A heterogeneous wind farm power optimization method based on agent model and LLM auxiliary mechanism

Abstract

This invention discloses a power optimization method for heterogeneous wind farms based on a surrogate model and an LLM-assisted mechanism, belonging to the field of wind power generation control and intelligent optimization. It includes: establishing a power optimization problem based on a non-uniform three-dimensional wake model; initializing the population and external database, and executing an iterative optimization loop. This loop integrates dynamic hyperparameter decision-making driven by a Large Language Model (LLM), high-quality particle selection based on a random forest surrogate model, adaptive center perturbation replacement combined with population state, and a differentiated particle swarm update strategy; finally, evaluating new solutions using a high-precision model and updating the database, repeating the loop until the optimal unit control strategy is output. This invention significantly improves optimization efficiency, global optimization capability, and engineering practicality by intelligently integrating physical models, surrogate learning, and LLM decision-making, providing an efficient and reliable solution for maximizing the power generation of heterogeneous wind farms.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation, wind power generation control and intelligent optimization, and specifically relates to a method for collaboratively optimizing the power generation of a heterogeneous wind farm containing different types of turbines. Background Technology

[0002] Driven by the "dual carbon" goals, the rapid development of the new energy industry has made upgrading and retrofitting existing wind farms an important way to improve power generation efficiency and revitalize existing resources. This has led to the emergence of a large number of heterogeneous wind farms with mixed arrangements of multiple turbine types. Compared with traditional wind farms with the same turbine type, the significant differences in turbine capacity, rotor diameter, and hub height in heterogeneous wind farms result in complex three-dimensional non-uniform characteristics in the wake field, and the aerodynamic coupling effect between turbines is greatly enhanced.

[0003] Currently, wind farm power optimization and wake control technologies have formed a relatively complete research system. Early wake models were mostly based on the assumption of uniform inflow, such as the classic Jensen model and the two-dimensional Gaussian model. Although they were computationally efficient, their prediction errors were significantly higher in heterogeneous scenarios. Subsequent CFD methods, such as large eddy simulation, could accurately reconstruct the wake evolution process, but they were computationally expensive and time-consuming. To balance accuracy and efficiency, three-dimensional wake models have been gradually proposed and adopted in engineering: Li et al. proposed a three-dimensional analytical wake model that considers the influence of atmospheric stability. By introducing a stability parameter, they corrected the traditional wake expansion rate, providing physical support for all-weather wind farm power prediction; Chen et al. developed a three-dimensional wake perception layout optimization framework called WAKE-NET, proving that a three-dimensional spatial staggered layout can significantly improve the annual power generation of the entire farm.

[0004] In terms of optimized control, existing technologies mostly employ yaw control, axial induction factor adjustment, or a combination of both strategies to reduce wake loss and increase total power. Bossanyi et al. developed a dynamically robust active wake control strategy to address the dynamic variability of actual wind conditions; Tamaro et al. proposed a robust active power control algorithm that maximizes the power tracking margin of wind farms under wake conditions by combining axial induction control with wake yaw steering; Liu et al. systematically studied the synchronous control strategy of yaw angle, rotor speed, and pitch angle, revealing that multi-variable combinations can better utilize the "secondary wake steering effect" to guide airflow, enabling downstream units to obtain more inflow energy. However, traditional optimization algorithms require frequent calls to high-precision wake models to calculate fitness. In large-scale wind farms with a large number of units and high control variable dimensionality, computational costs rise sharply and convergence speed decreases significantly, making it difficult to meet engineering practicality requirements.

[0005] Furthermore, most existing technologies are geared towards homogeneous wind farm design, lacking systematic optimization methods for on-site replacement of heterogeneous turbines. They fail to fully consider practical constraints such as turbine differences, layout variations, wind direction and speed fluctuations, and construction downtime, resulting in gaps between optimization results and engineering applications, and failing to fully unleash the power enhancement potential of wind farm upgrades. Simultaneously, existing intelligent optimization algorithms often employ preset linear decreasing strategies for hyperparameters, relying on manual experience for setting. They cannot dynamically adjust hyperparameters based on real-time population diversity, convergence speed, and wake model prediction errors, leading to premature convergence and insufficient robustness in complex heterogeneous wind farm scenarios.

