A wind farm cluster multi-objective coordinated dispatching method considering predicted power correction

Abstract

The application relates to the technical field of offshore wind power dispatching, and discloses a wind farm cluster multi-target coordinated dispatching method considering predicted power correction, which comprises the following steps: a physical baseline model is constructed, and XGBoost is used for residual compensation to obtain conventional state predicted power; for extreme weather such as typhoon, a typhoon wind field model and unit cut-out protection logic are combined to correct the predicted boundary; a dispatching model containing four targets of system operation safety, unit life protection, power generation fairness and transmission channel utilization rate is constructed; after the four targets are normalized, a double-target model of an internal operation quality comprehensive target and a transmission channel utilization rate target is reconstructed; an adaptive penalty function NSGA-II is used for solving, and a compromise dispatching scheme is output. The application realizes consistency maintenance of the predicted boundary and actual available capacity of units, and adaptive switching of optimization targets with dominant risks.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power dispatching technology, and in particular to a multi-objective collaborative dispatching method for wind farm clusters that considers predicted power correction. Background Technology

[0002] Offshore wind power development is gradually shifting from independent development of single wind farms to large-scale, clustered, and collaborative operation. Wind power clusters are characterized by wide spatial coverage, long power transmission distances, and large total installed capacity. Their active power output characteristics are affected by both the random fluctuations in wind speed at wind farms, the spatial heterogeneity of wind energy resources at different sites, and the inherent errors in numerical weather forecasts, as well as by strong disturbances caused by extreme and unconventional conditions such as typhoons, sudden changes in wind speeds, and failures in power transmission channels.

[0003] Under normal operating conditions, the dispatch and control center determines the feasible range of wind power output based on numerical weather forecasts, historical measured operating data, and unit power characteristic curves, and completes the coordinated allocation of active power among various wind farms within this constrained range. However, sudden scenarios such as typhoon attacks, sudden drops in output power, safe shutdown of units, and line fault isolation can cause the feasible boundary of system power generation to shrink sharply or change abruptly in a short period of time, resulting in the incompatibility of traditional power output prediction methods and supporting dispatch strategies.

[0004] Furthermore, there are significant multi-objective game-like checks and balances among the four optimization objectives: system safety and stable operation, wind turbine lifecycle loss management, balance of power generation rights among various power plants, and capacity utilization efficiency of power transmission channels. Different scheduling priorities need to be formulated for different extreme scenarios: under conditions of severe power fluctuations, the core objective is to smooth out rapid power output oscillations; under typhoon warning conditions, long-term full-capacity operation of large-capacity power plants should be avoided; under conditions of power transmission channel failure, feasible power allocation schemes under fault constraints should be prioritized.

[0005] Existing wind power output prediction technologies mainly include physical methods based on ideal power curves and machine learning methods based on historical data. Physical methods can characterize basic mechanisms such as cut-in, rated operation, and cut-out, but they are difficult to fully consider wake, electrical losses, unit control, and local meteorological deviations. Although purely data-driven methods can fit historical patterns, they are prone to losing physical consistency under conditions of few samples, such as typhoons, strong winds, protection shutdowns, and recovery.

[0006] Existing active power allocation strategies for offshore wind power clusters are mostly geared towards conventional operating conditions, primarily focusing on a single objective such as dispatch command tracking, total power smoothing, or maximizing power generation, or using fixed weights to combine multiple objectives. These methods fail to uniformly map numerical weather forecast biases, typhoon-induced cut-off protection, and sudden drops in transmission capacity to the current feasible domain, easily leading to dispatch ceilings exceeding actual generating capacity.

[0007] Furthermore, directly solving the four original objectives of system operation safety, unit life protection, power generation fairness, and transmission channel utilization together increases the complexity of Pareto front search, solution set selection, and engineering decision-making. Simply weighting the four objectives into a single objective masks the contradiction between internal operational quality and external transmission capacity. Therefore, existing technologies lack a cluster scheduling method that integrates prediction boundary correction, weight setting under known operating conditions, dimensionality reduction reconstruction from four objectives to two objectives, and constraint optimization solutions. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-objective collaborative scheduling method for wind farm clusters that considers predicted power correction.

[0009] The objective of this invention is achieved through the following technical solution: a multi-objective cooperative scheduling method for wind farm clusters that considers predicted power correction, the method comprising the following steps:

[0010] S1. Construct a wind farm output prediction model under normal conditions, and obtain the predicted wind farm output under normal conditions based on the wind farm output prediction model.

[0011] S2. Based on the wind farm predicted output obtained from S1, the wind farm intensity distortion caused by extreme weather and the impact of unit cut-out are corrected to obtain the corrected prediction boundary.

[0012] S3. Construct a multi-objective optimization scheduling model for offshore wind power clusters, including a system operation safety objective function aimed at minimizing power fluctuations, a unit life protection objective function aimed at minimizing the improved load rate of the units, a power generation fairness objective function aimed at minimizing the fatigue distribution coefficient, and a transmission channel utilization objective function aimed at minimizing the curtailment rate of the transmission channels.

[0013] S4. Based on the multi-objective optimization scheduling model of offshore wind power clusters, an improved bi-objective optimization model for solving NSGA-II is constructed. The system operation safety objective function, unit life protection objective function, power generation fairness objective function and transmission channel utilization objective function established in S3 are normalized respectively. The normalized system operation safety objective value, normalized unit life protection objective value and normalized power generation fairness objective value are weighted and merged according to preset weights to construct the first optimization objective with the goal of minimizing the comprehensive internal operation quality value. The normalized transmission channel utilization objective value is constructed as the second optimization objective with the goal of minimizing the transmission channel abandonment rate.

[0014] S5. Solve the improved bi-objective optimization model based on NSGA-II, taking the first optimization objective and the second optimization objective as the two optimization objectives of NSGA-II, and using an adaptive penalty function to handle constraint violations, search for cluster power allocation schemes within the feasible region formed by physical constraints, and output the cluster collaborative optimization scheduling results.

[0015] Specifically, S1 includes:

[0016] S11. Acquire and process wind farm operation data, meteorological data, and equipment parameters, and establish a physical baseline model, represented as follows:

[0017] ;

[0018] ;

[0019] ;

[0020] In the formula, It contributes to the physical model of the wind farm; For single-machine physical output; This refers to the number of wind turbines. This is the overall efficiency coefficient; This refers to the rated power of a single unit. The equivalent wind speed at the hub height; , and These are the cut-in wind speed, rated wind speed, and cut-out wind speed, respectively. Wind shear index; For height Wind speed at the location; Wheel hub height;

[0021] S12. Construct meteorological physical characteristics and time periodic characteristics, establish an XGBoost data-driven compensation model to compensate for the residuals of the physical baseline model, and obtain the predicted power under normal conditions.

