Wind turbine generator variable pitch system energy management method and system based on super capacitor

Abstract

The invention provides a wind turbine generator variable pitch system energy management method and system based on a supercapacitor, and relates to the technical field of energy management, and the method comprises the steps: collecting the operation data of a wind turbine generator, and building a BEM mapping model of a pitch angle and variable pitch power, and an equivalent circuit model of the supercapacitor. A model prediction control algorithm is adopted to predict the variable pitch power requirement, and the charge state of the supercapacitor is evaluated. An improved alternating direction multiplier method is adopted to cooperatively optimize the BEM model and the equivalent circuit model, and an environment compensation mechanism is adopted to correct an optimization result. A three-layer energy management architecture is constructed, and a fuzzy decision tree is adopted in the top layer to determine an operation mode; the middle layer dynamically adjusts a supercapacitor capacity distribution scheme through reinforcement learning; and the bottom layer carries out real-time energy regulation and control by adopting self-adaptive predictive control based on a BEM model.

Description

Technical Field

[0001] The present invention relates to energy management technologies, and particularly to an energy management method and system for a pitch control system of a wind turbine based on a supercapacitor. Background Art

[0002] In order to improve the efficiency and reliability of wind power generation, pitch control systems are widely used in wind turbines. However, the pitch control system needs to be frequently started and stopped, and the pitch angle needs to be quickly adjusted, which will cause large fluctuations in its power demand and have a certain impact on the stability of the power grid. The existing energy management methods for the pitch control system of wind turbines mainly have the following deficiencies:

[0003] Traditional energy management methods usually adopt rule-based control strategies and cannot effectively cope with complex and variable external environmental factors such as wind speed, resulting in low energy management efficiency.

[0004] Most of the existing methods consider the power demand prediction of the pitch control system and the state of charge assessment of the supercapacitor separately, lacking a collaborative optimization mechanism and making it difficult to achieve optimal control of energy management.

[0005] Traditional energy management methods usually adopt a fixed capacity allocation scheme and cannot dynamically adjust the energy allocation of the supercapacitor according to the actual operating conditions, restricting the performance of the supercapacitor. Summary of the Invention

[0006] Embodiments of the present invention provide an energy management method and system for a pitch control system of a wind turbine based on a supercapacitor, which can solve the problems in the prior art.

[0007] In the first aspect of the embodiments of the present invention,

[0008] An energy management method for a pitch control system of a wind turbine based on a supercapacitor is provided, including:

[0009] Collecting the operation data of the pitch control system of the wind turbine, where the operation data includes the voltage data, current data, temperature data of the supercapacitor, the power data, wind speed data, wind direction data and pitch angle data of the pitch control system; establishing a BEM mapping model between the pitch angle and the pitch power and an equivalent circuit model of the supercapacitor according to the operation data; predicting the power demand sequence of the pitch control system by using a model predictive control algorithm based on the BEM mapping model and combining the wind condition data and the pitch angle data; evaluating the state of charge of the supercapacitor based on the equivalent circuit model and combining the voltage data, current data and temperature data.

[0010] The improved alternating direction multiplier method is used to co-optimize the BEM mapping model and the equivalent circuit model, including: establishing a co-objective function with the power demand sequence deviation, the state of charge deviation, and the power correction amount as the optimization objectives; within each prediction period, iteratively updating the Lagrange multiplier, correcting the power demand sequence and the state of charge evaluation until the relative change amounts of the power demand sequence and the state of charge evaluation value are less than the preset threshold; using an environmental compensation mechanism to correct the power demand sequence and the state of charge after co-optimization;

[0011] Construct a three-layer energy management architecture. At the top layer, based on the corrected power demand sequence and the state of charge, a fuzzy decision tree is used to determine the operating mode of the pitch control system of the wind turbine; at the middle layer, according to the operating mode, the capacity allocation scheme of the emergency reserve area, the power buffer area, and the energy optimization area of the supercapacitor is dynamically adjusted through a reinforcement learning method; at the bottom layer, based on the capacity allocation scheme, an adaptive predictive control based on the BEM mapping model is used to perform real-time energy regulation on the supercapacitor.

[0012] In an alternative embodiment,

[0013] The steps of establishing a BEM mapping model between the pitch angle and the pitch power and an equivalent circuit model of the supercapacitor according to the operation data; predicting the power demand sequence of the pitch control system by using a model predictive control algorithm based on the BEM mapping model in combination with the wind condition data and the pitch angle data; and evaluating the state of charge of the supercapacitor based on the equivalent circuit model in combination with the voltage data, the current data, and the temperature data include:

[0014] The wind turbine blade is divided into multiple blade element micro-elements along the span direction. For each blade element micro-element, the inflow angle is calculated based on the local tip speed ratio and the twist angle, and the Reynolds number is determined in combination with the relative wind speed; the lift coefficient and the drag coefficient are obtained according to the Reynolds number, and a relationship equation between the axial force coefficient and the tangential force coefficient is established; based on the relationship equation, in combination with the Prandtl loss factor considering the tip loss and the hub loss, a blade element mechanical equation including the axial induction factor and the tangential induction factor is established; the blade element mechanical equation is used for iterative calculation. When the axial induction factor is greater than the set threshold, the Glauert correction is introduced, and the inflow angle and the angle of attack are updated until the axial induction factor and the tangential induction factor converge simultaneously; based on the converged axial induction factor and tangential induction factor, the power contribution of each blade element micro-element is calculated, and a BEM mapping model between the pitch angle, the wind speed, and the wind turbine rotational speed and the total pitch power is established through integral operation;

[0015] An equivalent circuit model of the supercapacitor including a series resistor and two parallel RC branches is established. The relationship between the temperature and the circuit parameters is established according to the voltage state equation, and the circuit parameters are corrected by using the temperature compensation coefficient to obtain a second-order RC dynamic characteristic model considering the temperature influence;

[0016] Based on the mapping relationship of the BEM mapping model, an autoregressive moving average model is used to predict the wind speed and wind direction sequences, and the current pitch angle is combined and input into the BEM mapping model to obtain the power demand sequence of the pitch system;

[0017] Based on the second-order RC dynamic characteristic model, an extended Kalman filter algorithm is used to construct a state observer. The voltage data, current data, and temperature data are input into the state observer, and the internal state estimation value including the voltage components of two RC branches is obtained through recursive update. The state of charge of the supercapacitor is calculated according to the internal state estimation value.

[0018] In an alternative implementation,

[0019] The BEM mapping model and the equivalent circuit model are co-optimized by using an improved alternating direction method of multipliers, including: establishing a collaborative objective function with the power demand sequence deviation, the state of charge deviation, and the power correction amount as the optimization objectives; in each prediction period, through iterative updating of the Lagrange multiplier, the corrected power demand sequence, and the corrected state of charge evaluation until the relative change amounts of the power demand sequence and the state of charge evaluation value are less than the preset threshold, the steps include:

[0020] Construct a collaborative objective function including the pitch system and the supercapacitor. The collaborative objective function is composed of a power demand sequence deviation term, a state of charge deviation term, and a power correction amount term, and constraint conditions are set. The constraint conditions include fast-scale dynamic constraints, slow-scale energy balance constraints, and cross-scale coupling constraints; an augmented Lagrangian function is constructed based on the constraint conditions;

[0021] The augmented Lagrangian function is decomposed by using an improved alternating direction method of multipliers, and the original optimization problem is decomposed into a pitch system sub-problem and a supercapacitor sub-problem, and the optimization solution is carried out by alternately solving the two sub-problems and updating the Lagrange multiplier; the improved alternating direction method of multipliers includes designing an adaptive neighborhood trust region constraint to dynamically adjust the trust region radius according to the local curvature of the augmented Lagrangian function; introducing an adaptive weight mechanism for the quadratic penalty term to dynamically adjust the penalty factor based on the constraint violation degree; adopting a heuristic step size selection strategy to determine the optimal step size by combining the Armijo criterion and the Wolfe condition; introducing a momentum term in the Lagrange multiplier update process;

[0022] The improved alternating direction multiplier method is used for solving. In the fast time scale, by constructing the control Hamilton function of the pitch system and based on the Karush-Kuhn-Tucker conditions, the quadratic programming problem of minimizing the power fluctuation of the pitch system is solved. In the slow time scale, by constructing the discrete-time optimality conditions of the supercapacitor and based on the stochastic dynamic programming method, the optimization problem of minimizing the energy loss of the supercapacitor is solved. A coupled correction equation set including the deviation of the power demand sequence and the deviation of the state of charge is constructed, and the Newton-Raphson method is used to iteratively solve the coupled correction equation set to perform co-optimization on the fast time scale and the slow time scale.

[0023] The Kalman filter is used to estimate the states of the pitch system and the supercapacitor and dynamically adjust the weight coefficients of the co-optimization objective function. The time scale separation solution process is repeatedly executed until the relative change amount of the power demand sequence and the evaluated value of the state of charge is less than the preset threshold.