[0006] In summary, current heterogeneous wind farm power optimization technologies generally suffer from problems such as insufficient wake modeling accuracy, high computational cost of high-dimensional optimization, premature convergence of algorithms, and weak engineering practicality, making it impossible to achieve efficient, stable, and globally optimal power optimization. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a heterogeneous wind farm power optimization method based on a surrogate model and an LLM-assisted mechanism, in order to solve the technical problems that are common in current heterogeneous wind farm power optimization technologies, such as insufficient wake modeling accuracy, high computational cost of high-dimensional optimization, premature convergence of algorithms, and weak engineering practicality.

[0008] The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism of this invention includes:

[0009] Step 1): Using the axial induction factor and yaw angle of each wind turbine in the heterogeneous wind farm as decision variables, establish an optimization problem with maximizing the total output power of the wind farm as the objective function. The total output power of the wind farm is calculated based on a non-uniform three-dimensional wake model according to the following steps:

[0010] 11) Wake Velocity Field Calculation: For any point within the wind farm, the directional wind speed affected by the wake of the upstream turbines is calculated using the following formula:

[0011]

[0012] in, These are the coordinates of the point in the downstream direction, horizontal direction, and vertical direction;

[0013] To account for the initial wind speed caused by wind shear, the calculation formula is as follows:

[0014]

[0015] in For free-flowing wind speed; The wind shear index. This refers to the height of the wind turbine hub;

[0016] This is the lateral offset of the wake centerline relative to the rotor center;

[0017] and These represent the characteristic widths of the wake in the lateral and vertical directions, respectively, and satisfy a linear expansion relationship:

[0018]

[0019] in, The diameter of the upstream wind turbine rotor. The starting position of the far wake. and These represent the lateral and vertical wake expansion rates, respectively. and The initial wake width at the far wake initiation position;

[0020] The velocity deficit coefficient at the center of the wake. Initial thrust coefficient:

[0021]

[0022] in, It is the axial induction factor;

[0023] 12) Calculation of multiple wake superposition: For downstream points located in the wake superposition area of ​​several upstream units, the total velocity loss is superimposed based on the velocity loss caused by each upstream unit individually, according to the principle of conservation of average kinetic energy loss.

[0024] 13) Calculation of equivalent wind speed for a single unit: For the first... The Taiwanese generator set discretizes its rotor plane into... For each measuring point, calculate the effective incoming air velocity based on steps 11) and 12). The rotor's average effective incoming air velocity is calculated using the following formula. :

[0025]

[0026] 14) Calculation of total wind farm power: based on the equivalent wind speed of all units. Their respective power curves and yaw angle Calculate the total output power of the wind farm. :

[0027]

[0028] in This represents the total number of wind turbines in a heterogeneous wind farm. The yaw power loss index;

[0029] Step 2): Initialize a population and an external database; wherein each particle in the population represents a set of decision variables for all wind turbine generators; the external database is used to store the particles and their corresponding historical fitness values, wherein the fitness values ​​are the total output power of the wind farm calculated in Step 1);

[0030] Step 3): Based on the initialized particle swarm and external database, execute an iterative optimization loop, performing the following steps in each iteration:

[0031] 31) Dynamic hyperparameter decision-making: The current state feature vector of the population is collected and input into the controller assisted by the large language model. The state feature vector includes at least population diversity, convergence slope and surrogate model prediction error. The controller infers and outputs a set of dynamic hyperparameters in real time through nonlinear mapping based on the state feature vector, including perturbation factor, surrogate selection ratio and adaptive particle swarm optimization scaling factor.

[0032] 32) Proxy-assisted selection: Train a random forest proxy model based on the external database, and use the random forest proxy model to predict the fitness value of each particle in the current population; sort the predicted fitness values ​​of all particles in the current population in descending order, and calculate the number of high-quality particles to be selected in this iteration according to the proxy selection ratio to form a subset of high-quality particles;

[0033] 33) Calculate the perturbation and replacement solution of the center: Calculate the geometric center of the high-quality particle subset in the decision space, and generate a perturbation replacement solution after boundary correction based on the perturbation scaling factor and the standard deviation of the current population in each decision variable dimension;

[0034] 34) Adaptive Particle Swarm Optimization Update: For particles not selected into the high-quality particle subset, their position and velocity are updated according to the standard particle swarm optimization algorithm rules, wherein the velocity update term is scaled by the adaptive particle swarm optimization scaling factor; for particles selected into the high-quality particle subset, the particle with the worst fitness value in the high-quality particle subset is replaced with the perturbation replacement solution with boundary correction in step 33).