[0022] Specifically, the XGBoost data-driven compensation model is as follows:

[0023] ;

[0024] In the formula, For the number of regression trees, Let r be the r-th regression tree; The predicted power of object k at sample time t or date d at time t; The input feature vector;

[0025] The input to the XGBoost data-driven compensation model is:

[0026] ;

[0027] In the formula, The input feature vector; It is a set of meteorological and physical characteristics; It is a set of time and periodic features; It contributes to the physical model of the wind farm;

[0028] The XGBoost data-driven compensation model is trained using loss function terms and model complexity regularization terms:

[0029] ;

[0030] In the formula, To train the objective function for the overall model built for the prediction object k; The prediction error loss function; This is the set of training samples corresponding to the predicted object; The actual power of sample d at time t; Let be the complexity regularization term for the r-th regression tree;

[0031] In the current iteration, we introduce the first and second derivatives of the loss function with respect to the predicted values:

[0032] ;

[0033] In the formula, Let be the first derivative of the loss function with respect to the predicted value of the i-th sample; Let be the second derivative of the predicted value of the i-th sample; This represents the actual power of the i-th sample. t′−1 is the predicted value obtained in the model iteration round; t′ is the model iteration round.

[0034] Specifically, S2 includes:

[0035] S21. Establish a typhoon wind field model based on the typhoon path and structural parameters, obtain the latitude and longitude of the typhoon center, the maximum wind speed of the typhoon center and the maximum wind speed radius, calculate the distance between the wind farm and the typhoon center, use the Rankine composite vortex model to approximate the radial wind speed of the typhoon, and obtain the effective wind speed of the typhoon at the location of the wind farm by combining the environmental wind field and the typhoon rotating wind field.

[0036] ;

[0037] In the formula, The effective typhoon wind speed at the location of the wind farm; This represents the distance between the wind farm and the center of the typhoon. This is a Rankine complex vortex model;

[0038] S22. Based on the conventional predicted power of S1, and combined with the typhoon's impact range and the unit's disconnection protection logic, a revised prediction boundary is formed, specifically as follows:

[0039] ;

[0040] In the formula, This represents the revised predicted power boundary under the influence of the typhoon. This is the baseline predicted power obtained from numerical weather prediction; For numerical weather forecast wind speed; Cut off the wind speed for the unit; and This is a correction factor.

[0041] Specifically, the objective function for system operation safety is:

[0042] ;

[0043] ;

[0044] In the formula, The objective function for ensuring system operation safety; This represents the total number of wind turbines in the wind farm. Power fluctuation coefficient; , Based on the closest to the current time The average and standard deviation of the unit power calculated over a historical period; To prevent the introduction of positive numbers with a denominator of zero; The length of the scheduling time domain;

[0045] The objective function for unit life protection is:

[0046] ;

[0047] ;

[0048] In the formula, The objective function for protecting the unit's lifespan; For the first Taiwanese crew The instantaneous fatigue coefficient of the machine at a given moment; For the first Taiwanese crew The actual active power at any given moment; For the first Rated installed capacity of the unit;

[0049] The objective function for power generation fairness is:

[0050] ;

[0051] ;

[0052] In the formula, The objective function is the fairness of power generation. for Fatigue distribution coefficient of wind farm at any given time;

[0053] The objective function for the utilization rate of the outgoing channel is:

[0054] ;

[0055] ;

[0056] In the formula, The objective function is the output channel utilization rate; The abandonment rate of the unit's power supply channel; This represents the maximum active power that the unit can theoretically generate under actual wind conditions at the current moment.

[0057] Specifically, the constraints of the multi-objective optimization scheduling model for offshore wind power clusters include:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] In the formula, The maximum permissible climbing rate; This is the minimum grid-connected power of the unit; This is the short-time overload factor; This represents the maximum transmission capacity limit of the outgoing channel; The upper limit of the generateable power at time t is determined by the prediction results; Let t be the reference value for the total power of the wind farm cluster at time t.

[0064] Specifically, the normalization in S4 is performed in the following manner:

[0065] ;

[0066] In the formula, For the first One objective function; To normalize the lower bound; The normalized upper bound; For the j-th original target f j The normalized target value;

[0067] The first optimization objective and the second optimization objective are expressed as follows:

[0068] ;

[0069] In the formula, The primary optimization objective is to achieve this. The second optimization objective; , and These correspond to weighting coefficients for system operational safety, unit lifespan, and wind farm power generation fairness, respectively, and satisfy the following conditions: .

[0070] Specifically, S5 includes:

[0071] S51. Encode the hourly output allocation ratio with real numbers and decode the chromosome into the actual active power of each scheduling unit.

[0072] S52. The improved bi-objective optimization model is solved iteratively using the constraint-controlled NSGA-II;

[0073] S53. Select a compromise scheduling scheme from the set of feasible Pareto optimal solutions.

[0074] Specifically, the chromosome is decoded into the actual active power of each scheduling unit as follows:

[0075] ;

[0076] In the formula, This refers to the actual active power. The actual reference power allocated to the cluster level in the t-th time period; For the first The scheduling unit in the first... Output allocation code for each time period; For the first The set of operable scheduling units for each time period; To prevent encoding tolerance for division by zero.

[0077] Specifically, S52 includes:

[0078] S521. Initialize the parent population and calculate the biobjective value and constraint violation degree:

[0079] Write the inequality constraint as Write the equality constraints as and respectively according to the allowable scale and Dimensionless transformation:

[0080] ;

[0081] In the formula, , q and r are the normalized violation degrees of inequality and equality constraints, respectively; q and r are the constraint indices. , which is the q-th inequality constraint; , is the r-th equality constraint; For candidate scheduling solutions; , These are the allowable scale and reference scale used for dimensionless transformation of the corresponding constraint violation degree, respectively;

[0082] Based on the current known operating condition s, the weights of various constraints are set to form the adaptive total constraint violation degree:

[0083] ;

[0084] In the formula, The degree of total constraint violation under adaptive operating conditions; and These are the violation weights of inequality constraints and equality constraints under operating condition s, respectively. , These represent the number of inequality constraints and the number of equality constraints, respectively. Current running status / scenario;

[0085] S522. Compare candidate solutions based on constraint dominance relationships:

[0086] ;

[0087] In the formula, and Candidate solution; For candidate solutions The objective value on the m-th optimization objective; For candidate solutions The objective value on the m-th optimization objective;

[0088] S523, Perform quick non-dominated sort:

[0089] ;

[0090] In the formula, Let be the constraint domination relationship; be the set of individuals constrained and dominated by individual p; be the number of individuals dominating individual p; p and q are the population individual indices; be the first non-dominated front.