[0024] In an optional implementation manner,

[0025] The steps of constructing a three-layer energy management architecture, where the top layer determines the operating mode of the wind turbine pitch system based on the corrected power demand sequence and the state of charge using a fuzzy decision tree; the middle layer dynamically adjusts the capacity allocation scheme of the emergency reserve area, the power buffer area, and the energy optimization area of the supercapacitor through a reinforcement learning method according to the operating mode; and the bottom layer performs real-time energy regulation on the supercapacitor using adaptive predictive control based on the BEM mapping model based on the capacity allocation scheme include:

[0026] Construct a fuzzy decision tree, use the power demand sequence and its change rate, the state of charge and its change rate as input feature vectors, establish the membership degree of the power demand deviation through an adaptive Gaussian function, construct the membership degree of the state of charge deviation using an S-shaped function introducing a dynamic adjustment factor, where the dynamic adjustment factor adaptively adjusts the slope and intercept of the function according to the state of charge change rate, and determine the operating mode based on the combination of the power demand deviation membership degree and the state of charge deviation membership degree;

[0027] Based on the operating mode, construct a dual-time-scale Actor-Critic network, where the slow-scale Actor network performs global energy allocation based on the corrected state of charge, the corrected power demand, the operating mode, and weather prediction, the fast-scale Actor network performs local power allocation based on the real-time state of charge, power demand, pitch angle, and wind speed, and the Critic network evaluates the value of the global energy allocation strategy and the local power allocation strategy based on a composite reward function constructed based on the charge-discharge efficiency, power quality, capacity life, and pitch stress of the supercapacitor;

[0028] Optimize the state-action distribution of the global energy allocation strategy and the local power allocation strategy using the Wasserstein distance, and dynamically allocate the capacities of the supercapacitor emergency reserve area, the power buffer area, and the energy optimization area based on the optimized global energy allocation strategy and local power allocation strategy;

[0029] Based on the capacity allocation scheme and the BEM mapping model, establish a dual-time-scale predictive control framework, where the fast time scale predicts power fluctuations and the slow time scale predicts energy trends; based on the prediction results, use a rolling optimization algorithm that adaptively adjusts according to the prediction error to perform real-time energy regulation on the supercapacitor. By adaptively adjusting the prediction step according to the operation mode and power change trend, and at the same time setting hard constraints on the state of charge safety boundary and power limit, as well as soft constraints on capacity allocation objectives and efficiency optimization, perform real-time updates of the energy regulation;

[0030] Construct an inter-layer collaborative decision-making framework, use a deep convolutional neural network and a multi-layer decision tree for fault detection, dynamically adjust the backup control strategy based on the fault detection results, and combine the event-triggered mechanism to adjust the power coordination weight between the pitch actuator and the supercapacitor.

[0031] In an alternative implementation,

[0032] The steps for the Critic network to evaluate the value of the global energy allocation strategy and the local power allocation strategy based on a composite reward function constructed from the charge-discharge efficiency, power quality, capacity life, and pitch stress of the supercapacitor include:

[0033] Collect the state of charge and charge-discharge power of the supercapacitor, use a quadratic polynomial to fit the charge efficiency coefficient and the discharge efficiency coefficient, and take the product of the charge-discharge efficiency coefficient and the corresponding charge-discharge power as the charge-discharge efficiency reward value; collect real-time power data, calculate the ratio of the power fluctuation amplitude to the rated power, and obtain the power harmonic components through Fourier transform, calculate the total harmonic distortion rate, and take the weighted sum of the ratio of the power fluctuation amplitude to the rated power and the total harmonic distortion rate as the power quality reward value; use the rainflow counting algorithm to cyclically identify the charge-discharge process of the supercapacitor, obtain the number of cycles and discharge depth data, and calculate the capacity loss in combination with the material attenuation characteristic coefficient, and take the negative value of the capacity loss as the life assessment reward value; collect real-time data of the wind turbine pitch angle, calculate the pitch angular velocity and acceleration, and at the same time monitor the change rate of the pitch torque, and take the weighted negative value of the pitch angular velocity and acceleration and the change rate of the pitch torque as the stress assessment reward value;

[0034] Construct a composite reward function, combine the charge-discharge efficiency reward value, the power quality reward value, the life assessment reward value, and the stress assessment reward value through an adaptive weight coefficient, and the adaptive weight coefficient is updated in real time by the gradient descent method;

[0035] Discount and accumulate the historical data sequence of the composite reward function to obtain the value estimate of the global energy allocation strategy; calculate the composite reward function value at the current moment, and estimate the expected cumulative reward at the current and next moments based on the state of charge value, charge and discharge power, and pitch angle state. By calculating the temporal difference between the reward value at the current moment and the expected cumulative reward, obtain the value estimate of the local power allocation strategy.

[0036] In an alternative embodiment,

[0037] The steps of using the Wasserstein distance to optimize the state-action distributions of the global energy allocation strategy and the local power allocation strategy, and dynamically allocating the capacities of the supercapacitor emergency reserve area, power buffer area, and energy optimization area based on the optimized global energy allocation strategy and local power allocation strategy include:

[0038] Construct the state-action space for global energy allocation. Use the corrected state of charge, corrected power demand, operating mode, and weather prediction data as the state space, and use the capacity allocation ratios of each area of the supercapacitor as the action space. Use the Wasserstein distance to measure the difference between the true distribution and the policy distribution, and make the policy distribution approximate the optimal distribution by iteratively optimizing the policy network parameters to obtain the optimized global energy allocation strategy;

[0039] Construct the state-action space for local power allocation. Use the real-time state of charge, power demand, pitch angle, and wind speed as the state space, and use the real-time power allocation as the action space. Based on the Wasserstein distance, construct the optimization objective, and obtain the optimized local power allocation strategy by iteratively optimizing the policy network parameters;

[0040] Calculate the benchmark capacity allocation ratios of the emergency reserve area, power buffer area, and energy optimization area according to the optimized global energy allocation strategy, and calculate the capacity adjustment amount based on the optimized local power allocation strategy;

[0041] Combine the benchmark capacity allocation ratios and the capacity adjustment amounts through the allocation weight coefficients to obtain the final capacity allocation plan for each area of the supercapacitor, and set the final capacity allocation plan to satisfy the sum of one and the upper and lower limit constraints of the capacities of each area;

[0042] Fuse the charge and discharge efficiency, power quality, life loss, and pitch stress of the supercapacitor through dynamic weight coefficients to obtain the capacity allocation effect evaluation function. Based on the capacity allocation effect evaluation function, use the gradient descent method to dynamically update the allocation weight coefficients and adaptively optimize the final capacity allocation plan.

[0043] In an alternative embodiment,

[0044] Steps for constructing an inter-layer collaborative decision-making framework, using a deep convolutional neural network and a multi-layer decision tree for fault detection, dynamically adjusting the backup control strategy based on the fault detection results, and combining an event-triggering mechanism to adjust the power coordination weights of the pitch actuator and the supercapacitor include:

[0045] Construct a deep convolutional neural network including multi-scale convolutional branches and residual connections. The multi-scale convolutional branches include a feature dimension reduction branch using a small-scale convolutional kernel, a local feature extraction branch using a medium-scale convolutional kernel, and a large-scale feature extraction branch using a large-scale convolutional kernel. Adaptive weight allocation is performed on the output features of each branch through a channel attention mechanism. A batch normalization layer and a Swish activation function are set after each convolutional layer of the deep convolutional neural network, and a fault feature vector including time-domain statistical features, frequency-domain energy features, and time-frequency joint features is extracted.

[0046] Input the fault feature vector into a multi-layer fault diagnosis model based on ensemble learning. In the first layer, a decision tree with an L1 regularization term is used to perform a preliminary classification of the fault. In the second layer, a gradient boosting tree with unilateral sampling is used to divide the fault types. In the third layer, a classification boosting tree with one-hot encoding is used to judge the faulty components. The diagnosis results of each layer are fused through a Softmax voting mechanism to obtain a fault type probability vector and a fault severity evaluation index.

[0047] Construct a control performance evaluation function including a power smoothness index, an energy utilization rate index of the energy storage device, and a device fatigue life index; based on the fault type probability vector and the fault severity evaluation index, use a hierarchical reinforcement learning method to construct a backup control strategy library, and select a basic strategy through a policy network and output a control instruction.

[0048] Use the control performance evaluation function to evaluate the execution result of the control instruction to obtain a control performance evaluation value. Combine the fault type probability vector to construct a multi-objective event-triggering criterion. Design a state trigger based on the Lyapunov stability theory. The state trigger triggers communication updates according to the state deviation of the pitch actuator. Design a time trigger based on the performance decay of the pitch actuator. The time trigger adjusts the sampling period according to the degree of performance decay. Use a fuzzy logic controller to dynamically adjust the trigger thresholds of the state trigger and the time trigger according to the fault severity evaluation index; adjust the power coordination weights of the pitch actuator and the supercapacitor based on the multi-objective event-triggering criterion.

[0049] In the second aspect of the embodiments of the present invention,

[0050] Provide a pitch system energy management system for a wind turbine based on a supercapacitor, including:

[0051] The first unit is used to collect the operation data of the pitch system of a wind turbine. The operation data includes the voltage data, current data, temperature data of the supercapacitor, the power data, wind speed data, wind direction data and pitch angle data of the pitch system. Based on the operation data, a BEM mapping model between the pitch angle and the pitch power and an equivalent circuit model of the supercapacitor are established. Based on the BEM mapping model, combined with the wind condition data and the pitch angle data, a model predictive control algorithm is used to predict the power demand sequence of the pitch system. Based on the equivalent circuit model, combined with the voltage data, current data and temperature data, the state of charge of the supercapacitor is evaluated.

[0052] The second unit is used to co-optimize the BEM mapping model and the equivalent circuit model by using an improved alternating direction method of multipliers, including: establishing a collaborative objective function with the power demand sequence deviation, the state of charge deviation and the power correction amount as the optimization objectives; within each prediction period, by iteratively updating the Lagrange multiplier, correcting the power demand sequence and the state of charge evaluation until the relative change amount of the power demand sequence and the state of charge evaluation value is less than a preset threshold; using an environmental compensation mechanism to correct the co-optimized power demand sequence and the state of charge.

[0053] The third unit is used to construct a three-layer energy management architecture. The top layer, based on the corrected power demand sequence and the state of charge, uses a fuzzy decision tree to determine the operation mode of the pitch system of the wind turbine. The middle layer, according to the operation mode, dynamically adjusts the capacity allocation scheme of the emergency reserve area, the power buffer area and the energy optimization area of the supercapacitor by using a reinforcement learning method. The bottom layer, based on the capacity allocation scheme, uses an adaptive predictive control based on the BEM mapping model to perform real-time energy regulation on the supercapacitor.

[0054] In the third aspect of the embodiments of the present invention,

[0055] There is provided an electronic device, including:

[0056] A processor;

[0057] A memory for storing instructions executable by the processor;

[0058] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0059] In the fourth aspect of the embodiments of the present invention,

[0060] There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0061] Through the collaborative optimization of the BEM mapping model and the equivalent circuit model, as well as the model predictive control algorithm, the present invention can more accurately predict the power demand, and perform optimized control in combination with the state of charge of the supercapacitor, thereby improving the response speed and stability of the pitch system and reducing energy loss.

[0062] Based on the equivalent circuit model and the environmental compensation mechanism of the supercapacitor, the present invention can more accurately evaluate and correct the state of charge of the supercapacitor, avoid overcharging and over-discharging, and thus extend its service life.

