Indicates the start-up and shutdown status of the thermal power unit; P g min and P g max These represent the minimum and maximum output of the generator; R g u , R g d and belt max The superscript variables represent the normal and maximum start-stop ramp rates; T g on and T g off Minimum continuous start-up and shutdown time; E t,e Indicates the state of charge of the energy storage; or ch and or dis For charge and discharge efficiency; P ch max and P dis max This is for the maximum charge / discharge power limit.

[0043] Equation (5a) represents the upper and lower limits of the unit's output. To take into account the output trajectory of the unit during normal scheduling and start-up / climbing phases, a modified ramping constraint is introduced; the purpose of this constraint is to avoid unsolvable problems caused by the unit's minimum technical output exceeding its standard ramping rate, as shown in Equation (5b); to prevent severe mechanical fatigue caused by frequent start-up and shutdown of the unit, it is stipulated that once the unit starts or stops, it must maintain its corresponding operating / shutdown state for a preset minimum duration, as shown in Equation (5c); Equation (5d) defines the physical operating limits of the energy storage system (ESS) and the continuity constraint of its state of energy (SOE). The first two inequalities strictly define the upper and lower limits of the charging and discharging power of the energy storage device in any time period. The third equation represents the core energy conservation and transfer equation, which accurately accounts for the energy conversion efficiency ( or ch and or dis The energy loss generated by this is considered. The last two constraints are used to enforce a safe capacity operating range (to protect battery health and prevent overcharging and over-discharging) and to set the energy boundary at the start of scheduling, respectively.

[0044] 3) Wind power output and target penalty model considering Jensen wake effect: Assuming a wind turbine i Located upstream, and its wake region varies with downstream distance x It expands linearly. For those located downstream... x ij wind turbines at the location j The wake radius at its location is defined as: (6a) Based on momentum theory, wind turbine i downstream wind turbines j Single velocity loss factor caused at the location d ij It can be represented as: (6b) This rate of loss only applies to wind turbines. j Located in the wind turbine i Within the wake influence area (i.e., lateral distance) d ij lat ≤R wake + R 0) and is located downstream of it ( X ij It is only effective when the value is greater than 0; otherwise... d i j =0.

[0045] Given that wind turbines are typically affected by multiple upstream wind turbines, this application employs a superposition method based on energy conservation to calculate the wind turbine's performance. n At any moment t Equivalent inflow velocity: (6c) Based on the calculated equivalent inflow velocity Maximum available power of wind turbine generators Determined by the power curve of the wind turbine generator set, as follows: (6d) After determining the available wind power capacity Subsequently, this scheduling model allows for wind curtailment in cases of grid congestion or insufficient regulation capacity. The actual output of the wind turbine generators... With the amount of air curtailed Must meet: (6e) (6f) Total penalty cost C pen Including penalties for wind curtailment and load shedding: (6g) in, R wake , R 0、 D , k wake and C TThese represent the wake radius, rotor radius, rotor diameter, wake attenuation constant, and thrust coefficient, respectively. x ij This is the downstream axial distance; d ij This is the speed reduction factor; V t free Free-flow wind speed; P w,t,n avail , P curt,t,n These represent the maximum available power of the wind turbine and the power of the wind curtailed; v in , 、v out , v rate For cut-in, cut-out, and rated wind speed.

[0046] Step 3: Establish a two-level optimization model for the system based on Logical Benders Decomposition (LBBD). A two-layer optimization model is established for the coordinated operation of the sea-land interconnection system.

[0047] 1) Upper-level model (main problem): responsible for power generation scheduling and maintenance decisions, with the objective of minimizing total operating costs. J MP : (7a) in, C gen , C cess , C pen , C wind and These represent the costs of thermal power generation, energy storage cycle aging, wind curtailment and load shedding penalties, wind power operation, and drone fixed maintenance, respectively. or t A Benders proxy variable representing the drone routing cost passed down from the lower layer; This represents the set of all decision variables in the main problem.