[0035] Step 4): Calculate the actual fitness value of the new particles generated after the update in Step 34), and store these new particles and their actual fitness values ​​as data pairs in the external database for use in subsequent iterations to update the random forest proxy model;

[0036] Step 5): Repeat steps S3 to S4 until the preset termination condition is met, and output the optimal combination of decision variables that maximizes the objective function value.

[0037] Furthermore, the optimization problem described in step 1), which aims to maximize the total output power of the wind farm, is mathematically defined as follows:

[0038]

[0039] in, As an axial inducing factor, Yaw angle Let P0 be the objective function, and P0 be the wind farm reference power corresponding to each wind turbine under independent optimal operating conditions. For a given free-flowing wind speed and wind direction Below, corresponding to the decision variables The total output power of the wind farm is calculated at that time.

[0040] Furthermore, in step 31), the optimized state feature is a five-dimensional feature vector:

[0041]

[0042] in For population diversity; The slope of convergence; For fitness variance; The prediction error of the random forest proxy model is... This refers to the iteration progress.

[0043] Furthermore, in step 31), the large language model-assisted hyperparameter decision controller dynamically outputs the dynamic hyperparameters using the following formula:

[0044]

[0045] in, This is the perturbation scaling factor. Choose the ratio for the agent. To optimize the scaling factor for adaptive particle swarm optimization, , and For the preset constant boundary value, , , The coefficients are generated by the large language model based on the normalized feature vectors, with subscripts i=1,2,3.

[0046] Furthermore, in step S32, the random forest surrogate model is trained on samples selected from the historical database, and its prediction result is the average of the outputs of all decision trees; the number of particles in the high-quality particle subset is obtained by the following formula:

[0047]

[0048] Where t is the iteration number index. To dynamically adjust the number of particles selected, where N is the population size, Select a ratio for the agent.

[0049] Furthermore, in step 32), the fitness value of the particle is obtained by averaging the prediction results of all decision trees by the random forest surrogate model:

[0050]

[0051] in For the number of decision trees, For the first The combination of decision variables represented by each particle;

[0052] The number of high-quality particles in step 32) is calculated as follows:

[0053]

[0054] in: Where N is the number of high-quality particles in the current iteration step, and N is the population size. Choose the ratio for the agent. This is a rounding function.

[0055] Furthermore, in step 33), the geometric center of the high-quality particle subset in the decision space... It is generated by the following formula:

[0056]

[0057] Where t is the iteration number index. Describes the set of indices of high-quality particles selected in the t-th iteration. For the first high-quality particle subset The position of each particle;

[0058] The perturbation-adjusted replacement solution is generated using the following formula:

[0059]

[0060] in, This is the perturbation scaling factor. Let be a d-dimensional random perturbation vector, where d is the total number of decision variables, and each of its components is... The generation follows an adaptive mechanism:

[0061]

[0062] Where N represents a Gaussian distribution with a mean of 0 and a variance of 0. Let be the sample standard deviation of the entire population in the j-th decision variable dimension;

[0063] Replacement solution after perturbation Boundary corrections are performed to obtain the final perturbation replacement solution. :

[0064]

[0065] in, and These are the preset lower and upper limits of the j-th decision variable, respectively.

[0066] Furthermore, the position and velocity updates of the particles are defined by the following two formulas:

[0067]

[0068] in, Select a label for the agent when the particle When selected into a high-quality subset of particles, =1; otherwise =0;

[0069] Replace the label for the worst particle when the particle When an individual is determined to be the least fit in the current population and is replaced, =1, otherwise =0;

[0070] Optimize scaling factor for adaptive particle swarm optimization;

[0071] and They represent particles respectively The velocity vectors at the current and next moments. The inertia coefficient, For particles The best historical position This is the historically optimal position for the entire population. and These are the individual cognitive coefficient and the social learning coefficient, respectively. and Random disturbance coefficient.