[0063] The present invention adopts a three-layer energy management architecture and a reinforcement learning method, which can dynamically adjust the capacity allocation scheme of the supercapacitor according to the actual operating conditions and achieve adaptive predictive control, thereby improving the energy utilization efficiency and enhancing the self-adaptability and robustness of the system. Brief Description of the Drawings

[0064] Figure 1 It is a schematic flow chart of the energy management method for the pitch system of a wind turbine based on a supercapacitor according to an embodiment of the present invention;

[0065] Figure 2 It is a schematic structural diagram of the energy management system for the pitch system of a wind turbine based on a supercapacitor according to an embodiment of the present invention. Detailed Embodiments

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0067] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0068] Figure 1 It is a schematic flow chart of the energy management method for the pitch system of a wind turbine based on a supercapacitor according to an embodiment of the present invention, as Figure 1 shown, the method includes:

[0069] S1. Collect the operation data of the pitch control system of the wind turbine, where the operation data includes the voltage data, current data, temperature data of the supercapacitor, the power data of the pitch control system, the wind speed data, the wind direction data, and the pitch angle data; establish the BEM mapping model between the pitch angle and the pitch power and the equivalent circuit model of the supercapacitor according to the operation data; based on the BEM mapping model, combined with the wind condition data and the pitch angle data, use the model predictive control algorithm to predict the power demand sequence of the pitch control system; based on the equivalent circuit model, combined with the voltage data, current data, and temperature data, evaluate the state of charge of the supercapacitor;

[0070] S2. Use the improved alternating direction method of multipliers to co-optimize the BEM mapping model and the equivalent circuit model, including: establish a collaborative objective function with the power demand sequence deviation, the state of charge deviation, and the power correction amount as the optimization objectives; in each prediction period, through iterative updating of the Lagrange multiplier, correcting the power demand sequence and the state of charge evaluation until the relative change amounts of the power demand sequence and the state of charge evaluation value are less than the preset threshold; use the environmental compensation mechanism to correct the co-optimized power demand sequence and the state of charge;

[0071] S3. Construct a three-layer energy management architecture. The top layer, based on the corrected power demand sequence and the state of charge, uses a fuzzy decision tree to determine the operation mode of the pitch control system of the wind turbine; the middle layer, according to the operation mode, dynamically adjusts the capacity allocation scheme of the emergency reserve area, the power buffer area, and the energy optimization area of the supercapacitor through the reinforcement learning method; the bottom layer, based on the capacity allocation scheme, uses the adaptive predictive control based on the BEM mapping model to perform real-time energy regulation on the supercapacitor.

[0072] In an alternative embodiment,

[0073] The steps of establishing the BEM mapping model between the pitch angle and the pitch power and the equivalent circuit model of the supercapacitor according to the operation data; based on the BEM mapping model, combined with the wind condition data and the pitch angle data, using the model predictive control algorithm to predict the power demand sequence of the pitch control system; based on the equivalent circuit model, combined with the voltage data, current data, and temperature data, evaluating the state of charge of the supercapacitor include:

[0074] The wind turbine blade is divided into multiple chordwise elements along the span direction. For each chordwise element, the inflow angle is calculated based on the local tip speed ratio and the twist angle, and the Reynolds number is determined in combination with the relative wind speed. The lift coefficient and the drag coefficient are obtained according to the Reynolds number, and the relationship equations of the axial force coefficient and the tangential force coefficient are established. Based on the relationship equations, in combination with the Prandtl loss factor considering tip loss and hub loss, a chordwise element mechanical equation including the axial induction factor and the tangential induction factor is established. The chordwise element mechanical equation is used for iterative calculation. When the axial induction factor is greater than the set threshold, the Glauert correction is introduced, and the inflow angle and the angle of attack are updated until both the axial induction factor and the tangential induction factor converge. Based on the converged axial induction factor and tangential induction factor, the power contribution of each chordwise element is calculated, and a BEM mapping model of the pitch angle, the wind speed, and the wind turbine speed to the total pitch power is established through integral operation.

[0075] An equivalent circuit model of a supercapacitor including a series resistor and two parallel RC branches is established. The relationship between the temperature and the circuit parameters is established according to the voltage state equation, and the circuit parameters are corrected by using the temperature compensation coefficient to obtain a second-order RC dynamic characteristic model considering the influence of temperature.

[0076] Based on the mapping relationship of the BEM mapping model, the autoregressive moving average model is used to predict the wind speed and wind direction sequences, and the power demand sequence of the pitch system is obtained by inputting the current pitch angle into the BEM mapping model.

[0077] Based on the second-order RC dynamic characteristic model, an extended Kalman filter algorithm is used to construct a state observer. The voltage data, current data, and temperature data are input into the state observer, and the internal state estimation values including the voltage components of the two RC branches are obtained through recursive update. The state of charge of the supercapacitor is calculated according to the internal state estimation values.

[0078] Exemplarily, first, a blade aerodynamic model is established to obtain the mapping relationship between the pitch power, the pitch angle, the wind speed, and the wind turbine speed. The wind turbine blade is discretized into a number of small chordwise elements along the span direction. For each chordwise element, the local tip speed ratio is calculated, and the inflow angle is calculated according to the blade twist angle. In combination with the oncoming wind speed, the relative wind speed at the position of the chordwise element is calculated, and then the Reynolds number at this position is determined. According to the Reynolds number, the corresponding lift coefficient and drag coefficient are found from the airfoil database or calculated through CFD simulation, and the relationship between the axial force coefficient and the tangential force coefficient and the lift coefficient and the drag coefficient is established.

[0079] Next, consider the eddy current losses at the blade tip and hub, and introduce the Prandtl loss factor to correct the mechanical properties of the blade element. Establish a blade element mechanical equation that includes the axial induction factor and the tangential induction factor. Solve this equation using an iterative method. Initially, set the axial induction factor and the tangential induction factor to zero. During the iteration process, if the axial induction factor is greater than a certain set threshold, then introduce the Glauert correction to update the inflow angle and the angle of attack. Repeat the iterative calculation until both the axial induction factor and the tangential induction factor converge.

[0080] Based on the converged axial induction factor and tangential induction factor, calculate the power contribution generated by each blade element microelement, and integrate the power contributions of all blade element microelements to obtain the total power of the wind turbine. By changing the pitch angle, wind speed, and wind turbine speed, and repeating the above calculation process, a mapping relationship between the pitch angle, wind speed, wind turbine speed, and total pitch power can be established, that is, the BEM (Blade Element Momentum Theory) mapping model. For example, when the wind speed is 10 m / s, the speed is 15 rpm, and the pitch angle is 5 degrees, the pitch power is 100 kW.

[0081] Then, establish an equivalent circuit model of the supercapacitor to evaluate its state of charge. This model consists of a series resistor and two parallel RC branches, which represent the internal resistance of the supercapacitor and different time constant characteristics respectively. Based on the voltage state equation, establish the relationship between temperature and circuit parameters (resistance and capacitance). Since temperature affects the performance of the supercapacitor, a temperature compensation coefficient is introduced to correct the circuit parameters, thereby obtaining a second-order RC dynamic characteristic model considering the influence of temperature. For example, at 25 degrees Celsius, the series resistance is 0.1 ohm, the resistances of the two RC branches are 0.01 ohm and 0.001 ohm respectively, and the capacitances are 100 farads and 1000 farads respectively.

[0082] Next, predict the power demand sequence of the pitch system. Based on the established BEM mapping model, combined with historical wind speed and wind direction data, use the autoregressive moving average model to predict future wind speed and wind direction. Input the predicted wind speed, the current wind turbine speed, and the pitch angle into the BEM mapping model to obtain the power demand sequence of the pitch system. For example, if the predicted wind speed for the next 10 minutes is 12 m / s, the current speed is 15 rpm, and the pitch angle is 5 degrees, then the predicted pitch power demand is 120 kW.

[0083] Finally, the state of charge of the supercapacitor is evaluated. Based on the established second-order RC dynamic characteristic model, an extended Kalman filter algorithm is used to construct a state observer. The voltage, current, and temperature data of the supercapacitor are input into the state observer. Through recursive updating, the state observer obtains the internal state estimation values including the voltage components of the two RC branches. Based on these two voltage components, the state of charge of the supercapacitor can be calculated.

[0084] By accurately predicting the power demand of the pitch system, the present invention can optimize the pitch control strategy, improve the wind energy capture efficiency, and reduce the mechanical wear of the pitch system; accurately evaluating the state of charge of the supercapacitor can avoid overcharging and over-discharging, thereby extending its service life and improving the reliability of the system; by optimizing wind energy capture and increasing the equipment life, the operating cost of the system can be reduced and the economic benefits of wind power generation can be improved.

[0085] In an alternative embodiment,

[0086] The improved alternating direction multiplier method is used to co-optimize the BEM mapping model and the equivalent circuit model, including: establishing a collaborative objective function with the power demand sequence deviation, the state of charge deviation, and the power correction amount as the optimization objectives; in each prediction period, by iteratively updating the Lagrange multiplier, correcting the power demand sequence, and correcting the state of charge evaluation until the relative change amounts of the power demand sequence and the state of charge evaluation value are less than the preset threshold, the steps include:

[0087] Construct a collaborative objective function including the pitch system and the supercapacitor, the collaborative objective function is composed of a power demand sequence deviation term, a state of charge deviation term, and a power correction amount term, and set constraint conditions, the constraint conditions include fast-scale dynamic constraints, slow-scale energy balance constraints, and cross-scale coupling constraints; construct an augmented Lagrangian function based on the constraint conditions;

[0088] The improved alternating direction multiplier method is used to decompose the augmented Lagrangian function, decompose the original optimization problem into a pitch system sub-problem and a supercapacitor sub-problem, and perform optimization by alternately solving the two sub-problems and updating the Lagrange multiplier; the improved alternating direction multiplier method includes designing an adaptive neighborhood trust region constraint, dynamically adjusting the trust region radius according to the local curvature of the augmented Lagrangian function; introducing an adaptive weight mechanism for the quadratic penalty term, dynamically adjusting the penalty factor based on the constraint violation degree; adopting a heuristic step size selection strategy, combining the Armijo criterion and the Wolfe condition to determine the optimal step size; introducing a momentum term in the Lagrange multiplier update process;

[0089] The improved alternating direction multiplier method is used for solving. In the fast time scale, by constructing the control Hamilton function of the pitch system, the quadratic programming problem of minimizing the power fluctuation of the pitch system is solved based on the Karush-Kuhn-Tucker conditions; in the slow time scale, by constructing the discrete-time optimality conditions of the supercapacitor, the optimization problem of minimizing the energy loss of the supercapacitor is solved based on the stochastic dynamic programming method; a coupled correction equation set including the deviation of the power demand sequence and the deviation of the state of charge is constructed, and the Newton-Raphson method is used to iteratively solve the coupled correction equation set to co-optimize the fast time scale and the slow time scale;

[0090] The Kalman filter is used to estimate the states of the pitch system and the supercapacitor and dynamically adjust the weight coefficients of the co-optimization objective function, and the time scale separation solving process is repeatedly executed until the relative change amount of the power demand sequence and the evaluated value of the state of charge is less than the preset threshold.