[0048] To address the nonlinearity of PINN, an adaptive dynamic tangent plane strategy is proposed, which extracts the physical gradient (sensitivity) using a chain rule. ): (7b) Transform it into a linear inequality constraint (Cut) and add it to the main problem: (7c) (7d) in, This represents the marginal physical gradient of wind turbine health on power adjustment; IHI t,n LP This represents the amount of degradation after linearization; and Indicates the reference power and the optimal pre-scheduled power; IHI true,t,n This represents the actual physical degradation value obtained from the forward propagation of the neural network.

[0049] The flowchart of the adaptive dynamic tangent plane strategy is as follows: Figure 2 As shown.

[0050] 2) Sub-problems (lower-level problems) Optimize drone routes based on upper-level task sets to minimize logistics costs. J sp,t: (8a) in, J sp,t: This represents the cost of drone flight for the lower-level subproblem; C deone_km This indicates the unit cost per kilometer of drone flight.

[0051] (8b) (8c) (8d) Among them, among them, Indicates the current maintenance decision variable Interpreting History Hamming distance between them; Formulas (8c) and (8d) represent the optimality cut and feasibility cut fed back from the lower model to the upper model, respectively; M benders It is a large positive number. In two adjacent iterations, when the difference between the objective functions of the upper and lower layers satisfies the convergence condition, the globally optimal scheduling scheme is obtained.

[0052] 3) Establish the solution process for the two-level optimization model The solution process for the two-layer model considering physics guidance and Benders decomposition constructed in this application can be summarized as follows (e.g.) Figure 3 (as shown) 1) Initialization: Load system operating parameters, including grid parameters, wind turbine layout parameters and pre-trained neural network weights, and then enter the Benders decomposition iteration process.

[0053] 2) Solve the upper-level master problem: Under the current constraint pool, solve the mixed-integer linear programming master problem. The objective is to minimize the total system cost, including thermal power generation, energy storage operation and maintenance costs, and proxy variables passed down from lower levels. Variables to be optimized include thermal power output, energy storage charging and discharging power, maintenance decisions, and wind turbine health index. Constraints on node power balance, energy storage continuity, and wake effects must be satisfied. After solving, extract the wind power dispatch plan and maintenance plan.

[0054] 3) Physical guidance interface verification: Input the wind power dispatch plan output from the upper layer into the neural network gradient check module, and calculate the actual health degradation increment by combining features such as wind speed, time, temperature, and humidity.

[0055] 4) Determine if a physical default has occurred: Compare the degradation amount estimated by the upper-level linearized model with the actual degradation amount output by the neural network. If the error exceeds a preset threshold, dynamically generate a physical guiding tangent plane, add it to the constraint pool of the upper-level main problem, and directly return to step 2 to solve the main problem again. If there is no default, it means that the current scheduling scheme conforms to physical laws, and the maintenance plan is passed to the lower-level subproblem.

[0056] 5) Solving lower-level subproblems: The lower-level vehicle path planning solver receives a fixed maintenance plan and optimizes the flight routes of multiple UAVs.

[0057] 6) Feedback Benders Cut Plane: Check the solution status of the lower-level subproblems. Based on the solution results, generate the corresponding Benders optimality cut or feasibility cut, feed it back to the upper level, and add it to the constraint pool of the main problem.

[0058] 7) Convergence Check: Calculate the iteration gap between the upper and lower objective functions to determine if the convergence condition is met. If it is met, terminate the iteration and output the global optimal result, including the optimal joint scheduling plan, UAV flight trajectory, and the cost of each part of the system. If it is not met, proceed to the next iteration and return to step 2 to solve the upper-level master problem with the new cutting plane.