[0091] S524. Calculate the crowding distance for individuals within each non-dominated front:

[0092] ;

[0093] In the formula, For the target scale; To achieve target normalization tolerance; , These are the maximum and minimum values ​​of the m-th objective in the current non-dominated frontier, respectively; , These are the target values ​​of adjacent individuals after sorting by the m-th target; Let i be the crowding distance for individual i;

[0094] S525, Selecting the parent generation through a binary tournament:

[0095] ;

[0096] In the formula, , These are two candidate individuals participating in the binary tournament; , It is a non-dominant class; , These are their respective crowding distances; To select priority relationships.

[0097] S526. Use arithmetic crossover and Gaussian mutation to generate the offspring population:

[0098] With crossover probability Perform crossover on individuals in the current parent population; for the two selected parent individuals, on the... Arithmetic crossover is used to generate two crossover offspring for each variable:

[0099] ;

[0100] In the formula, The arithmetic cross coefficient follows the following rules: (0,1) Uniformly distributed; , Let j be the value of the j-th variable of the two parent individuals; j is the index of the decision variable. Let be the values ​​of the two offspring generated after arithmetic crossover on the j-th decision variable, i.e. and , (0) and (1) are used to distinguish two cross-off offspring, and do not represent powers or numerical ranges;

[0101] With the probability of mutation Select the individuals to be mutated, and categorize them according to the proportion of variable variation in a single mutated individual. Select the variation location and use the standard deviation as . Gaussian perturbation generates mutated offspring:

[0102] ;

[0103] In the formula, For the mutated first One variable value; The first before the mutation One variable value; The standard deviation of the Gaussian perturbation; Let be a random variable that follows a standard normal distribution;

[0104] Perform truncation repair on components that exceed the upper and lower bounds of the variable after crossover or mutation:

[0105] ;

[0106] In the formula, , The first The lower and upper bounds of each decision variable.

[0107] S527. Merge parent and offspring populations and implement elite preservation:

[0108] ;

[0109] In the formula, This is the parent population of the kth generation; , These represent the crossover offspring and the mutated offspring generated in the k-th generation, respectively. Population size; For the next generation of population; To truncate and retain N according to non-dominance level ascending order and congestion distance descending order pop Individual.

[0110] The present invention has the following advantages:

[0111] 1. This invention integrates vertical extrapolation of wind speed, segmented wind turbine power characteristic curves, and wind farm cluster output aggregation to construct a physical benchmark prediction model. It also introduces an XGBoost model driven by meteorological and temporal characteristics to carry out prediction residual compensation, thereby achieving high-precision prediction of conventional wind power output and solving the technical defect that a single physical model is difficult to capture the prediction deviation caused by nonlinear meteorological disturbances.

[0112] 2. Based on conventional wind power prediction, this invention combines the typhoon path, Rankine composite vortex wind field model and wind turbine cut-out protection action logic to dynamically correct the upper and lower boundaries of wind power generation capacity, thereby achieving accurate constraint on wind power output boundaries under extreme typhoon conditions and overcoming the problem of distortion of output boundaries caused by conventional prediction methods not considering extreme strong wind protection actions.

[0113] 3. This invention integrates the issuance of dispatch instructions, the predicted boundary of wind power output after extreme weather correction, the available installed capacity of wind farms and the transmission limit of transmission channels, and constructs a cluster wind power reference allocation power calculation model and a multi-objective scheduling feasible domain solution method. It realizes a complete characterization of the scheduling feasible space under multiple types of operational boundary constraints, and solves the shortcomings of existing allocation strategies that only consider capacity constraints and ignore multi-dimensional boundary coupling restrictions.

[0114] 4. This invention dynamically configures target weights based on meteorological forecast information, real-time equipment operating status, protection actions, equipment fault signals, and preset scheduling operating parameters. It uses weighted fusion to form an internal comprehensive operating quality target by normalizing the results of system operation safety targets, unit fatigue life protection targets, and power generation fairness targets. It also uses the normalized index of transmission channel utilization rate as an independent optimization target to establish an improved dual-objective optimization model. This achieves hierarchical decoupling optimization of multi-dimensional operation demands and avoids the problems of unclear target priority and unbalanced optimization guidance caused by direct weighting of multiple targets.

[0115] 5. This invention improves the dual-objective optimization model by coupling reference power tracking constraints, unit ramp rate constraints, single unit output upper and lower limit constraints, meteorological correction prediction boundary constraints, station available capacity constraints, and transmission channel capacity constraints. It uses an adaptive penalty function combined with the NSGA-II multi-objective evolutionary algorithm to solve the model and outputs a Pareto optimal power allocation solution set that satisfies all-dimensional operational constraints. This solves the technical problem of convergence difficulty and lack of feasible solutions in traditional fixed penalty functions under multiple strong constraint scenarios.

[0116] 6. This invention designs a complete technical process for wind power cluster scheduling with full-cycle collaborative linkage. Within a single scheduling cycle, it completes the closed-loop connection of all stages, including conventional output prediction, output boundary correction for extreme / transitional weather, construction of multiple original optimization objectives, adaptive weight assignment for operating conditions, construction of a dual-objective optimization model adapted to NSGA-II solution, evolutionary algorithm model solution, and output of optimal power allocation results. This achieves integrated collaborative scheduling of prediction, modeling, solution, and allocation, breaking the limitations of existing technologies where each functional module operates independently and the timing is fragmented, resulting in insufficient scheduling timeliness. Attached Figure Description

[0117] Figure 1 This is a schematic diagram of the scheduling method of the present invention;

[0118] Figure 2 Comparison chart of predicted and corrected power output of the station during three typhoons;

[0119] Figure 3 The diagram shows the optimized scheduling results under high fluctuation conditions.

[0120] Figure 4 This is a diagram showing the optimized scheduling results under typhoon risk conditions.