[0091] Exemplarily, first, a co-optimization objective function is established. The objective function includes three parts: a power demand sequence deviation term, which is used to measure the gap between the actual generated power and the planned generated power; a state of charge deviation term, which is used to measure the gap between the actual state of charge of the supercapacitor and the target state of charge; and a power correction amount term, which is used to measure the degree of correction to the power demand sequence. The goal is to achieve the stable operation of the wind power generation system and the reasonable utilization of the supercapacitor by minimizing this objective function. The constraint conditions include: fast time scale dynamic constraints, such as the physical limitations of the pitch system actuators; slow time scale energy balance constraints, such as the charge and discharge limitations of the supercapacitor; and cross-scale coupling constraints, such as the relationship between the output power of the pitch system and the charge and discharge power of the supercapacitor.

[0092] Next, an augmented Lagrangian function is constructed. The objective function and the constraint conditions are combined to form the augmented Lagrangian function. This function includes the original objective function, the constraint conditions, and the Lagrange multipliers.

[0093] Then, the improved alternating direction multiplier method is used for solving. This method decomposes the original optimization problem into a pitch system sub-problem and a supercapacitor sub-problem, and realizes the overall optimization by alternately solving the two sub-problems. The improved alternating direction multiplier method includes the following key techniques: an adaptive neighborhood trust region constraint, which dynamically adjusts the trust region radius according to the local curvature of the augmented Lagrangian function to ensure the convergence of the algorithm; an adaptive weight mechanism for the quadratic penalty term, which dynamically adjusts the penalty factor according to the degree of constraint violation to improve the robustness of the algorithm; a heuristic step size selection strategy, which combines the Armijo criterion and the Wolfe condition to determine the optimal step size to accelerate the convergence speed of the algorithm; and a momentum term is introduced in the Lagrange multiplier update process to avoid the algorithm falling into a local optimal solution.

[0094] On a fast time scale, by constructing the control Hamiltonian function of the pitch system and solving the quadratic programming problem of minimizing the power fluctuation of the pitch system based on the Karush-Kuhn-Tucker conditions. For example, when the wind speed suddenly increases, the pitch angle is quickly adjusted to reduce the wind energy captured by the wind turbine, thereby limiting the fluctuation of the output power. On a slow time scale, by constructing the discrete-time optimality conditions of the supercapacitor and solving the optimization problem of minimizing the energy loss of the supercapacitor based on the stochastic dynamic programming method. For example, according to the predicted wind speed and power demand, a charging and discharging strategy for the supercapacitor is formulated to minimize the energy loss.

[0095] Construct a coupled correction equation set including the deviation of the power demand sequence and the deviation of the state of charge, and use the Newton-Raphson method to iteratively solve the coupled correction equation set. For example, assume that the initial deviation of the power demand sequence is 10 kW and the deviation of the state of charge is 5%. Through iterative solution using the Newton-Raphson method, the deviation of the power demand sequence is finally reduced to 1 kW and the deviation of the state of charge is reduced to 1%. Co-optimize the fast time scale and the slow time scale to achieve the coordinated control of the pitch system and the supercapacitor.

[0096] Use a Kalman filter to estimate the states of the pitch system and the supercapacitor. For example, using the data from the wind speed sensor, pitch angle sensor, and supercapacitor voltage and current sensors, estimate the real-time power output of the wind turbine and the real-time state of charge of the supercapacitor through the Kalman filter, and dynamically adjust the weight coefficients of the co-objective function. For example, when the wind speed fluctuates greatly, increase the weight of the power demand sequence deviation term to give priority to ensuring the stability of the system. Repeat the time scale separation solution process until the relative change amount of the power demand sequence and the state of charge evaluation value is less than the preset threshold. For example, when the relative change amount of the power demand sequence and the state of charge evaluation value is less than 0.1%, stop the iteration.

[0097] Through the co-optimization of pitch control and supercapacitor energy management, the present invention can effectively suppress the wind power fluctuation and improve the stability of the wind power generation system; by formulating a reasonable charging and discharging strategy, the energy loss of the supercapacitor can be minimized and its service life can be extended; by improving the stability of the wind power generation system and the operating efficiency of the supercapacitor, the cost of wind power generation can be reduced and its economic benefits can be improved.

[0098] In an alternative embodiment,

[0099] Build a three - layer energy management architecture. At the top layer, based on the corrected power demand sequence and state of charge, a fuzzy decision tree is used to determine the operating mode of the pitch system of the wind turbine; at the middle layer, according to the operating mode, the capacity allocation scheme of the emergency reserve area, power buffer area and energy optimization area of the supercapacitor is dynamically adjusted by the reinforcement learning method; the steps of real - time energy regulation of the supercapacitor based on the capacity allocation scheme by using adaptive predictive control based on the BEM mapping model include:

[0100] Build a fuzzy decision tree, take the power demand sequence and its change rate, state of charge and its change rate as input feature vectors, establish the membership degree of power demand deviation through an adaptive Gaussian function, and construct the membership degree of state of charge deviation by using an S - type function introducing a dynamic adjustment factor. The dynamic adjustment factor adaptively adjusts the slope and intercept of the function with the change rate of the state of charge, and determines the operating mode based on the combination of the membership degree of power demand deviation and the membership degree of state of charge deviation;

[0101] Build a two - time - scale Actor - Critic network based on the operating mode. Among them, the slow - scale Actor network conducts global energy allocation based on the corrected state of charge, corrected power demand, operating mode and weather prediction, the fast - scale Actor network conducts local power allocation based on the real - time state of charge, power demand, pitch angle and wind speed, and the Critic network evaluates the value of the global energy allocation strategy and local power allocation strategy based on a composite reward function constructed by the charge - discharge efficiency, power quality, capacity life and pitch stress of the supercapacitor;

[0102] Optimize the state - action distribution of the global energy allocation strategy and local power allocation strategy by using the Wasserstein distance, and dynamically allocate the capacity of the emergency reserve area, power buffer area and energy optimization area of the supercapacitor based on the optimized global energy allocation strategy and local power allocation strategy;

[0103] Establish a two - time - scale predictive control framework based on the capacity allocation scheme and the BEM mapping model, where the fast - time scale predicts power fluctuations and the slow - time scale predicts energy trends; based on the prediction results, use a rolling optimization algorithm that adaptively adjusts according to the prediction error to conduct real - time energy regulation of the supercapacitor. By adaptively adjusting the prediction step according to the operating mode and power change trend, and at the same time setting hard constraints on the state - of - charge safety boundary and power limit and soft constraints on the capacity allocation target and efficiency optimization, the real - time update of energy regulation is carried out;

[0104] Build an inter - layer collaborative decision - making framework, use a deep convolutional neural network and a multi - layer decision tree for fault detection, dynamically adjust the backup control strategy based on the fault detection results, and adjust the power coordination weight between the pitch actuator and the supercapacitor in combination with the event - trigger mechanism.

[0105] Exemplarily, first, a fuzzy decision tree is constructed for top-level operation mode decision-making. The decision tree takes the corrected power demand sequence and its change rate, state of charge and its change rate as input features. To describe the degree of power demand deviation, an adaptive Gaussian function is used to establish a membership function. The membership function of the state of charge deviation is constructed using an S-shaped function with a dynamically adjusted factor. This dynamically adjusted factor can adaptively adjust the slope and intercept of the S-shaped function according to the change rate of the state of charge, so as to more accurately reflect the change trend of the state of charge. For example, when the change rate of the state of charge is large, the dynamically adjusted factor will increase the slope of the S-shaped function, making the membership function more sensitive to the change of the state of charge deviation. Finally, the operation mode is determined according to the combination of the membership function of power demand deviation and the membership function of state of charge deviation. For example, when the power demand deviation is large and the state of charge is high, the system enters the "power priority" mode; when the power demand deviation is small and the state of charge is low, the system enters the "energy reserve" mode.

[0106] Next, a dual-time-scale Actor-Critic network is constructed for dynamic allocation of the middle-layer supercapacitor capacity. The slow-scale Actor network is responsible for global energy allocation, and its inputs are the corrected state of charge, the corrected power demand, the current operation mode, and weather prediction information. For example, if the weather forecast shows weak wind in the future, the slow-scale Actor network will increase the capacity allocation ratio of the supercapacitor energy optimization area to store more energy for emergencies. The fast-scale Actor network is responsible for local power allocation, and its inputs are the real-time state of charge, power demand, pitch angle, and wind speed information. For example, when the wind speed suddenly increases, the fast-scale Actor network will increase the capacity allocation ratio of the supercapacitor power buffer area to absorb excess energy and avoid excessive power fluctuations. The Critic network is used to evaluate the value of global and local power allocation strategies. Its evaluation indicators include the charge and discharge efficiency, power quality, capacity life, and pitch stress of the supercapacitor. The reward function of the Critic network will perform a weighted sum according to the comprehensive performance of these indicators. For example, the weight of charge and discharge efficiency is 0.4, the weight of power quality is 0.3, the weight of capacity life is 0.2, and the weight of pitch stress is 0.1. By optimizing the state-action distribution of global and local power allocation strategies through the Wasserstein distance, the dynamic allocation of the capacities of the supercapacitor emergency reserve area, power buffer area, and energy optimization area is finally realized.