[0072] The beneficial effects of this invention are:

[0073] 1. This invention employs a non-uniform three-dimensional wake model capable of characterizing three-dimensional aerodynamic interference between units, overcoming the limitations of insufficient accuracy of traditional two-dimensional or uniform wake models in heterogeneous scenarios. By introducing mechanisms such as wind shear, linear wake extension, and superposition of multiple wake kinetic energies, this model provides a physical calculation basis for subsequent optimization that approximates the real flow field, ensuring the accuracy of the optimization objective, which is a prerequisite for achieving effective power enhancement.

[0074] 2. This invention creatively introduces a random forest surrogate model and deeply couples it with a hyperparameter controller assisted by a Large Language Model (LLM), forming an intelligent computing resource allocation mechanism. Traditional methods use a fixed proportion of the surrogate model, which can easily lead to deviations in the early and mid-stage search direction. This invention uses LLM to analyze the prediction error and convergence state of the surrogate model in real time, dynamically adjusting the intervention ratio (k...) of the surrogate model. t The algorithm significantly utilizes the surrogate model in the early stages of the search and when the surrogate confidence is high to accelerate exploration, and then reverts to the physical model in the later stages of convergence or when the surrogate error is large to ensure accuracy. This dynamic, feedback-based surrogate-physical model collaboration mechanism fundamentally solves the contradiction between high-precision computation and optimization efficiency, and achieves a significant reduction in computational overhead.

[0075] 3. Another key innovation of this invention lies in using LLM to replace traditional linear / empirical formulas, enabling online sensing, real-time decision-making, and adaptive control of the optimization process. The LLM controller integrates multi-dimensional states such as population diversity (D) and convergence gradient (ΔG), and outputs the perturbation intensity (α) in real time through nonlinear mapping. t Particle swarm scaling factor (η) t Key parameters such as [list of parameters] are included. This allows the algorithm to dynamically adjust according to the search terrain: when population diversity decreases, it intelligently enhances perturbations to escape local optima; when convergence stalls, it adjusts the particle update step size to restore vitality. This upgrades the optimization algorithm from "fixed strategy execution" to "intelligent situational response," thoroughly enhancing the algorithm's global convergence ability and robustness in complex, high-dimensional, multi-peak heterogeneous wind farm optimization problems.

[0076] 4. This invention aims to maximize the overall power increase and coordinates the yaw angle and axial induction factor of each unit, which can significantly reduce wake superposition losses. Compared with the optimal single-unit operation mode, the total power generation of the wind farm is increased by a greater margin and the optimization effect is more prominent.

[0077] 5. The present invention has stronger engineering applicability and versatility. The method can be adapted to various heterogeneous wind farm layouts such as single row, regular grid, and large-scale staggered layout. It is compatible with actual engineering scenarios such as multiple wind directions, multiple wind speeds, on-site replacement of units, and construction shutdown. The optimization results are more in line with the needs of on-site applications and can provide a direct, usable, intelligent and stable operation optimization solution for the upgrading and transformation of heterogeneous wind farms. Attached Figure Description

[0078] Figure 1 The flowchart shows the adaptive particle swarm optimization algorithm. Detailed Implementation

[0079] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0080] The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism in this embodiment includes:

[0081] Step 1): Using the axial induction factor and yaw angle of each wind turbine in the heterogeneous wind farm as decision variables, establish an optimization problem with maximizing the total output power of the wind farm as the objective function. The optimization problem is mathematically defined as follows:

[0082]

[0083] in, As an axial inducing factor, Yaw angle Let P0 be the objective function, and P0 be the wind farm reference power corresponding to each wind turbine under independent optimal operating conditions. For a given free-flowing wind speed and wind direction Below, corresponding to the decision variables The total output power of the wind farm is calculated at that time.