[0121] Figure 5 The diagram shows the optimized scheduling results under the condition of outgoing channel failure.

[0122] Figure 6 This is a flowchart of the solution process for the multi-objective optimization scheduling model of the present invention. Detailed Implementation

[0123] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0124] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0125] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0126] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0127] like Figures 1 to 6 As shown, a multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction is characterized by the following steps:

[0128] S1. Construct a wind farm output prediction model under normal conditions, and obtain the predicted wind farm output under normal conditions based on the model. This step is used to obtain the predicted wind farm output under normal conditions, providing a benchmark predicted power for unconventional weather correction and cluster optimization scheduling.

[0129] S1 specifically includes:

[0130] S11. Acquire and process wind farm operation data, meteorological data, and equipment parameters to establish a physical baseline model. Offshore wind farms typically consist of multiple turbines. Due to differences in spatial location, inflow conditions, and wind farm structure, the effective inflow wind speed of different turbines is not entirely consistent. Furthermore, the relationship between wind speed and turbine active power output exhibits a typical nonlinear relationship. To characterize the basic conversion process from meteorological conditions to turbine power output, this embodiment first constructs a physical baseline model based on vertical extrapolation of wind speed and turbine power curves, setting a reference height. The wind speed at that location is The equivalent wind speed at the hub height is:

[0131] ;

[0132] Once the hub height wind speed is determined, the piecewise power curve can be used to describe the power output variation of the wind turbine near the cut-in, rated, and cut-out wind speeds. The physical output of a single unit can be expressed as:

[0133] ;

[0134] Based on this, the physical output of each unit is aggregated to the scale of the wind farm, and a comprehensive efficiency coefficient is introduced to correct for internal losses at the site, thus obtaining the output of the wind farm physical model:

[0135] ;

[0136] In the formula, It contributes to the physical model of the wind farm; For single-machine physical output; This refers to the number of wind turbines. This is the overall efficiency coefficient; This refers to the rated power of a single unit. The equivalent wind speed at the hub height; , and These are the cut-in wind speed, rated wind speed, and cut-out wind speed, respectively. Wind shear index; For height Wind speed at the location; This refers to the wheel hub height.

[0137] The physical output of the wind farm is spliced ​​together with wind speed, wind direction (sine and cosine), air pressure, temperature, humidity, wind power density, and time period characteristics. Gradient boosting trees are used to compensate for the residuals of the physical model to obtain the predicted power under normal conditions.

[0138] S12. Construct meteorological physical characteristics and time periodic characteristics, establish an XGBoost data-driven compensation model to compensate for the residuals of the physical baseline model, and obtain the predicted power under normal conditions.

[0139] Vertical wind shear reflects the characteristics of wind speed distribution at different heights and has a significant impact on the effective wind energy within the swept area of ​​large-capacity wind turbine rotors. This embodiment defines the wind shear index in logarithmic form:

[0140] ;

[0141] In the formula, and Wind speeds at heights of 100m and 10m, respectively;

[0142] To avoid the interference of sudden changes in wind direction values ​​on machine learning models, this invention uses sine and cosine forms to encode wind direction;

[0143] For any height wind direction :

[0144] ;

[0145] Air density directly affects wind energy input intensity. This embodiment calculates air density based on air pressure and temperature:

[0146] ;

[0147] In the formula, air density; The gas constant for dry air. and These are air pressure and absolute temperature, respectively.

[0148] Wind power density:

[0149] ;

[0150] In the formula, Wind power density describes the energy intensity contained in the incoming flow and has a more direct physical correlation with wind power output. Relative humidity is introduced. As a supplement to meteorological features, the final set of meteorological features is obtained. :

[0151] ;

[0152] In the formula, , Wind speeds at heights of 100 m and 10 m, respectively; , The wind directions at heights of 100 m and 10 m are respectively. Relative humidity; Wind shear index; air density; Wind power density; , These are the sine and cosine coding terms for wind direction at a height of 100 m, respectively. , These are the sine and cosine coding terms for wind direction at a height of 10 m, respectively.

[0153] Furthermore, wind power output typically exhibits significant periodic variations on both intraday and intrayear timescales, influenced by factors such as sea and land breezes, boundary layer evolution, and seasonal climate background. To characterize these periodic patterns, this embodiment constructs time features including hours, weeks, and months to describe the statistical distribution characteristics of wind power output.

[0154] To avoid abrupt changes in discrete-time variables, a sine and cosine form is used to periodically encode time. Taking hourly features as an example:

[0155] ;

[0156] In the formula, Let t be the hour number of the day; , These are the sine and cosine codes for the hourly cycle, respectively.

[0157] Taking the characteristics of day sequence within the year as an example:

[0158] ;

[0159] In the formula, Let t be the day number of time t in the year; 1 to 365 for common years and 1 to 366 for leap years. , These are the sine and cosine codes for the annual cycle, respectively.

[0160] The yearly date sequence in the formula is the sequential number of the current date within the year, with common years and leap years normalized according to the corresponding number of days in the year.

[0161] Furthermore, by incorporating weekday / month / seasonal indicator variables, the model's ability to identify differences in wind field background across different seasons is enhanced.

[0162] ;

[0163] In the formula, For working day indicator variables; For month indicator variables; This is an indicator variable for winter. This is an indicator variable for summer.

[0164] After constructing the meteorological and physical characteristics and temporal characteristics, the output of the wind farm physical model is concatenated with the above characteristics to form the input feature vector of the machine learning model:

[0165] ;

[0166] In the formula, The input feature vector; It is a set of meteorological and physical characteristics; It is a set of time and periodic features; It contributes to the physical model of the wind farm.

[0167] The XGBoost data-driven compensation model consists of an additive combination of multiple regression trees, and its output is expressed as follows:

[0168] ;

[0169] In the formula, This represents the number of regression trees; Let r be the r-th regression tree; The predicted power of object k at sample time t or date d at time t; The input feature vector;

[0170] The XGBoost data-driven compensation model is trained using loss function terms and model complexity regularization terms:

[0171] ;

[0172] In the formula, To train the objective function for the overall model built for the prediction object k; This is the set of training samples corresponding to the predicted object; The actual power of sample d at time t; Let be the complexity regularization term for the r-th regression tree; The preferred loss function for prediction error is the squared error, expressed as:

[0173] ;

[0174] In the formula, This is the squared error loss function; the smaller the value, the closer the predicted value is to the actual value. This refers to the actual power. Predict power for the model;

[0175] The model complexity regularization term is used to suppress overfitting.