[0059] This application has the following beneficial effects: (1) A collaborative optimization framework for "Operation & Maintenance (O&M)" for land-sea interconnection systems was constructed, realizing the collaborative optimization of accurate prediction of wind turbine health status and operation and maintenance scheduling. By unifying the modeling and joint solution of wind power scheduling considering the Jensen wake effect and multi-UAV inspection path planning, the maintenance delay caused by the harsh marine environment can be alleviated, and a more flexible balance can be achieved between system operating costs and long-term equipment lifespan. This avoids the disconnect between prediction and scheduling caused by optimizing only a single objective.

[0060] (2) A dynamic linearization method with the introduction of Physically Guided Neural Network (PINN) is proposed, which significantly improves the prediction accuracy of wind turbine health status. The gradient of relevant physical quantities is calculated at the current operating point using the chain rule, which transforms the originally difficult black box mapping into a linear tangent plane with clear theoretical boundary constraints. This allows the optimization solver to follow the actual physical degradation law and safe operation boundary of the wind turbine during iteration, thus solving the problems of instability and poor prediction accuracy of pure data-driven methods in extrapolation scenarios.

[0061] (3) A two-layer solution model based on Logical Benders Decomposition (LBBD) was designed, which significantly improved the solution efficiency and stability of operation and maintenance scheduling. By adaptively generating a physically consistent tangent plane and constructing a combinatorial cut using Hamming distance, the continuous power grid scheduling problem and the discrete UAV path planning problem are structurally decoupled. Compared with directly solving the coupled nonlinear model, the computational load is greatly reduced, the curse of dimensionality of traditional methods is avoided, and the solution efficiency and stability are higher, thus ensuring the real-time performance and reliability of collaborative optimization.

[0062] The following is a specific implementation example of this application: (1) Simulation system parameter settings and data sources In this embodiment, the proposed method is tested using an improved 6-node Garver transmission network, with a cluster of 20 offshore wind turbines connected at node 3. The system simulation has a time resolution of 1 hour and a rolling optimization cycle of 7 days (168 hours). The system's marine meteorological data (including real-time wind speed, ambient temperature, and humidity) is collected from the FINO1 (North Sea and Baltic Sea Research Platform 1) marine meteorological platform. The wind turbine health degradation dataset used to train the Physics-Guided Neural Network (PINN), as well as the detailed operating parameters of the wind turbines, are obtained from the Hill of Towie public wind farm dataset.

[0063] (2) Comparison of schemes and analysis of effects In this embodiment, three schemes are designed for comparison: Option 1: Adopt a two-layer operation and maintenance optimization method, but only configure a single drone (single drone operation and maintenance solution). Option 2: Adopt a two-layer operation and maintenance optimization method, but use a traditional linear model to predict the health status of wind turbines (traditional linear prediction scheme). Option 3: This embodiment proposes a two-layer operation and maintenance optimization method for land-sea interconnection systems that considers Physically Guided Neural Network (PINN) and multi-UAV collaboration.

[0064] The optimization results of the three schemes are explained in detail below with reference to Table 1 and the attached figures.

[0065] Table 1 Comparison of System Operation and Maintenance Cost Optimization Results for Schemes 1-3

[0066] 1. Break through the bottleneck of operation and maintenance resources to achieve optimal overall system cost. As shown in Table 1, Scheme 3 proposed in this embodiment achieves the lowest total system cost ($256,872.68), which is 6.02% and 1.55% lower than Scheme 1 and Scheme 2, respectively. Scheme 1 uses a single drone, resulting in insufficient maintenance resources. When multiple wind turbines simultaneously approach their critical health values, timely maintenance is impossible, necessitating shutdown and incurring a high load shedding penalty cost of $40,622.86. Scheme 2's linear degradation model fails to capture nonlinear fatigue, and its scheduling strategy is too conservative, leading to wind curtailment costs of $8,361.19 and a significant increase in thermal power costs. Scheme 3, through precise physical sensing and multi-drone collaboration, effectively balances power generation efficiency and equipment health.