[0084] The total output power of the wind farm is calculated based on a non-uniform three-dimensional wake model using the following steps:

[0085] 11) Wake Velocity Field Calculation: For any point within the wind farm, the directional wind speed affected by the wake of the upstream turbines is calculated using the following formula:

[0086]

[0087] in, These are the coordinates of the point in the downstream direction, horizontal direction, and vertical direction;

[0088] To account for the initial wind speed caused by wind shear, the calculation formula is as follows:

[0089]

[0090] in For free-flowing wind speed; The wind shear index. This refers to the height of the wind turbine hub;

[0091] This is the lateral offset of the wake centerline relative to the rotor center;

[0092] and These represent the characteristic widths of the wake in the lateral and vertical directions, respectively, and satisfy a linear expansion relationship:

[0093]

[0094] in, The diameter of the upstream wind turbine rotor. The starting position of the far wake. and These represent the lateral and vertical wake expansion rates, respectively. and The initial wake width at the far wake initiation position;

[0095] The velocity deficit coefficient at the center of the wake. Initial thrust coefficient:

[0096]

[0097] in, It is the axial induction factor.

[0098] 12) Multiple Wake Overlap Calculation: For downstream points located in the overlapping wake regions of several upstream turbines, the total velocity loss is calculated based on the velocity losses caused by each upstream turbine individually, and is overlaid according to the principle of conservation of average kinetic energy loss. The velocity loss at the p-th measuring point on the rotor surface of the i-th wind turbine can be calculated using the following formula:

[0099]

[0100] in, Indicates the upstream number Typhoon turbine at the wind measurement point The wake velocity at the point can be used to calculate the effective flow distribution of the downstream unit point by point on the rotor plane, providing a basis for subsequent calculations of rotor average wind speed and single unit power.

[0101] 13) Calculation of equivalent wind speed for a single unit: For the first... The Taiwanese generator set discretizes its rotor plane into... For each measuring point, calculate the effective incoming air velocity based on steps 11) and 12). The rotor's average effective incoming air velocity is calculated using the following formula. :

[0102]

[0103] 14) Calculation of total wind farm power: based on the equivalent wind speed of all units. Their respective power curves and yaw angle Calculate the total output power of the wind farm. :

[0104]

[0105] in This represents the total number of wind turbines in a heterogeneous wind farm. This is the yaw power loss index.

[0106] Step 2): Initialize a population and an external database; where each particle in the population represents a set of decision variables for all wind turbine generators (i.e., the set of axial induction factors and yaw angles); the external database is used to store particles and their corresponding historical fitness values, where the fitness values ​​are the total output power of the wind farm calculated in Step 1).

[0107] Step 3): Based on the initialized particle swarm and external database, execute an iterative optimization loop, performing the following steps in each iteration:

[0108] 31) Dynamic Hyperparameter Decision-Making: The current population's state feature vector is collected and input to the Large Language Model (LLM)-assisted controller. In this embodiment, the state feature vector includes at least population diversity (D), convergence slope (D), and other parameters. ) and surrogate model prediction error ( ), fitness variance ( ) and iteration progress ( The state feature vector expression is as follows:

[0109]

[0110] For each channel of the input features, feature normalization processing must be performed:

[0111]

[0112] In the formula, Indicates the first The original observations of each state feature in this iteration. and They represent the first The minimum and maximum values ​​of each state feature within the historical operation or preset range.

[0113] The controller, based on the state feature vector, infers in real time through nonlinear mapping and outputs a set of dynamic hyperparameters, including a disturbance factor ( ), Agent selection ratio ( ) and adaptive particle swarm optimization scaling factor ( The dynamic hyperparameter group expression is as follows:

[0114]

[0115] The large language model-assisted hyperparameter decision controller dynamically outputs the dynamic hyperparameters using the following formula:

[0116]

[0117] in, This is the perturbation scaling factor. Choose the ratio for the agent. To optimize the scaling factor for adaptive particle swarm optimization, , and For the preset constant boundary value, , , The coefficients are inferred by the large language model based on the normalized feature vectors, with subscripts i=1,2,3;

[0118]

[0119] Among them The inference mapping function follows the nonlinear transformation criterion of wind power dynamics and optimization theory to break the traditional linear evolution constraint of parameters.