[0176] In the current iteration, we introduce the first and second derivatives of the loss function with respect to the predicted values:

[0177] ;

[0178] In the formula, Let be the first derivative of the loss function with respect to the predicted value of the i-th sample; Let be the second derivative of the predicted value of the i-th sample; This represents the actual power of the i-th sample. t′−1 is the predicted value obtained in the model iteration round; t′ is the model iteration round.

[0179] The first derivative represents the gradient of the loss function with respect to the current predicted value, and the second derivative represents the corresponding curvature. Both are used together to solve for the new regression tree.

[0180] The newly added regression tree is solved using a second-order approximation method, achieving a balance between fitting accuracy and generalization ability. The trained XGBoost model is applied to the physical model residuals (i.e., the deviation between the actual power output and the physical model output), and the residual prediction value is output. The residual prediction value is superimposed with the physical model output to obtain the predicted power of the wind farm under normal conditions, which is used as the input of S2.

[0181] S2. Based on the wind farm predicted output obtained in S1, correct for wind farm intensity distortion caused by extreme weather and the impact of turbine cut-out, to obtain the corrected prediction boundary. This step uses the conventional predicted power obtained in S1 as a basis to correct for wind farm intensity distortion caused by extreme or transitional weather such as typhoons and the impact of turbine cut-out, to obtain the corrected prediction boundary. This corrected prediction boundary will serve as a key input in S3 to determine the actual cluster-level allocation reference power and establish prediction boundary constraints. In other words, S2 superimposes the physical constraints of extreme weather on the conventional prediction of S1, ensuring that the power boundary used for subsequent scheduling is consistent with the actual available generating capacity. Specifically, it includes:

[0182] S21. Establish a typhoon wind field model based on the typhoon's path and structural parameters. For strong wind-type extreme weather such as typhoons, it is necessary to further characterize the wind-receiving characteristics of the wind farm. First, obtain the latitude and longitude of the typhoon center, the maximum wind speed at the typhoon center, and the maximum wind speed radius. Then, use the spherical distance formula to calculate the distance between the wind farm and the typhoon center.

[0183] ;

[0184] In the formula, This represents the distance between the wind farm and the center of the typhoon. The coordinates of the typhoon's center are latitude and longitude. The coordinates of the center of the wind farm are latitude and longitude. The average radius of the Earth;

[0185] The radial wind speed of the typhoon is approximated using the Rankine complex vortex model:

[0186] ;

[0187] In the formula, This is the maximum wind speed at the center of the typhoon; The radius of maximum wind speed; Radial distance; It is an exponent that indicates the decrease in wind speed with distance. The tangential wind speed is located at a radial distance r from the typhoon center.

[0188] Substituting the aforementioned typhoon structural parameters and spherical distance into the composite vortex model to synthesize the environmental wind field and the typhoon rotating wind field, the effective typhoon wind speed at the wind farm location is obtained.

[0189] ;

[0190] In the formula, The effective typhoon wind speed at the location of the wind farm; This represents the distance between the wind farm and the center of the typhoon. This is the Rankine complex vortex model.

[0191] S22. Based on the conventional predicted power of S1, a modified prediction boundary is formed by combining the typhoon's impact range and the unit's cut-out protection logic.

[0192] When the distance between a wind farm and the center of a typhoon meets the following conditions, it is determined that the period of typhoon influence has begun:

[0193] ;

[0194] In the formula, Threshold for determining the impact range of a typhoon; when At that time, it was assumed that the wind farm was within the typhoon's influence range and a predicted power boundary correction was performed.

[0195] Dynamically correct conventional predicted power based on the unit segmented power curve and cut-off protection logic:

[0196] ;

[0197] In the formula, This represents the revised predicted power boundary under the influence of the typhoon. This is the baseline predicted power obtained from numerical weather prediction; For numerical weather forecast wind speed; Cut off the wind speed for the unit; and As a correction factor, when When using correction factors ,when When using correction factors .

[0198] This revised model incorporates the typhoon's influence radius, rotational wind speed, and turbine cut-out protection mechanism into the power output prediction process, providing physical constraints for correcting the prediction boundaries during the typhoon's influence period.

[0199] The corrected prediction boundary can characterize the strong wind load reduction, shutdown, and recovery process after the typhoon passes. After the typhoon's impact conditions are lifted, the prediction link is restored to the conventional model after deviation correction, and the corrected prediction boundary is passed to S3 as the basis for prediction boundary constraints and reference power calculation in the scheduling model.

[0200] S3. Construct a multi-objective optimization scheduling model for offshore wind power clusters, including a system operation safety objective function aimed at minimizing power fluctuations, a unit life protection objective function aimed at minimizing unit improved load rates, a power generation fairness objective function aimed at minimizing fatigue distribution coefficients, and a transmission channel utilization objective function aimed at minimizing transmission channel curtailment rates. This step is used to determine the reference power, objective functions, and feasible power allocation range for optimal scheduling. This step receives the corrected prediction boundary from S2 and, combined with upper-level scheduling instructions and available capacity at the wind farms, jointly determines the actual cluster-level allocation reference power; simultaneously, it establishes four objective functions and six types of constraints. The output of this step is the objective functions and constraints, which will be passed to S4 for objective normalization and bi-objective reconstruction, and the reference power will be passed to S5 for chromosome decoding. Specifically, it includes:

[0201] S31, combined with scheduling instructions, corrected prediction boundaries and available station capacity, to determine the actual allocation reference power at the cluster level;

[0202] To ensure that the scheduling target is within the current feasible domain, the actual allocation reference power at the cluster level is taken as the minimum value among the upper-level scheduling instruction, the corrected prediction boundary, and the available capacity of the site:

[0203] ;

[0204] In the formula, The actual reference power is allocated at the cluster level; , , These are the upper-level scheduling instruction boundary, the corrected cluster total prediction boundary, and the current total available capacity of the site, respectively. Taking the minimum of these values ​​ensures that the reference power always lies within the current feasible region. The outgoing channel capacity is limited by subsequent independent constraints. This reference power will be passed to S5 as the total base value when the chromosome is decoded into the actual active power of each scheduling unit.