[0067] 2. Health perception and proactive defense scheduling under extreme weather conditions (combined) Figure 4 and Figure 5 verify) Figure 4 The chart shows the changes in the wind turbine health index and the number of drone dispatch orders over 7 days. On the 3rd day, an extreme storm occurred with wind speeds reaching 21.5 m / s, and the sea conditions were too poor for drones to take off. Figure 5 The wind power output dispatch strategies of Scheme 2 and Scheme 3 on the day of the storm were compared: Scheme 2 (traditional model) did not consider nonlinear fatigue under extreme wind speeds, allowing the turbines to operate at full power. This resulted in an irreversible and rapid decline in the turbine health index, making forced shutdowns highly likely. Scheme 3 (this embodiment), through the sensitivity feedback of the PINN model, found that the degradation cost of continuing high-load operation of the turbines at this point exceeded the cost of supplementing with thermal power. Therefore, during the peak of the storm, from 8 to 12 o'clock, the system proactively reduced turbine output to only about 60% of its rated capacity. This short-term reduction in power generation significantly alleviated the mechanical stress on the turbines and slowed down equipment aging, demonstrating that this scheme, under extreme weather conditions, can ensure both short-term economic viability and long-term protection of equipment assets.

[0068] 3. High-concurrency spatial collaborative routing of multiple UAVs under safe weather windows Following the storm, wind speeds dropped below 13 m / s on the fourth day, creating a suitable meteorological window for operations. Figure 5 At this time, the system had 12 previously backlogged fans that needed to be inspected. Figure 6This is the parallel flight path diagram of multiple drones in the first hour of day 4 in this embodiment. Under the calculation of the lower-level Benders subproblem, the system clusters the wind turbines scattered in various locations according to the distance matrix. Three drones are simultaneously scheduled, and each is assigned a non-overlapping closed-loop route with the shortest total distance, strictly adhering to the restriction that each drone can inspect a maximum of three wind turbines. This spatial coordination method completely eliminates unnecessary empty travel distances, minimizing drone routing costs to only $832.95. Simultaneously, all backlogged inspection tasks are quickly completed, avoiding forced shutdowns due to safety issues. This demonstrates that this embodiment is highly robust and has high engineering practical value when handling high-concurrency operation and maintenance tasks in large offshore wind farms.

[0069] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A neural network-driven dual-layer operation and maintenance method for offshore and onshore wind power based on wind turbine health prediction, characterized in that, Includes the following steps: S10, acquire the operating parameters of the land-sea interconnection system, and establish a wind turbine health prediction model based on a physical guidance neural network, a maintenance path planning subsystem model based on UAVs, and a two-layer optimization model for power generation scheduling and UAV inspection path optimization; the upper layer model of the two-layer optimization model optimizes power generation scheduling and maintenance decisions with the goal of minimizing the overall system operating cost, and the lower layer model optimizes the inspection path based on the UAV-based maintenance path planning subsystem model with the goal of minimizing the UAV flight cost. S20, solve the two-layer optimization model iteratively based on logical Benders decomposition: In each iteration, input the scheduling plan obtained by the upper-layer model into the physical guidance neural network, extract the physical gradient and generate linear tangent plane constraints to be added to the upper-layer model, pass the maintenance decision output by the upper-layer model to the lower-layer model for solution, and feed back the Benders cutting plane to the upper-layer model according to the solution state of the lower-layer model. S30, repeat step S20 until the iteration gap between the upper and lower objective functions meets the convergence condition, and output the optimal scheduling plan and UAV path.

2. The method for dual-layer operation and maintenance of onshore and offshore wind power driven by neural network-based wind turbine health prediction as described in claim 1, characterized in that, In step S10, the wind turbine health prediction model based on the physical guided neural network is used to calculate the health index and health degradation increment of each wind turbine to quantify the degree of wind turbine degradation and trigger maintenance decisions; the health index and health degradation increment are calculated by the following formula: in, HI t,n express t Time Fan n The health index is 1 for healthy and 0 for needing to be shut down. ΔHI t,n Indicates the increase in health deterioration; FNN ) This represents a physically guided neural network function; V t,n eff , P w,t,n , τ t , Tem and Hum These represent the equivalent inflow wind speed, the actual dispatch power of the fan, the cumulative operating time, the ambient temperature, and the ambient humidity, respectively.