[0107] Finally, a dual-time-scale predictive control framework is established based on the capacity allocation scheme and the BEM mapping model for real-time bottom-layer energy regulation. The fast time scale predicts power fluctuations, and the slow time scale predicts energy trends. For example, the fast time scale predicts the power fluctuation situation in the next 1 minute, and the slow time scale predicts the energy trend in the next 1 hour. Based on the prediction results, a rolling optimization algorithm that adaptively adjusts based on the prediction error is used to perform real-time energy regulation on the supercapacitor. This algorithm adaptively adjusts the prediction step size according to the operating mode and the power change trend. For example, when the system is in the "power priority" mode and the power changes violently, the prediction step size will be shortened to respond more quickly to power fluctuations. At the same time, to ensure the safe and stable operation of the system, the algorithm sets hard constraints on the state of charge safety boundary and power limit, as well as soft constraints on the capacity allocation target and efficiency optimization.

[0108] To further enhance the reliability of the system, an inter-layer collaborative decision-making framework is constructed. This framework uses a deep convolutional neural network and a multi-layer decision tree for fault detection. For example, by analyzing the vibration signals, temperature data, etc. of the wind turbine, it is judged whether there are faults such as gearbox faults and bearing faults. Once a fault is detected, the system dynamically adjusts the backup control strategy and adjusts the power coordination weights of the pitch actuator and the supercapacitor in combination with the event-triggered mechanism. For example, when a pitch system fault is detected, the system reduces the power coordination weight of the pitch actuator and increases the power coordination weight of the supercapacitor to maintain the stable operation of the wind turbine.

[0109] The present invention can effectively smooth the output of wind power, reduce the phenomenon of wind abandonment and power curtailment, thereby improving the utilization rate of wind energy, effectively suppressing the fluctuation of wind power, improving the power quality of the power grid, and enhancing the stability of the power grid by dynamically adjusting the capacity allocation and energy regulation strategy of the supercapacitor; by optimizing the charge and discharge strategy of the supercapacitor, the number and depth of its charge and discharge can be reduced, thereby extending its service life. At the same time, this method can also reduce the stress of the pitch system and extend its service life.

[0110] In an alternative embodiment,

[0111] The steps for the Critic network to evaluate the value of the global energy allocation strategy and the local power allocation strategy based on a composite reward function constructed based on the charge and discharge efficiency, power quality, capacity life, and pitch stress of the supercapacitor include:

[0112] Collect the state of charge and charge-discharge power of the supercapacitor, use a quadratic polynomial to fit the charge efficiency coefficient and the discharge efficiency coefficient, and take the product of the charge-discharge efficiency coefficient and the corresponding charge-discharge power as the charge-discharge efficiency reward value; collect the real-time power data, calculate the ratio of the power fluctuation amplitude to the rated power, and obtain the power harmonic components through Fourier transform, calculate the total harmonic distortion rate, and take the weighted sum of the ratio of the power fluctuation amplitude to the rated power and the total harmonic distortion rate as the power quality reward value; use the rainflow counting algorithm to cyclically identify the charge-discharge process of the supercapacitor, obtain the cycle times and discharge depth data, and calculate the capacity loss in combination with the material attenuation characteristic coefficient, and take the negative value of the capacity loss as the life assessment reward value; collect the real-time data of the wind turbine blade pitch angle, calculate the blade pitch angular velocity and acceleration, and simultaneously monitor the change rate of the pitch torque, and take the weighted negative value of the blade pitch angular velocity and acceleration and the change rate of the pitch torque as the stress assessment reward value;

[0113] Construct a composite reward function, combine the charge-discharge efficiency reward value, the power quality reward value, the life assessment reward value, and the stress assessment reward value through an adaptive weight coefficient, and the adaptive weight coefficient is updated in real time by the gradient descent method;

[0114] Discount and accumulate the historical data sequence of the composite reward function to obtain the value estimate of the global energy allocation strategy; calculate the value of the composite reward function at the current moment, and estimate the expected cumulative reward at the current and next moments based on the state of charge value, the charge-discharge power, and the blade pitch angle state. By calculating the temporal difference between the reward value at the current moment and the expected cumulative reward, obtain the value estimate of the local power allocation strategy.

[0115] Exemplarily, first, collect the state of charge data and charge-discharge power data of the supercapacitor. To evaluate the charge-discharge efficiency, use a quadratic polynomial to fit the charge efficiency coefficient and the discharge efficiency coefficient. Specifically, through a large amount of experimental data, establish a relationship curve between the charge-discharge efficiency and the corresponding power, and fit it with a quadratic polynomial to obtain the corresponding coefficients. For example, through experiments, it is measured that the charge efficiency is 90% at a power of 50kW, 92% at a power of 100kW, 91% at a power of 150kW, and so on. After obtaining multiple groups of data, perform quadratic polynomial fitting. Multiply the charge-discharge efficiency coefficient by the corresponding charge-discharge power to obtain the charge-discharge efficiency reward value.

[0116] Next, collect real-time power data and evaluate power quality. Calculate the ratio of the power fluctuation amplitude to the rated power. For example, if the real-time power fluctuates by 5% above and below the rated power, then this ratio is 0.05. At the same time, obtain the power harmonic components through Fourier transform and calculate the total harmonic distortion rate. For example, the calculated total harmonic distortion rate is 3%. Perform a weighted sum of the ratio of the power fluctuation amplitude to the rated power and the total harmonic distortion rate to obtain the power quality reward value.

[0117] To evaluate the lifespan of the supercapacitor, use the rainflow counting algorithm to perform cycle identification on the charge and discharge process of the supercapacitor, and obtain the cycle count and discharge depth data. For example, 1000 cycles are identified and the discharge depth is 80%. Combine with the material attenuation characteristic coefficient. For example, assume the material attenuation characteristic coefficient is 0.0001 and calculate the capacity loss. For example, the capacity loss is 0.5%. Since the capacity loss represents a performance decline, the negative value of the capacity loss is used as the lifespan evaluation reward value, that is, -0.5%.

[0118] In addition, collect real-time data of the wind turbine pitch angle and evaluate the pitch stress. Calculate the pitch angular velocity and acceleration. For example, the pitch angular velocity is 2° / s and the acceleration is 0.5° / s 2 . At the same time, monitor the change rate of the pitch torque. For example, the change rate of the pitch torque is 10 Nm / s. Take the negative value after performing a weighted sum of the pitch angular velocity, acceleration, and the change rate of the pitch torque as the stress evaluation reward value.

[0119] Combine the above charge-discharge efficiency reward value, power quality reward value, lifespan evaluation reward value, and stress evaluation reward value through an adaptive weight coefficient to construct a composite reward function. The adaptive weight coefficient is updated in real time through the gradient descent method to dynamically adjust the weights of different indicators. For example, the initial weight is set to 0.25 and then adjusted according to the system operation conditions. Perform a discounted accumulation on the historical data sequence of the composite reward function to obtain the value estimate of the global energy allocation strategy. Calculate the value of the composite reward function at the current moment, and based on the state of charge value, charge-discharge power, and pitch angle state, estimate the expected cumulative reward at the current and next moments. Obtain the value estimate of the local power allocation strategy by calculating the temporal difference between the reward value at the current moment and the expected cumulative reward.

[0120] The present invention maximizes the charge-discharge efficiency of the supercapacitor, reduces energy loss, effectively suppresses power fluctuations and harmonics, and improves the stability and reliability of the power grid by optimizing the charge-discharge strategy; by reducing the capacity loss, it prolongs the service life of the supercapacitor and reduces the system maintenance cost.

[0121] In an alternative embodiment,

[0122] Optimize the state-action distribution of the global energy allocation strategy and the local power allocation strategy using the Wasserstein distance. The steps for dynamically allocating the capacities of the supercapacitor emergency reserve area, power buffer area, and energy optimization area based on the optimized global energy allocation strategy and local power allocation strategy include:

[0123] Construct the state-action space for global energy allocation. Use the corrected state of charge, corrected power demand, operating mode, and weather prediction data as the state space, and the capacity allocation ratios of each area of the supercapacitor as the action space. Measure the difference between the true distribution and the policy distribution using the Wasserstein distance, and approximate the optimal distribution by iteratively optimizing the policy network parameters to obtain the optimized global energy allocation strategy;

[0124] Construct the state-action space for local power allocation. Use the real-time state of charge, power demand, pitch angle, and wind speed as the state space, and the real-time power allocation as the action space. Build an optimization objective based on the Wasserstein distance, and obtain the optimized local power allocation strategy by iteratively optimizing the policy network parameters;

[0125] Calculate the benchmark capacity allocation ratios of the emergency reserve area, power buffer area, and energy optimization area according to the optimized global energy allocation strategy, and calculate the capacity adjustment amount based on the optimized local power allocation strategy;

[0126] Combine the benchmark capacity allocation ratios and the capacity adjustment amount through an allocation weight coefficient to obtain the final capacity allocation plan for each area of the supercapacitor, and set the final capacity allocation plan to satisfy the sum of one and the upper and lower limit constraints of the capacities of each area;

[0127] Weightedly fuse the charge and discharge efficiency, power quality, life loss, and pitch stress of the supercapacitor through a dynamic weight coefficient to obtain a capacity allocation effect evaluation function. Based on the capacity allocation effect evaluation function, use the gradient descent method to dynamically update the allocation weight coefficient and adaptively optimize the final capacity allocation plan.

[0128] Exemplarily, first, establish an energy management system model of the supercapacitor. This model includes three areas: the emergency reserve area, power buffer area, and energy optimization area of the supercapacitor, which are used to cope with emergencies, smooth power fluctuations, and optimize energy utilization respectively. The model contains parameters such as the charge and discharge characteristics, efficiency, and life loss of the supercapacitor.