[0120] When the controller detects a decrease in population distribution entropy, i.e. a decrease in diversity, it nonlinearly increases the perturbation gain. To force the maintenance of search space coverage; when the convergence gradient approaches zero and is in the early stage of the search, it is determined to be a premature convergence state, and the controller generates a pulsed perturbation command. Guide the population to escape local optima; when residuals increase, instruct to reduce the proxy ratio. The correctness of the search direction is ensured by reverting to a high-precision physical model; as the iteration progresses, the controller reduces [its sensitivity] in the later stages. and This allows the algorithm to smoothly switch from global oscillation to local fine-grained optimization; at the same time, the controller can integrate population diversity and fitness variance indices to dynamically correct the adaptive particle swarm optimization scaling factor, thereby achieving condition-adaptive adjustment of particle inertia and social learning ability.

[0121] 32) Proxy-assisted selection: When the number of sample data pairs stored in the external database reaches or exceeds a preset threshold, a random forest proxy model is trained based on the external database, and the fitness value of each particle in the current population is predicted using the random forest proxy model; the predicted fitness values ​​of all particles in the current population are sorted in descending order, and the selection is based on the proxy selection ratio ( Calculate the number of high-quality particles to be selected in this iteration, and form a subset of high-quality particles.

[0122] The fitness value of the particle is obtained by averaging the prediction results of all decision trees by the random forest surrogate model:

[0123]

[0124] in For the number of decision trees, For the first The combination of decision variables represented by each particle.

[0125] Predicted fitness values ​​for all particles in the t-th iteration Sort in descending order:

[0126]

[0127] Where N represents the population size. This represents the sorted index sequence. (Before selection) A subset of high-quality particles The number of high-quality particles is calculated as follows:

[0128]

[0129] Where: N is the population size. Choose the ratio for the agent. This is a rounding function.

[0130] Based on high-quality particle subsets Define the label for each particle. for:

[0131]

[0132] In the early stages of iteration, a more lenient selection mechanism is used to enhance exploration capabilities; in the later stages, a stricter selection mechanism is adopted to accelerate convergence. This is achieved by dynamically adjusting the number of selected particles. It has achieved a transition from an early search phase focused on global exploration to a later search phase focused on local development.

[0133] 33) Calculate the perturbation and replacement solution of the center: Calculate the geometric center of the high-quality particle subset in the decision space, and generate a perturbation replacement solution after boundary correction based on the perturbation scaling factor and the standard deviation of the current population in each decision variable dimension.

[0134] The geometric center of the high-quality particle subset in the decision space It is generated by the following formula:

[0135]

[0136] Where t is the iteration number index. Describes the set of indices of high-quality particles selected in the t-th iteration. For the first high-quality particle subset The position of each particle.

[0137] The perturbation-adjusted replacement solution is generated using the following formula:

[0138]

[0139] in, This is the perturbation scaling factor. Let be a d-dimensional random perturbation vector, where d is the total number of decision variables, and each of its components is... The generation follows an adaptive mechanism:

[0140]

[0141] Where N represents a Gaussian distribution with a mean of 0 and a variance of 0. Let be the sample standard deviation of the entire population in the j-th decision variable dimension.

[0142] Replacement solution after perturbation Boundary corrections are performed to obtain the final perturbation replacement solution. :

[0143]

[0144] in, and These are the preset lower and upper limits of the j-th decision variable, respectively.

[0145] 34) Adaptive Particle Swarm Optimization Update: For particles not selected into the high-quality particle subset, their position and velocity are updated according to the standard particle swarm optimization algorithm rules, wherein the velocity update term is determined by the adaptive particle swarm optimization scaling factor (…). ) Scaling is performed; for particles selected into the high-quality particle subset, the particle with the worst fitness value in the high-quality particle subset is replaced with the perturbation replacement solution with boundary correction in step 33).

[0146] The position and velocity updates of the particles are defined by the following two formulas:

[0147]

[0148] in, Select a label for the agent when the particle When selected into a high-quality subset of particles, =1; otherwise =0;

[0149] Replace the label for the worst particle when the particle When an individual is determined to be the least fit in the current population and is replaced, =1, otherwise =0;

[0150] Optimize scaling factor for adaptive particle swarm optimization;

[0151] and They represent particles respectively The velocity vectors at the current and next moments. The inertia coefficient, For particles The best historical position This is the historically optimal position for the entire population. and These are the individual cognitive coefficient and the social learning coefficient, respectively. and Random disturbance coefficient.