3. The method for dual-layer operation and maintenance of onshore and offshore wind power driven by neural network-based wind turbine health prediction as described in claim 2, characterized in that, The health index and health degradation increment are used to establish maintenance logic and state transition constraints to determine the binary variables of maintenance decisions output by the upper-level model. Z t,n The constraint includes: in, HI init,n This indicates the initial health value of the fan after it has fully recovered. HI min Indicates the minimum safe and healthy threshold; u w,t,n and Z t,n These represent the operating status and maintenance decision-making variables of the wind turbine, respectively. K uav Indicates the total capacity of the drone fleet; M big and M HI All of these are large positive constants used for Big-M relaxation.

4. The method for dual-layer operation and maintenance of onshore and offshore wind power driven by neural network-based wind turbine health prediction as described in claim 1, characterized in that, In step S10, the upper-level model aims to minimize the overall system operating cost, and the objective function is: in, C gen , C cess , C pen , C wind and These represent the costs of thermal power generation, energy storage cycle aging, wind curtailment and load shedding penalties, wind power operation, and drone fixed maintenance, respectively. η t A Benders proxy variable representing the drone routing cost passed down from the lower layer; This represents the set of all decision variables in the upper-level model.

5. The method for dual-layer operation and maintenance of onshore and offshore wind power driven by neural network-based wind turbine health prediction as described in claim 4, characterized in that, In step S20, the extraction of physical gradients and generation of linear tangent plane constraints specifically includes: The marginal physical gradient of wind turbine health on power adjustment is calculated using the chain rule. ; Transforming it into the following linear inequality constraint and incorporating it into the upper-level model, the calculation formula is as follows: in, ΔHI t,n LP This represents the amount of degradation after linearization; and Indicates the reference power and the optimal pre-scheduled power; ΔHI true,t,n This represents the actual physical degradation value obtained from the forward propagation of the neural network.

6. The method for dual-layer operation and maintenance of onshore and offshore wind power driven by neural network-based wind turbine health prediction as described in claim 1, characterized in that, In step S20, the objective function of the lower-level model and the Benders cutting plane of the feedback satisfy the following relationship: Lower-level model objective function: ; in, J sp,t: This represents the flight cost of the drone in the lower-level model; C deone_km This indicates the unit cost of flight per kilometer for a drone; d ij Represents a node i To the node j Flight distance; y k,i,j Indicates drone k From node i Fly to the node j A binary variable; The optimality cut or feasibility cut fed back from the lower-level model to the upper-level model is represented as: in, Indicates the current maintenance decision variable Interpreting History Hamming distance between them; M benders It is a large positive number.

7. The method for dual-layer operation and maintenance of onshore and offshore wind power driven by neural network-based wind turbine health prediction as described in claim 1, characterized in that, In step S10, the sea-land interconnection system operation and wake effect subsystem model is used to calculate the equivalent inflow wind speed and the maximum available power of the wind turbine, so as to constrain the actual wind power output. The actual output of the wind power It constitutes part of the scheduling plan described in step S20; The model includes the following formulas: Wake radius: ; Single speed loss factor: ; Equivalent inflow velocity: ; The maximum available power is determined by the power curve of the wind turbine generator set, using the following formula: ; in, R wake , R 0、 D , k wake and C T These represent the wake radius, rotor radius, rotor diameter, wake attenuation constant, and thrust coefficient, respectively. x ij This is the downstream axial distance; δ ij This is the speed reduction factor; V t free Free-flow wind speed; P w,t,n avail , P curt,t,n These represent the maximum available power of the wind turbine and the power of the wind curtailed; v in , 、v out , v rate For cut-in, cut-out, and rated wind speed.