[0129] Next, collect the operation data of the system. These data include the state of charge (SOC) of the supercapacitor, power demand, pitch angle of the wind turbine, wind speed, operation mode, and weather prediction data, etc. The state of charge can be expressed as a percentage. For example, an SOC of 80% means that the current stored electricity of the supercapacitor is 80% of its rated capacity. The power demand can be expressed in kW. For example, the current power demand is 100 kW. The pitch angle can be expressed in degrees. For example, the pitch angle is 15 degrees. The wind speed can be expressed in m / s. For example, the current wind speed is 8 m / s. The operation mode can be divided into normal mode, emergency mode, etc. The weather prediction data can include wind speed prediction, light intensity prediction, etc. Assume that the current SOC is 70%, the power demand is 90 kW, the pitch angle is 10 degrees, the wind speed is 7 m / s, the operation mode is normal mode, and the weather prediction is sunny with relatively stable wind speed.

[0130] Then, construct the global energy allocation strategy. Use the corrected state of charge, corrected power demand, operation mode, and weather prediction data as the input of the state space. The corrected state of charge and power demand refer to adjusting the current state of charge and power demand according to historical data and prediction data. For example, if it is predicted that the power demand will increase in the future period, the corrected power demand can be set higher. Use the capacity allocation ratios of the three regions of the supercapacitor as the action space. For example, the capacity ratio of the emergency reserve area can be set to 20%, the capacity ratio of the power buffer area can be set to 30%, and the capacity ratio of the energy optimization area can be set to 50%. Utilize the reinforcement learning algorithm, such as the deep reinforcement learning algorithm, and adopt the Wasserstein distance as the loss function. By iteratively optimizing the parameters of the policy network, make the policy distribution approach the optimal distribution, and finally obtain the global energy allocation strategy.

[0131] Meanwhile, construct the local power allocation strategy. Use the real-time state of charge, power demand, pitch angle, and wind speed as the input of the state space. Use the real-time power allocation of the three regions of the supercapacitor as the action space. For example, the power allocated to the emergency reserve area at the current moment can be set to 0 kW, the power allocated to the power buffer area can be set to 40 kW, and the power allocated to the energy optimization area can be set to 50 kW. Also utilize the reinforcement learning algorithm and adopt the Wasserstein distance as the loss function. By iteratively optimizing the parameters of the policy network, finally obtain the local power allocation strategy.

[0132] Calculate the benchmark capacity allocation ratios for the three regions according to the global energy allocation strategy. For example, according to the current state and the policy network, the calculated benchmark capacity allocation ratios for the emergency reserve area, the power buffer area, and the energy optimization area are 20%, 30%, and 50% respectively. Calculate the capacity adjustment amounts according to the local power allocation strategy. For example, according to the current state and the policy network, the calculated capacity adjustment amounts for the three regions are +2%, -1%, and -1% respectively.

[0133] Combine the benchmark capacity allocation ratios and the capacity adjustment amounts through the allocation weight coefficients to obtain the final capacity allocation plan. For example, assuming the allocation weight coefficient is 0.8, the final capacity allocation ratios are as follows: Emergency reserve area: 20% + 0.8×2% = 21.6%; Power buffer area: 30% + 0.8×(-1%) = 29.2%; Energy optimization area: 50% + 0.8×(-1%) = 49.2%. And set the final capacity allocation plan to meet the conditions that the sum is one and the upper and lower limits of the capacities of each region are constrained. For example, set the lower limit of the capacity ratio of the emergency reserve area to be 10% and the upper limit to be 30%.

[0134] Finally, construct an evaluation function for the capacity allocation effect. The charge and discharge efficiency, power quality, life loss, and pitch stress of the supercapacitor are weighted and fused through dynamic weight coefficients to obtain the evaluation function for the capacity allocation effect. For example, assuming the current charge and discharge efficiency is 95%, the power quality is 90%, the life loss is 5%, and the pitch stress is 10%, and the dynamic weight coefficients are 0.4, 0.3, 0.2, and 0.1 respectively, the value of the evaluation function for the capacity allocation effect is 0.4×95% + 0.3×90% + 0.2×5% + 0.1×10% = 66%. Dynamically update the allocation weight coefficients using the gradient descent method based on the evaluation function for the capacity allocation effect, and adaptively optimize the final capacity allocation plan.

[0135] Through dynamically adjusting the capacity allocation of each region of the supercapacitor, the present invention can better match the power output characteristics of wind power generation, improve the energy utilization efficiency, and reduce energy waste; a reasonable capacity allocation strategy can avoid frequent deep charge and discharge of the supercapacitor, thereby extending its service life and reducing the maintenance cost; through the setting of the emergency reserve area and the adjustment of the power buffer area, it can effectively respond to emergencies and power fluctuations, and improve the stability and reliability of the wind power generation system.

[0136] In an alternative embodiment,

[0137] The steps of constructing an inter-layer collaborative decision-making framework, using a deep convolutional neural network and a multi-layer decision tree for fault detection, dynamically adjusting the backup control strategy based on the fault detection results, and combining an event-triggered mechanism to adjust the power coordination weight between the pitch actuator and the supercapacitor include:

[0138] Construct a deep convolutional neural network that includes multi-scale convolutional branches and residual connections. The multi-scale convolutional branches include a feature dimensionality reduction branch using a small-scale convolutional kernel, a local feature extraction branch using a medium-scale convolutional kernel, and a large-scale feature extraction branch using a large-scale convolutional kernel. Adaptive weight allocation is performed on the output features of each branch through a channel attention mechanism. After each convolutional layer of the deep convolutional neural network, a batch normalization layer and a Swish activation function are set, and a fault feature vector containing time-domain statistical features, frequency-domain energy features, and time-frequency joint features is extracted;

[0139] Input the fault feature vector into a multi-layer fault diagnosis model based on ensemble learning. In the first layer, a decision tree with an L1 regularization term is used to perform a preliminary classification of the fault. In the second layer, a gradient boosting tree with unilateral sampling is used to divide the fault types. In the third layer, a classification boosting tree with one-hot encoding is used to judge the faulty components. The diagnosis results of each layer are fused through a Softmax voting mechanism to obtain a fault type probability vector and a fault severity evaluation index;

[0140] Construct a control performance evaluation function that includes a power smoothness index, an energy utilization rate index of the energy storage device, and a device fatigue life index; Based on the fault type probability vector and the fault severity evaluation index, a hierarchical reinforcement learning method is used to construct a backup control strategy library, and a basic strategy is selected through a policy network and a control instruction is output;

[0141] Use the control performance evaluation function to evaluate the execution result of the control instruction to obtain a control performance evaluation value. Combine the fault type probability vector to construct a multi-objective event trigger criterion. Design a state trigger based on the Lyapunov stability theory. The state trigger triggers communication updates according to the state deviation of the pitch actuator. Design a time trigger based on the performance decay of the pitch actuator. The time trigger adjusts the sampling period according to the degree of performance decay. Use a fuzzy logic controller to dynamically adjust the trigger thresholds of the state trigger and the time trigger according to the fault severity evaluation index; Adjust the power coordination weight between the pitch actuator and the supercapacitor based on the multi-objective event trigger criterion.

[0142] Exemplarily, first, a deep convolutional neural network is constructed to extract fault features. The network contains multiple branches to capture features at different scales. A branch with a small-scale convolutional kernel is used to reduce the feature dimension. For example, a 3x3 convolutional kernel with a stride of 1 is used, and the dimension is further reduced through a max-pooling layer. A branch with a medium-scale convolutional kernel, such as a 5x5 convolutional kernel with a stride of 1, extracts local features and focuses on local fault information. A branch with a large-scale convolutional kernel, such as a 7x7 convolutional kernel with a stride of 1, extracts large-scale features and captures global fault information. The output features of each branch are weighted and fused through a channel attention mechanism. The channel attention mechanism adaptively adjusts the weights of each channel according to the activation degree of each channel, highlighting important feature channels and suppressing unimportant feature channels. For example, the global average pooling value of each channel can be calculated, and then the weight coefficient of each channel is obtained through two fully connected layers and a Sigmoid activation function. After each convolutional layer of the deep convolutional neural network, a batch normalization layer and a Swish activation function are connected to accelerate the network training speed and improve the network's non-linear expression ability. Finally, the network outputs a fault feature vector containing time-domain statistical features (such as mean, variance, kurtosis, etc.), frequency-domain energy features (such as energy values in different frequency bands), and time-frequency joint features (such as wavelet transform coefficients). For example, this vector can be a 128-dimensional vector.

[0143] Next, the extracted fault feature vector is input into a multi-layer fault diagnosis model for fault diagnosis. The model adopts an ensemble learning method and contains three layers. The first layer uses a decision tree with an L1 regularization term for preliminary classification. For example, faults are classified into three major categories: mechanical faults, electrical faults, and other faults. The L1 regularization term can reduce the complexity of the model and prevent overfitting. The second layer uses a gradient boosting tree with one-sided sampling for fault type division. For example, mechanical faults are further divided into bearing faults, gearbox faults, etc. One-sided sampling can effectively handle the problem of data imbalance. The third layer uses a classification boosting tree with one-hot encoding to determine the faulty component. For example, it determines which specific bearing or gear has failed. One-hot encoding can convert categorical variables into numerical variables, facilitating model processing. Finally, the diagnostic results of the three layers are fused through a Softmax voting mechanism to obtain a fault type probability vector and a fault severity evaluation index. For example, the fault type probability vector can represent the probabilities of various fault types, and the fault severity evaluation index can be a numerical value between 0 and 1. The larger the value, the more severe the fault. Suppose in a case, the model determines that the probability of bearing fault is 0.8, the probability of gearbox fault is 0.1, the probability of other faults is 0.1, and the fault severity evaluation index is 0.7.

[0144] Then, construct a backup control strategy library and select an appropriate control strategy. First, construct a control performance evaluation function that includes power smoothness index, energy utilization rate index of energy storage devices, and device fatigue life index. For example, the power smoothness index can be measured by the standard deviation of power output, the energy utilization rate index of energy storage devices can be measured by the ratio of actual output energy to theoretical maximum output energy, and the device fatigue life index can be measured by the cumulative damage degree. Then, based on the fault type probability vector and the fault severity evaluation index, use the hierarchical reinforcement learning method to construct a backup control strategy library. The strategy library contains various control strategies, such as reducing power output, switching to standby devices, etc. Select the optimal basic strategy through the policy network and output control instructions. For example, if the diagnosis result is a bearing fault and the fault severity is relatively high, the policy network may select the control strategy of reducing power output.