[0152] Select tabs This determines whether to update particles using a surrogate-assisted update strategy or according to the Adaptive Particle Swarm Optimization (APSO) rule. For particles predicted by the surrogate model to have high optimization potential, a value is assigned... Furthermore, a center perturbation strategy is incorporated for refined searching; the remaining particles complete the iteration according to the APSO update process. This selective update mechanism can prioritize the allocation of limited computing resources to more promising regions in the search space while maintaining population diversity, thereby improving the overall search efficiency of the algorithm. At the same time, the APSO scaling factor, as a nonlinear adjustment operator, replaces the fixed velocity weight in the traditional algorithm, and is used to adjust the particle's step size in real time according to the convergence slope and fitness variance of the current population.

[0153] Step 4): Calculate the actual fitness value of the new particles generated after the update in Step 34), and store these new particles and their actual fitness values ​​as data pairs in the external database for use in subsequent iterations to update the random forest proxy model.

[0154] Step 5): Repeat steps S3 to S4 until the preset termination condition is met, and output the optimal combination of decision variables that maximizes the objective function value.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for power optimization of heterogeneous wind farms based on surrogate models and LLM-assisted mechanisms, characterized in that: include: Step 1): Using the axial induction factor and yaw angle of each wind turbine in the heterogeneous wind farm as decision variables, establish an optimization problem with maximizing the total output power of the wind farm as the objective function. The total output power of the wind farm is calculated based on a non-uniform three-dimensional wake model according to the following steps: 11) Wake Velocity Field Calculation: For any point within the wind farm, the directional wind speed affected by the wake of the upstream turbines is calculated using the following formula: in, These are the coordinates of the point in the downstream direction, horizontal direction, and vertical direction; To account for the initial wind speed caused by wind shear, the calculation formula is as follows: in For free-flowing wind speed; The wind shear index, This refers to the height of the wind turbine hub. This is the lateral offset of the wake centerline relative to the rotor center; and These represent the characteristic widths of the wake in the lateral and vertical directions, respectively, and satisfy a linear expansion relationship: in, The diameter of the upstream wind turbine rotor. The starting position of the far wake. and These represent the lateral and vertical wake expansion rates, respectively. and The initial wake width at the far wake initiation position; The velocity deficit coefficient at the wake center. Initial thrust coefficient: in, It is the axial induction factor; 12) Calculation of multiple wake superposition: For downstream points located in the wake superposition area of ​​several upstream units, the total velocity loss is superimposed based on the velocity loss caused by each upstream unit individually, according to the principle of conservation of average kinetic energy loss. 13) Calculation of equivalent wind speed for a single unit: For the first... The Taiwanese generator set discretizes its rotor plane into... For each measuring point, calculate the effective incoming air velocity based on steps 11) and 12). The rotor's average effective incoming air velocity is calculated using the following formula. : 14) Calculation of total wind farm power: based on the equivalent wind speed of all units. Their respective power curves and yaw angle Calculate the total output power of the wind farm. : in This represents the total number of wind turbines in a heterogeneous wind farm. The yaw power loss index; Step 2): Initialize a population and an external database; wherein each particle in the population represents a set of decision variables for all wind turbine generators; the external database is used to store the particles and their corresponding historical fitness values, wherein the fitness values ​​are the total output power of the wind farm calculated in Step 1); Step 3): Based on the initialized particle swarm and external database, execute an iterative optimization loop, performing the following steps in each iteration: 31) Dynamic hyperparameter decision-making: The current state feature vector of the population is collected and input into the controller assisted by the large language model. The state feature vector includes at least population diversity, convergence slope and surrogate model prediction error. The controller infers and outputs a set of dynamic hyperparameters in real time through nonlinear mapping based on the state feature vector, including perturbation factor, surrogate selection ratio and adaptive particle swarm optimization scaling factor. 32) Proxy-assisted selection: Train a random forest proxy model based on the external database, and use the random forest proxy model to predict the fitness value of each particle in the current population; sort the predicted fitness values ​​of all particles in the current population in descending order, and calculate the number of high-quality particles to be selected in this iteration according to the proxy selection ratio to form a subset of high-quality particles; 33) Calculate the perturbation and replacement solution of the center: Calculate the geometric center of the high-quality particle subset in the decision space, and generate a perturbation replacement solution after boundary correction based on the perturbation scaling factor and the standard deviation of the current population in each decision variable dimension; 34) Adaptive Particle Swarm Optimization Update: For particles not selected into the high-quality particle subset, their position and velocity are updated according to the standard particle swarm optimization algorithm rules, wherein the velocity update term is scaled by the adaptive particle swarm optimization scaling factor; for particles selected into the high-quality particle subset, the particle with the worst fitness value in the high-quality particle subset is replaced with the perturbation replacement solution with boundary correction in step 33). Step 4): Calculate the actual fitness value of the new particles generated after the update in Step 34), and store these new particles and their actual fitness values ​​as data pairs in the external database for use in subsequent iterations to update the random forest proxy model; Step 5): Repeat steps S3 to S4 until the preset termination condition is met, and output the optimal combination of decision variables that maximizes the objective function value.