8. The method for dual-layer operation and maintenance of onshore and offshore wind power driven by neural network-based wind turbine health prediction as described in claim 1, characterized in that, In step S10, the UAV-based maintenance path planning subsystem model is used to constrain the task allocation, flight distance and inspection quantity of multiple UAVs in order to generate the path constraint conditions required for the lower-level model to solve. The model includes the following constraints: in, and Represent the set of all physical nodes and the set of target nodes maintained by the UAV, respectively; K sub Indicates the total number of available drones; y k,i,j Indicates drone k From node i Fly to the node j A binary variable; u k Indicates drone k The utilization status; d i,j Represents a node i To the node j Flight distance; D max and N max These represent the maximum flight range of the drone and the maximum number of wind turbines that can be inspected in a single operation, respectively. order k,i This is an auxiliary variable representing the node access order for eliminating sub-loops.

9. The method for dual-layer operation and maintenance of onshore and offshore wind power driven by neural network-based wind turbine health prediction as described in claim 1, characterized in that, In step S20, the specific process of iterative solution includes: Initialization: Load system operating parameters, including grid parameters, wind turbine layout parameters, and pre-trained neural network weights; Solve the upper-level master problem: Solve the mixed-integer linear programming master problem under the current constraint pool, and extract the wind power dispatch plan and maintenance plan; Physical guidance interface verification: Input the wind power dispatch plan into the neural network gradient check module, and calculate the actual health degradation increment by combining wind speed, time, temperature and humidity characteristics; Determine if there is a physical default: Compare the degradation amount estimated by the upper linearization model with the actual degradation amount output by the neural network. If the error exceeds a preset threshold, dynamically generate a physical guiding tangent plane, add it to the constraint pool of the upper master problem, and resolve the upper master problem. If there is no default, pass the maintenance plan to the lower subproblem. Solve the sub-problem at the next level: Given a fixed maintenance plan, optimize the flight paths of multiple UAVs; Feedback Benders Cut Plane: Generate corresponding Benders optimality cuts or feasibility cuts based on the solution status of the lower-level subproblems and feed them back to the upper level, adding them to the constraint pool of the main problem; Convergence check: Calculate the iteration gap between the upper and lower objective functions. If the convergence condition is met, terminate the iteration and output the global optimal result, including the optimal joint scheduling plan, UAV flight trajectory and the cost of each part of the system. If the condition is not met, proceed to the next iteration and solve the upper master problem with the new cutting plane.

10. A neural network-driven dual-layer operation and maintenance system for offshore and onshore wind power with wind turbine health prediction, characterized in that, To implement the method of any one of claims 1 to 9, comprising: The parameter acquisition module is used to acquire the operating parameters of the sea-land interconnection system; The model building module is used to build a wind turbine health prediction model based on a physical guided neural network, a maintenance path planning subsystem model based on UAVs, and a two-layer optimization model for power generation scheduling and UAV inspection path optimization. The upper layer model of the two-layer optimization model optimizes power generation scheduling and maintenance decisions with the goal of minimizing the overall system operating cost, while the lower layer model optimizes the inspection path based on the UAV-based maintenance path planning subsystem model with the goal of minimizing the UAV flight cost. The iterative solution module is used to iteratively solve the two-layer optimization model based on logical Benders decomposition: In each iteration, the scheduling plan obtained by the upper-layer model is input into the physical guidance neural network, the physical gradient is extracted and a linear tangent plane constraint is generated and added to the upper-layer model, the maintenance decision output by the upper-layer model is passed to the lower-layer model for solution, and the Benders cutting plane is fed back to the upper-layer model according to the solution state of the lower-layer model. The convergence determination and output module is used to iterate repeatedly until the iteration gap between the upper and lower objective functions meets the convergence condition, and output the optimal scheduling plan and UAV path.

Images