[0145] Finally, dynamically adjust the control strategy according to the fault detection result and control performance evaluation. Use the control performance evaluation function to evaluate the execution result of the control instruction to obtain the control performance evaluation value. Construct a multi-objective event trigger criterion in combination with the fault type probability vector. Design a state trigger based on the Lyapunov stability theory. The state trigger triggers communication update according to the state deviation of the pitch actuator. For example, communication update is only performed when the state deviation exceeds the preset threshold. Design a time trigger based on the performance decay of the pitch actuator. The time trigger adjusts the sampling period according to the degree of performance decay. For example, as the performance decay intensifies, the sampling period is shortened. Use a fuzzy logic controller to dynamically adjust the trigger thresholds of the state trigger and the time trigger according to the fault severity evaluation index. For example, when the fault severity is relatively high, lower the trigger threshold and perform control adjustment more frequently. Finally, adjust the power coordination weight between the pitch actuator and the supercapacitor based on the multi-objective event trigger criterion. For example, when the energy of the supercapacitor is sufficient and the fault severity is relatively low, the power output weight of the supercapacitor can be increased.

[0146] Through the synergistic effect of the deep convolutional neural network and the multi-layer decision tree, the present invention can effectively extract fault features and achieve high-precision fault diagnosis; dynamically adjusting the backup control strategy based on the fault detection result can effectively cope with various fault situations and ensure the stable operation of the system; the event trigger mechanism can reduce unnecessary communication and control operations, improve the system efficiency, and at the same time, combined with the control performance evaluation function, can optimize the overall performance of the system.

[0147] Figure 2 FIG. [X] is a schematic structural diagram of a pitch system energy management system of a wind turbine based on a supercapacitor according to an embodiment of the present invention, as Figure 2 shown, the system includes:

[0148] The first unit is used to collect the operation data of the pitch system of a wind turbine. The operation data includes the voltage data, current data, temperature data of a supercapacitor, the power data, wind speed data, wind direction data and pitch angle data of the pitch system; establish a BEM mapping model of the pitch angle and pitch power and an equivalent circuit model of the supercapacitor according to the operation data; based on the BEM mapping model, combine the wind condition data and pitch angle data, and use a model predictive control algorithm to predict the power demand sequence of the pitch system; based on the equivalent circuit model, combine the voltage data, current data and temperature data to evaluate the state of charge of the supercapacitor.

[0149] The second unit is used to co-optimize the BEM mapping model and the equivalent circuit model by using an improved alternating direction multiplier method, including: establishing a collaborative objective function with the power demand sequence deviation, state of charge deviation and power correction amount as the optimization objectives; in each prediction period, iteratively update the Lagrange multiplier, correct the power demand sequence and correct the state of charge evaluation until the relative change amounts of the power demand sequence and the state of charge evaluation value are less than a preset threshold; use an environment compensation mechanism to correct the co-optimized power demand sequence and state of charge.

[0150] The third unit is used to construct a three-layer energy management architecture. The top layer, based on the corrected power demand sequence and state of charge, uses a fuzzy decision tree to determine the operation mode of the pitch system of the wind turbine; the middle layer, according to the operation mode, dynamically adjusts the capacity allocation scheme of the emergency reserve area, power buffer area and energy optimization area of the supercapacitor by using a reinforcement learning method; the bottom layer, based on the capacity allocation scheme, performs real-time energy regulation on the supercapacitor by using adaptive predictive control based on the BEM mapping model.

[0151] In the third aspect of the embodiments of the present invention,

[0152] A kind of electronic device is provided, including:

[0153] A processor;

[0154] A memory for storing instructions executable by the processor;

[0155] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0156] In the fourth aspect of the embodiments of the present invention,

[0157] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0158] The present invention may be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0159] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A supercapacitor-based wind turbine pitch control system energy management method, characterized in that: include: Collecting the operating data of the variable pitch system of the wind turbine, the operating data including the voltage data, current data, temperature data of the supercapacitor, the power data of the variable pitch system, wind speed data, wind direction data and pitch angle data; establishing a BEM mapping model of the pitch angle and the variable pitch power and an equivalent circuit model of the supercapacitor according to the operating data; based on the BEM mapping model, in combination with the wind condition data and the pitch angle data, using a model predictive control algorithm to predict the power demand sequence of the variable pitch system; based on the equivalent circuit model, in combination with the voltage data, current data and temperature data, evaluating the charge state of the supercapacitor; The BEM mapping model and the equivalent circuit model are collaboratively optimized by using an improved alternating direction multiplier method, including: establishing a collaborative objective function with power demand sequence deviation, state of charge deviation and power correction as optimization targets; in each prediction cycle, iteratively updating the Lagrange multiplier, correcting the power demand sequence and correcting the state of charge evaluation until the relative change of the power demand sequence and the state of charge evaluation value is less than a preset threshold; and using an environmental compensation mechanism to correct the power demand sequence and state of charge after collaborative optimization; A three-layer energy management architecture is constructed. The top layer uses a fuzzy decision tree to determine the operating mode of the wind turbine pitch system based on the corrected power demand sequence and charge state; the middle layer dynamically adjusts the capacity allocation plan of the emergency reserve area, power buffer area and energy optimization area of ​​the supercapacitor through a reinforcement learning method according to the operating mode; the bottom layer uses an adaptive predictive control based on the BEM mapping model to perform real-time energy regulation on the supercapacitor based on the capacity allocation plan.

2. The method according to claim 1, characterized in that The steps of establishing a BEM mapping model of pitch angle and pitch power and an equivalent circuit model of a supercapacitor according to the operation data; predicting a power demand sequence of a pitch system by using a model predictive control algorithm based on the BEM mapping model in combination with wind condition data and pitch angle data; and evaluating a state of charge of a supercapacitor based on the equivalent circuit model in combination with voltage data, current data and temperature data include: The wind rotor blade is divided into multiple blade element micro-elements along the span direction. For each blade element micro-element, the inflow angle is calculated based on the local tip speed ratio and the twist angle, and the Reynolds number is determined in combination with the relative wind speed; the lift coefficient and the drag coefficient are obtained according to the Reynolds number, and the relationship equation between the axial force coefficient and the tangential force coefficient is established; based on the relationship equation, combined with the Prandtl loss factor considering the tip loss and the hub loss, a blade element mechanics equation including the axial induction factor and the tangential induction factor is established; the blade element mechanics equation is used for iterative calculation, and when the axial induction factor is greater than the set threshold, the Glauert correction is introduced, and the inflow angle and the angle of attack are updated until the axial induction factor and the tangential induction factor converge at the same time; the power contribution of each blade element micro-element is calculated based on the converged axial induction factor and the tangential induction factor, and a BEM mapping model of the pitch angle, wind speed, and the wind rotor speed with the total variable pitch power is established through integral calculation; An equivalent circuit model of a supercapacitor including a series resistor and two parallel RC branches is established, a relationship between temperature and circuit parameters is established according to a voltage state equation, the circuit parameters are corrected using a temperature compensation coefficient, and a second-order RC dynamic characteristic model considering the influence of temperature is obtained; Based on the mapping relationship of the BEM mapping model, an autoregressive moving average model is used to predict the wind speed and wind direction sequence, and the current pitch angle is input into the BEM mapping model to obtain the power demand sequence of the variable pitch system; Based on the second-order RC dynamic characteristic model, an extended Kalman filter algorithm is used to construct a state observer, voltage data, current data and temperature data are input into the state observer, and an internal state estimation value including two RC branch voltage components is obtained by recursive updating, and the charge state of the supercapacitor is calculated according to the internal state estimation value.

3. The method according to claim 1, characterized in that The BEM mapping model and the equivalent circuit model are collaboratively optimized by using an improved alternating direction multiplier method, including: establishing a collaborative objective function with power demand sequence deviation, state of charge deviation and power correction as optimization targets; in each prediction cycle, iteratively updating the Lagrange multiplier, correcting the power demand sequence and correcting the state of charge evaluation until the relative change of the power demand sequence and the state of charge evaluation value is less than a preset threshold, including the following steps: Constructing a collaborative objective function including a pitch system and a supercapacitor, wherein the collaborative objective function is composed of a power demand sequence deviation term, a charge state deviation term, and a power correction term, and setting constraints, wherein the constraints include fast-scale dynamic constraints, slow-scale energy balance constraints, and cross-scale coupling constraints; constructing an augmented Lagrangian function based on the constraints; The improved alternating direction multiplier method is used to decompose the augmented Lagrangian function, and the original optimization problem is decomposed into a variable pitch system subproblem and a supercapacitor subproblem, and the optimization solution is performed by alternately solving the two subproblems and updating the Lagrangian multiplier; the improved alternating direction multiplier method includes designing an adaptive neighborhood trust region constraint, dynamically adjusting the trust region radius according to the local curvature of the augmented Lagrangian function; introducing an adaptive weight mechanism of the quadratic penalty term, and dynamically adjusting the penalty factor based on the degree of constraint violation; adopting a heuristic step size selection strategy, combining the Armijo criterion and the Wolfe condition to determine the optimal step size; introducing a momentum term in the Lagrangian multiplier update process; The improved alternating direction multiplier method is used for solving. In a fast time scale, the control Hamiltonian function of the variable pitch system is constructed, and the quadratic programming problem of minimizing power fluctuation of the variable pitch system is solved based on the Karush-Kuhn-Tucker condition. In a slow time scale, the discrete time optimality condition of the supercapacitor is constructed, and the optimization problem of minimizing energy loss of the supercapacitor is solved based on the stochastic dynamic programming method. A coupled correction equation group containing power demand sequence deviation and charge state deviation is constructed, and the coupled correction equation group is iteratively solved by the Newton-Raphson method, and the fast time scale and the slow time scale are collaboratively optimized. The Kalman filter is used to estimate the state of the variable pitch system and the state of the supercapacitor and dynamically adjust the weight coefficient of the collaborative objective function. The time scale separation solution process is repeated until the relative change of the power demand sequence and the state of charge evaluation value is less than the preset threshold.