[0205] S32. Construct system operation safety, unit life protection, power generation fairness and transmission channel utilization indicators and objective functions;

[0206] The operational safety indicator uses the power mean and standard deviation within a sliding window to define the power fluctuation coefficient. The sliding window mean is:

[0207] ;

[0208] The standard deviation of the sliding window is:

[0209] ;

[0210] The power fluctuation coefficient is:

[0211] ;

[0212] In the formula, The mean of the sliding window; This represents the number of sampling points within the sliding time window. , Based on the closest to the current time The average and standard deviation of the unit power calculated over a historical period; For the unit At any moment The actual active power; The standard deviation of the sliding window; To prevent the introduction of extremely small positive numbers with a denominator of zero; This is the power fluctuation coefficient. The larger the unit power fluctuation coefficient, the more severe the power oscillation.

[0213] The objective function for system operation safety is to minimize the sum of the power fluctuation coefficients of all units at all times.

[0214] ;

[0215] In the formula, The objective function for ensuring system operation safety; The length of the scheduling time domain.

[0216] The unit life protection index uses an improved load factor to characterize life loss related to high load operation:

[0217] ;

[0218] In the formula, For the first Taiwanese crew The instantaneous fatigue coefficient of the machine at a given moment; For the first Taiwanese crew The actual active power (MW) at any given time; For the first Rated installed capacity (MW) of the unit.

[0219] The objective function for unit life protection aims to minimize the sum of improved load rates for all units at all times:

[0220] ;

[0221] In the formula, This is the objective function for protecting the unit's lifespan.

[0222] The fatigue distribution coefficient is defined as the degree of dispersion of the improved load rate of each power station or unit relative to the average value, which is used as the power generation fairness index.

[0223] ;

[0224] In the formula, for The fatigue distribution coefficient of the wind farm at any given time ranges from [0,1]. The smaller the value, the more uniform the fatigue of each unit in the wind farm. Conversely, the larger the value, the greater the difference in fatigue between the units, and one or more units may fail prematurely.

[0225] The objective function for power generation fairness aims to minimize the sum of fatigue distribution coefficients at all times.

[0226] ;

[0227] In the formula, The objective function is the fairness of power generation.

[0228] The utilization rate of the transmission channel is represented by the abandonment rate of the transmission channel.

[0229] ;

[0230] In the formula, The abandonment rate of the unit's power supply channel; This represents the maximum active power that the unit can theoretically generate under actual wind conditions at the current moment. To achieve the goal of "high utilization rate of the transmission channel," power waste caused by scheduling restrictions should be minimized as much as possible. The transmission channel abandonment rate is used to characterize the degree of underutilization of the wind turbine's transmission capacity under restricted scheduling conditions; therefore, this indicator should be as small as possible.

[0231] The objective function for transmitting channel utilization uses the transmitting channel abandonment rate to characterize the transmitting channel utilization rate.

[0232] ;

[0233] In the formula, The objective function is the output channel utilization rate.

[0234] The constraints of the multi-objective optimization scheduling model for offshore wind power clusters include:

[0235] ;

[0236] In the formula, The total power of the wind farm cluster at time t is the reference value. The sum of the actual active power of each scheduling unit should strictly track the actual allocated reference power at the cluster level at each scheduling time.

[0237] ;

[0238] In the formula, The maximum permissible climbing rate.

[0239] ;

[0240] In the formula, The overload factor is taken as the short-term overload factor. Considering the short-term over-generation capacity of offshore wind power under specific wind conditions, this model allows for a small overload. ; This is the minimum grid-connected power of the unit;

[0241] ;

[0242] In the formula, The upper limit of the power that can be generated at time t is determined by the prediction results; the upper limit of the prediction boundary is given by the aforementioned conventional prediction, meteorological deviation correction and typhoon dynamic correction link, and is used to limit the power that the cluster can generate at each scheduling time.

[0243] ;

[0244] In the formula, This represents the maximum transmission capacity limit for the outgoing channel. In the event of an outgoing channel failure, this limit will dynamically shrink based on the failure status.

[0245] In the event of a fault in the outgoing channel, the available capacity of the corresponding station or the upper limit of the outgoing channel capacity is dynamically reduced based on the fault isolation results, so that the optimizer only searches within the feasible region after the fault. The above six types of constraints will be passed to S5 as objects to be processed by the adaptive penalty function in the NSGA-II solution process.

[0246] S4. An improved dual-objective optimization model for solving NSGA-II is constructed based on the multi-objective optimization scheduling model of offshore wind power clusters. The system operation safety objective function, unit life protection objective function, power generation fairness objective function, and transmission channel utilization objective function established in S3 are normalized respectively. The normalized system operation safety objective value, normalized unit life protection objective value, and normalized power generation fairness objective value are weighted and merged according to preset weights to construct the first optimization objective with the goal of minimizing the comprehensive internal operation quality value. The normalized transmission channel utilization objective value is constructed as the second optimization objective with the goal of minimizing the transmission channel abandonment rate. This dual-objective reconstruction reduces the Pareto solution set selection complexity when directly solving the four objectives together, and avoids the problem that the trade-off between internal operation quality and transmission capacity is obscured after merging the four objectives into a single objective.

[0247] To eliminate differences in the dimensions and orders of magnitude of the objectives, the four objective functions are normalized:

[0248] ;

[0249] In the formula, For the j-th original target f j The normalized target value; For the first One objective function; To normalize the lower bound; The upper bound for normalization is defined as the upper and lower bounds of the corresponding objective in the current candidate solution set. The normalization result is used to eliminate the differences in dimensions and orders of magnitude among the four types of objectives.

[0250] The three normalized objectives—system operation safety, unit life protection, and power generation fairness—are weighted accordingly. , and The weighted objective is constructed as F1, and the normalization objective of the outgoing channel utilization is constructed as F2, thus forming an improved bi-objective optimization model for solving NSGA-II:

[0251] ;

[0252] In the formula, The primary optimization objective is to achieve this. The second optimization objective; , and These correspond to weighting coefficients for system operational safety, unit lifespan, and wind farm power generation fairness, respectively, and satisfy the following conditions: .

[0253] The weights in this embodiment are set according to engineering priorities and simulation scenarios. All weights are non-negative, and the sum of all weights is 1. Under normal operating conditions, the objectives are relatively balanced, and the fairness weight is appropriately increased to reduce long-term power distribution imbalances between power stations. Under high-fluctuation operating conditions, the focus is on suppressing rapid power fluctuations, and the safety weight is set to 0.8. Under typhoon risk operating conditions, the focus is on reducing the continuous high-load exposure of critical power stations, and the lifetime weight is set to 0.6. Under transmission channel failure operating conditions, the focus is on meeting failure constraints and restoring feasible power distribution, and the safety weight is set to 0.7. A set of feasible operating conditions and weight mappings is shown in Table 1.