2. The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism according to claim 1, characterized in that: The optimization problem described in step 1), which aims to maximize the total output power of the wind farm, is mathematically defined as follows: in, It is an axial inducing factor. Yaw angle Let P0 be the objective function, and P0 be the wind farm reference power corresponding to each wind turbine under independent optimal operating conditions. For a given free-flowing wind speed and wind direction Below, corresponding to the decision variables The total output power of the wind farm is calculated at that time.

3. The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism according to claim 2, characterized in that: In step 31), the optimized state feature is a five-dimensional feature vector: in For population diversity; The slope of convergence; For fitness variance; The prediction error of the random forest proxy model is... This refers to the iteration progress.

4. The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism according to claim 3, characterized in that: In step 31), the large language model-assisted hyperparameter decision controller dynamically outputs the dynamic hyperparameters using the following formula: in, This is the perturbation scaling factor. Choose the ratio for the agent. To optimize the scaling factor for adaptive particle swarm optimization, , and For the preset constant boundary value, , , The coefficients are generated by the large language model based on the normalized feature vectors, with subscripts i=1,2,3.

5. The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism according to claim 4, characterized in that: In step S32, the random forest surrogate model is trained on samples selected from the historical database, and its prediction result is the average of the outputs of all decision trees; the number of particles in the high-quality particle subset is obtained by the following formula: Where t is the iteration number index. To dynamically adjust the number of particles selected, where N is the population size, Select a ratio for the agent.

6. The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism according to claim 5, characterized in that: Step 32) The fitness value of the particle is obtained by averaging the prediction results of all decision trees by the random forest surrogate model: in For the number of decision trees, For the first The combination of decision variables represented by each particle; The number of high-quality particles in step 32) is calculated as follows: in: Where N is the number of high-quality particles in the current iteration step, and N is the population size. Choose the ratio for the agent. This is a rounding function.

7. The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism according to claim 6, characterized in that: In step 33), the geometric center of the high-quality particle subset in the decision space It is generated by the following formula: Where t is the iteration number index. Describes the set of indices of high-quality particles selected in the t-th iteration. For the first high-quality particle subset The position of each particle; The perturbation-adjusted replacement solution is generated using the following formula: in, This is the perturbation scaling factor. Let be a d-dimensional random perturbation vector, where d is the total number of decision variables, and each of its components is... The generation follows an adaptive mechanism: Where N represents a Gaussian distribution with a mean of 0 and a variance of 0. Let be the sample standard deviation of the entire population in the j-th decision variable dimension; Replacement solution after perturbation Boundary corrections are performed to obtain the final perturbation replacement solution. : in, and These are the preset lower and upper limits of the j-th decision variable, respectively.

8. The heterogeneous wind farm power optimization method based on surrogate model and LLM-assisted mechanism according to any one of claims 1-7, characterized in that: The position and velocity updates of the particles are defined by the following two formulas: in, Select a label for the agent when the particle When selected into a high-quality subset of particles, =1; otherwise =0; Replace the label for the worst particle when the particle When an individual is determined to be the least fit in the current population and is replaced, =1, otherwise =0; Optimize scaling factor for adaptive particle swarm optimization; and They represent particles respectively The velocity vectors at the current and next moments. The inertia coefficient, For particles The best historical position This is the historically optimal position for the entire population. and These are the individual cognitive coefficient and the social learning coefficient, respectively. and Random disturbance coefficient.

Images