4. The method according to claim 1, characterized in that A three-layer energy management architecture is constructed. The top layer uses a fuzzy decision tree to determine the operation mode of the wind turbine pitch system based on the corrected power demand sequence and charge state. The middle layer dynamically adjusts the capacity allocation scheme of the supercapacitor emergency reserve area, power buffer area and energy optimization area through reinforcement learning method according to the operation mode. The bottom layer is based on the capacity allocation scheme, and the steps of using adaptive predictive control based on the BEM mapping model to perform real-time energy regulation on the supercapacitor include: A fuzzy decision tree is constructed, and the power demand sequence and its change rate, the state of charge and its change rate are used as input feature vectors. The membership of the power demand deviation is established through an adaptive Gaussian function, and the membership of the state of charge deviation is constructed by using an S-type function that introduces a dynamic adjustment factor. The dynamic adjustment factor adaptively adjusts the function slope and intercept with the state of charge change rate, and determines the operation mode based on the combination of the power demand deviation membership and the state of charge deviation membership; Based on the operation mode, a dual-time-scale Actor-Critic network is constructed, wherein the slow-scale Actor network performs global energy allocation based on the corrected state of charge, corrected power demand, operation mode and weather forecast, and the fast-scale Actor network performs local power allocation based on the real-time state of charge, power demand, pitch angle and wind speed. The Critic network evaluates the value of the global energy allocation strategy and the local power allocation strategy based on a composite reward function constructed based on the charge and discharge efficiency, power quality, capacity life and pitch stress of the supercapacitor; The Wasserstein distance is used to optimize the state-action distribution of the global energy allocation strategy and the local power allocation strategy. Based on the optimized global energy allocation strategy and the local power allocation strategy, the capacity of the supercapacitor emergency reserve area, power buffer area and energy optimization area is dynamically allocated. A dual-time-scale predictive control framework is established based on the capacity allocation scheme and the BEM mapping model, wherein the fast time scale predicts power fluctuations and the slow time scale predicts energy trends; based on the prediction results, a rolling optimization algorithm based on adaptive adjustment of prediction errors is used to perform real-time energy regulation on supercapacitors, and the prediction step size is adaptively adjusted according to the operating mode and power change trend, while hard constraints of the state of charge safety boundary and power limit and soft constraints of the capacity allocation target and efficiency optimization are set to update the energy regulation in real time; An inter-layer collaborative decision-making framework is constructed, and deep convolutional neural networks and multi-layer decision trees are used for fault detection. The backup control strategy is dynamically adjusted based on the fault detection results, and the power coordination weight of the variable pitch actuator and the supercapacitor is adjusted in combination with the event trigger mechanism.

5. The method according to claim 4, characterized in that The steps of evaluating the value of the global energy allocation strategy and the local power allocation strategy by using the composite reward function constructed by the critic network based on the supercapacitor's charge and discharge efficiency, power quality, capacity life, and pitch stress include: Collect the state of charge and charge and discharge power of the supercapacitor, use quadratic polynomials to fit the charging efficiency coefficient and the discharge efficiency coefficient, and use the product of the charging and discharging efficiency coefficient and the corresponding charging and discharging power as the charging and discharging efficiency reward value; collect real-time power data, calculate the ratio of the power fluctuation amplitude to the rated power, and obtain the power harmonic component through Fourier transform, calculate the total harmonic distortion rate, and use the weighted sum of the ratio of the power fluctuation amplitude to the rated power and the total harmonic distortion rate as the power quality reward value; use the rain flow counting algorithm to perform cyclic identification on the charging and discharging process of the supercapacitor, obtain the number of cycles and discharge depth data, calculate the capacity loss in combination with the material attenuation characteristic coefficient, and use the negative value of the capacity loss as the life assessment reward value; collect real-time data on the pitch angle of the wind turbine, calculate the pitch angular velocity and acceleration, and monitor the pitch torque change rate at the same time, and use the weighted negative value of the pitch angular velocity and acceleration and the pitch torque change rate as the stress assessment reward value; Constructing a composite reward function, combining the charge and discharge efficiency reward value, the power quality reward value, the life assessment reward value and the stress assessment reward value through an adaptive weight coefficient, wherein the adaptive weight coefficient is updated in real time through a gradient descent method; The historical data sequence of the compound reward function is discounted and accumulated to obtain a value estimate of the global energy allocation strategy. The value of the compound reward function at the current moment is calculated, and based on the state of charge value, the charge and discharge power, and the pitch angle state, the expected cumulative rewards at the current and next moments are estimated. The value estimate of the local power allocation strategy is obtained by calculating the time difference between the reward value at the current moment and the expected cumulative reward.

6. The method according to claim 4, characterized in that The Wasserstein distance is used to optimize the state-action distribution of the global energy allocation strategy and the local power allocation strategy. The steps of dynamically allocating the capacity of the supercapacitor emergency reserve area, the power buffer area, and the energy optimization area based on the optimized global energy allocation strategy and the local power allocation strategy include: The state-action space of global energy allocation is constructed. The corrected state of charge, corrected power demand, operating mode and weather forecast data are used as the state space, and the capacity allocation ratio of each region of the supercapacitor is used as the action space. The Wasserstein distance is used to measure the difference between the real distribution and the policy distribution. The policy distribution is made close to the optimal distribution by iteratively optimizing the policy network parameters, and the optimized global energy allocation strategy is obtained. Construct the state-action space of local power allocation, take the real-time state of charge, power demand, pitch angle and wind speed as the state space, take the real-time power allocation as the action space, construct the optimization target based on Wasserstein distance, and obtain the optimized local power allocation strategy by iteratively optimizing the strategy network parameters; The base capacity allocation ratios of the emergency reserve area, the power buffer area, and the energy optimization area are calculated based on the optimized global energy allocation strategy, and the capacity adjustment amount is calculated based on the optimized local power allocation strategy; The reference capacity allocation ratio and the capacity adjustment amount are combined by means of an allocation weight coefficient to obtain a final capacity allocation plan for each region of the supercapacitor, and the final capacity allocation plan is set to satisfy the constraints of a sum of one and upper and lower limits of the capacity of each region; The charging and discharging efficiency, power quality, life loss and pitch stress of the supercapacitor are weighted and fused through dynamic weight coefficients to obtain a capacity allocation effect evaluation function. Based on the capacity allocation effect evaluation function, the allocation weight coefficient is dynamically updated using the gradient descent method to adaptively optimize the final capacity allocation scheme.

7. The method according to claim 4, characterized in that The steps of constructing an inter-layer collaborative decision-making framework, using a deep convolutional neural network and a multi-layer decision tree for fault detection, dynamically adjusting the backup control strategy based on the fault detection results, and adjusting the power coordination weight of the variable pitch actuator and the supercapacitor in combination with the event trigger mechanism include: Construct a deep convolutional neural network including multi-scale convolution branches and residual connections, wherein the multi-scale convolution branches include a feature dimension reduction branch using a small-scale convolution kernel, a local feature extraction branch using a medium-scale convolution kernel, and a large-scale feature extraction branch using a large-scale convolution kernel. The output features of each branch are adaptively weighted through a channel attention mechanism. A batch normalization layer and a Swish activation function are set after each convolution layer of the deep convolutional neural network to extract a fault feature vector including time-domain statistical features, frequency-domain energy features, and time-frequency joint features. The fault feature vector is input into a multi-layer fault diagnosis model based on ensemble learning. The first layer uses a decision tree with an L1 regularization term to preliminarily classify the fault. The second layer uses a gradient boosting tree with single-side sampling to classify the fault type. The third layer uses a classification boosting tree with one-hot encoding to determine the faulty component. The diagnosis results of each layer are fused through a Softmax voting mechanism to obtain a fault type probability vector and a fault severity evaluation index. Construct a control performance evaluation function including a power smoothness index, an energy storage device energy utilization index, and an equipment fatigue life index; based on the fault type probability vector and the fault severity evaluation index, a hierarchical reinforcement learning method is used to construct a backup control strategy library, and a basic strategy is selected through a strategy network to output a control instruction; The control performance evaluation function is used to evaluate the execution result of the control instruction to obtain a control performance evaluation value. A multi-objective event triggering criterion is constructed in combination with the fault type probability vector. A state trigger is designed based on the Lyapunov stability theory. The state trigger triggers communication update according to the state deviation of the pitch actuator. A time trigger is designed based on the performance attenuation of the pitch actuator. The time trigger adjusts the sampling period according to the degree of performance attenuation. A fuzzy logic controller is used to dynamically adjust the triggering thresholds of the state trigger and the time trigger according to the fault severity evaluation index. The power coordination weight of the pitch actuator and the supercapacitor is adjusted based on the multi-objective event triggering criterion.

8. A wind turbine pitch system energy management system based on supercapacitors, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect the operation data of the variable pitch system of the wind turbine, wherein the operation data includes the voltage data, current data, temperature data of the supercapacitor, the power data of the variable pitch system, wind speed data, wind direction data and pitch angle data; a BEM mapping model of the pitch angle and the variable pitch power and an equivalent circuit model of the supercapacitor are established according to the operation data; based on the BEM mapping model, in combination with the wind condition data and the pitch angle data, a model predictive control algorithm is used to predict the power demand sequence of the variable pitch system; based on the equivalent circuit model, in combination with the voltage data, current data and temperature data, the charge state of the supercapacitor is evaluated; The second unit is used to collaboratively optimize the BEM mapping model and the equivalent circuit model by using an improved alternating direction multiplier method, including: establishing a collaborative objective function with power demand sequence deviation, state of charge deviation and power correction as optimization targets; in each prediction cycle, iteratively updating the Lagrange multiplier, correcting the power demand sequence and correcting the state of charge evaluation until the relative change of the power demand sequence and the state of charge evaluation value is less than a preset threshold; and using an environmental compensation mechanism to correct the power demand sequence and state of charge after collaborative optimization; The third unit is used to construct a three-layer energy management architecture. The top layer uses a fuzzy decision tree to determine the operating mode of the wind turbine variable pitch system based on the corrected power demand sequence and charge state; the middle layer dynamically adjusts the capacity allocation plan of the supercapacitor's emergency reserve area, power buffer area and energy optimization area through a reinforcement learning method according to the operating mode; the bottom layer uses adaptive predictive control based on the BEM mapping model to perform real-time energy regulation on the supercapacitor based on the capacity allocation plan.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images