[0254] Table 1 Target Weight Mapping Table under Different Known Operating Conditions

[0255] Normal operating conditions 0.3 0.3 0.4 Smooth scheduling and peak load sharing High fluctuation conditions 0.8 0.1 0.1 Safety and rapid fluctuation suppression Typhoon risk conditions 0.3 0.6 0.1 Lifetime protection and high-load exposure control Sending channel failure conditions 0.7 0.2 0.1 Safe and feasible reallocation under fault constraints

[0256] The weights listed in Table 1 are a set of feasible parameters set according to project priority and simulation scenario in this embodiment, and are not the only limitation on the weight values. In practical applications, the weights of each target can be adjusted according to the scale of the wind farm, equipment characteristics, transmission channel conditions, and scheduling requirements, while maintaining the priority of the dominant target under the corresponding operating conditions.

[0257] S5. Solve the improved bi-objective optimization model based on NSGA-II, taking the first and second optimization objectives as the two optimization objectives of NSGA-II, and using an adaptive penalty function to handle constraint violations. Search for cluster power allocation schemes within the feasible region formed by physical constraints, and output the cluster collaborative optimization scheduling results. This step uses F1 and F2 constructed in S4 as the two optimization objectives of NSGA-II, and uses an adaptive penalty function to handle constraint violations, searching for cluster power allocation schemes within the feasible region formed by physical constraints. The improved NSGA-II optimization method referred to in this invention combines the bi-objective reconstruction and adaptive constraint handling in step four with the NSGA-II solution process; the non-dominated sorting, congestion distance calculation, crossover, and mutation of NSGA-II still use its conventional mechanisms.

[0258] S51. Encode the hourly output allocation ratio with real numbers and decode the chromosome into the actual active power of each scheduling unit.

[0259] Suppose the cluster contains N wind farms or wind turbines, and the scheduling cycle contains T time periods. The [number]th [unit / item] will be [then / ... The scheduling unit in the first The output allocation code for each time period is denoted as follows: Individual chromosomes are represented as:

[0260] ;

[0261] make Indicates the first The set of operable scheduling units for each time period. This indicates the encoding tolerance to prevent division by zero. Encoding values ​​within the same time period are normalized; when the encoding sum is not greater than the tolerance, it is evenly distributed among operable units; when no operable units exist, the output of each unit is set to zero.

[0262] ;

[0263] In the formula, This refers to the actual active power. The actual reference power allocated to the cluster level in the t-th time period; For the first The scheduling unit in the first Output allocation code for each time period; For the first The set of operable scheduling units for each time period; To prevent encoding tolerance for division by zero;

[0264] The integrated internal operational quality target of S4 and the target of transmission channel utilization Forming a dual-objective vector:

[0265] ;

[0266] S52. The improved biobjective optimization model is iteratively solved using the constraint-controlled NSGA-II; specifically including:

[0267] S521. Initialize the parent population and calculate the biobjective value and constraint violation degree:

[0268] Randomly initialize the parent population within the upper and lower bounds of the variables, and rewrite the inequality constraints as follows: Write the equality constraints as and respectively according to the allowable scale and Dimensionless transformation:

[0269] ;

[0270] In the formula, , q and r are the normalized violation degrees of inequality and equality constraints, respectively; q and r are the constraint indices. , which is the q-th inequality constraint; , is the r-th equality constraint; For candidate scheduling solutions; , These are the allowable scale and reference scale used for dimensionless transformation of the corresponding constraint violation degree, respectively, and are not directly equivalent to the upper limit of the constraint.

[0271] Based on the current known operating condition s, the weights of various constraints are set to form the adaptive total constraint violation degree:

[0272] ;

[0273] In the formula, The degree of total constraint violation under adaptive operating conditions; and These are the violation weights of inequality constraints and equality constraints under operating condition s, respectively; when When the time comes, the candidate solution is determined to be a feasible solution. To allow for feasibility tolerance; , These represent the number of inequality constraints and the number of equality constraints, respectively. Current running status / scenario;

[0274] S522. Compare candidate solutions based on constraint dominance relationships. For candidate solutions... and ,when Feasible When not feasible, Superior When neither is feasible, the solution with the lower total constraint violation takes precedence; when both are feasible, they are compared according to the bi-objective Pareto dominance relation. The dominance relation between feasible solutions is as follows:

[0275] ;

[0276] In the formula, and Candidate solution; For candidate solutions The objective value on the m-th optimization objective; For candidate solutions The objective value for the m-th optimization objective; m∈{1,2};

[0277] S523. Perform a fast non-dominated sort and record the dominant solution set for each individual p. and the number of individuals that dominate that individual. . The individuals with zero dominance constitute the first non-dominated frontier; by subtracting the dominance count layer by layer, the remaining non-dominated frontiers are obtained.

[0278] ;

[0279] In the formula, To constrain the relationship of domination; Let p be the set of individuals that are governed by individual p. q represents the number of individuals that dominate individual p; p and q are the population individual indices; The first non-dominated frontier, i.e., n p A set of individuals whose value is 0;

[0280] S524. Calculate the crowding distance of individuals within each non-dominated frontier. Sort each target separately, setting the crowding distance of boundary individuals to infinity, and the crowding distance of internal individuals... The crowded distance is:

[0281] ;

[0282] In the formula, For the target scale; To achieve target normalization tolerance; , These are the maximum and minimum values ​​of the m-th objective in the current non-dominated frontier, respectively; , These are the target values ​​of adjacent individuals after sorting by the m-th target; Let be the crowding distance of individual i; when two individuals have the same non-dominance rank, the individual with the larger crowding distance is given priority in order to maintain the uniformity of the Pareto front distribution.

[0283] S525. Parent selection is achieved through a binary tournament. Two candidate individuals are randomly selected from the current population, prioritizing those with lower non-dominant rank; if ranks are the same, those with greater crowding distance are selected. The selection criteria are expressed as follows:

[0284] ;

[0285] In the formula, , These are two candidate individuals participating in the binary tournament; , It is a non-dominant class; , These are their respective crowding distances; To select priority relationships, those with lower non-dominant rank take precedence, and when ranks are the same, those with greater crowding distance take precedence.

[0286] S526. Use arithmetic crossover and Gaussian mutation to generate the offspring population:

[0287] With crossover probability Perform crossover on individuals in the current parent population; for the two selected parent individuals, on the... Two crossover offspring are generated using arithmetic crossover on each variable:

[0288] ;

[0289] In the formula, The arithmetic cross coefficient follows the following rules: (0,1) Uniformly distributed; , Let j be the value of the j-th variable of the two parent individuals; j is the index of the decision variable. Let be the values ​​of the two offspring generated after arithmetic crossover on the j-th decision variable, i.e. and , (0) and (1) are used to distinguish two cross-off offspring, and do not represent powers or numerical ranges;

[0290] With the probability of mutation Select the individuals to be mutated, and categorize them according to the proportion of variable variation in a single mutated individual. Select the variation location and use the standard deviation as . Gaussian perturbation generates mutated offspring:

[0291] ;

[0292] In the formula, For the mutated first One variable value; For the first time before the mutation One variable value; The standard deviation of the Gaussian perturbation; Let be a random variable that follows a standard normal distribution;

[0293] Perform truncation repair on components that exceed the upper and lower bounds of the variable after crossover or mutation:

[0294] ;

[0295] In the formula, , The first The lower and upper bounds of each decision variable.

[0296] S527. Merge the parent and offspring populations and implement elite preservation, cross-generational generation. Mutant subalgebras and temporary merging of populations The sizes are respectively:

[0297] ;

[0298] The parent population, crossover offspring population, and mutated offspring population of generation k are merged. The merged population is then re-executed with constrained dominance sorting and crowding distance calculation. Finally, the populations are truncated according to non-dominance level from low to high and crowding distance from large to small. Individual:

[0299] ;

[0300] In the formula, This is the parent population of the kth generation; , These represent the crossover offspring and the mutated offspring generated in the k-th generation, respectively. Population size; For the next generation of population; To truncate and retain N according to non-dominance level ascending order and congestion distance descending order pop Individual.

[0301] S528. Determine the termination condition:

[0302] Let k be the current iteration algebra, The maximum number of iterations. When When returning to S525 to continue generating offspring; The iteration terminates and the current population is output.

[0303] S53. Select a compromise scheduling scheme from the set of feasible Pareto optimal solutions.

[0304] First, retain the candidate solutions with a non-dominated level of 1 and a total constraint violation degree no greater than the feasibility tolerance:

[0305] ;

[0306] In the formula, This is the set of feasible Pareto candidate solutions; This is the i-th candidate solution; It is a non-dominant class;

[0307] To prevent normalization division by zero when a certain objective has the same value in the candidate set, a target scale is defined for selecting a compromise solution:

[0308] ;

[0309] In the formula, To prevent the target scale from being divided by zero during normalization;

[0310] Calculate the normalized distance from each feasible solution to the bi-objective ideal point, and select the compromise solution with the smallest distance:

[0311] ;

[0312] In the formula, #imgpt498# and #imgpt499# are the minimum and maximum values ​​of the m-th objective in the feasible Pareto solution set, respectively; #imgpt500# is the compromise scheduling scheme with the minimum distance. The selection of the ideal point distance belongs to the decision-making process after the formation of the Pareto solution set and is not considered as an improvement to the NSGA-II basic search mechanism.

[0313] Taking a deep-sea offshore wind farm cluster in my country as an example, this study uses SCADA operation data at the farm level from 2021 to 2024 and historical meteorological reanalysis data for model development and threshold calibration. Operational forecast boundaries are generated using bias-corrected numerical weather prediction. The cluster comprises five equivalent farm-level dispatch units with capacities of 147.7MW, 102.0MW, 206.4MW, 100.8MW, and 112.0MW, respectively, for a total installed capacity of 668.9MW.

[0314] Without meteorological bias correction, the MAE, RMSE, coefficient of determination R², and capacity normalization accuracy of representative station 3 were 23.548 MW, 32.969 MW, 0.738, and 88.591%, respectively. After meteorological bias correction, these improved to 21.213 MW, 30.023 MW, 0.783, and 89.722%, respectively. Further typhoon dynamic correction resulted in 21.161 MW, 29.919 MW, 0.784, and 89.748%, respectively. Figure 2 As shown, the main function of typhoon dynamic correction is to make the prediction boundaries more consistent with actual protection behavior during periods of sudden strong wind changes, power drops, and recovery.

[0315] Under high volatility conditions, the optimized average total ramp rate decreased from 135.26 MW / h to 129.56 MW / h; the maximum improved load factor range decreased from 96.29% to 62.83%, and the average improved load factor standard deviation decreased from 18.20% to 9.70%. This indicates that the dispatching strategy mainly reduces the imbalance between stations through internal reallocation, as shown in the results. Figure 3 As shown, Figure 3 In the diagram, (a) is the stacked diagram of the unit power distribution after optimization for 24 hours, (b) is the comparison diagram of the total ramp amount before and after optimization, and (c) is the comparison diagram of the load distribution balance before and after optimization.

[0316] Under typhoon risk conditions, the maximum improved load rate range decreased from 99.58% to 55.46%, and the average improved load rate standard deviation decreased from 18.02% to 8.39%. The number of periods when the load rate of station 3 exceeded 90% decreased from 13 to 2, indicating that increasing the lifetime protection weight can reduce the continuous high load exposure of critical stations during sudden drops and recovery processes. Figure 4 As shown, Figure 4In the diagram, (a) is the stacked diagram of the unit power distribution after optimization for 24 hours, (b) is the comparison diagram of the total ramp amount before and after optimization, and (c) is the comparison diagram of the load distribution balance before and after optimization.

[0317] After power station 3 was isolated due to the fault, the available power transmission capacity of the cluster decreased from 668.9MW to 462.5MW. During the initial two periods of the fault, unavoidable power shortfalls of 33MW and 67MW occurred respectively. After the remaining operational power stations reached their respective capacity limits, the system largely resumed tracking of the feasible boundary, as shown in the following results. Figure 5 As shown, Figure 5 In the diagram, (a) shows the optimized 24-hour unit power allocation stacking diagram, (b) shows the comparison of total ramp-up before and after optimization, and (c) shows the comparison of load distribution balance before and after optimization. Under this condition, some balance indicators may deteriorate, which is a feasibility cost caused by the sudden drop in available capacity. The evaluation focuses on whether a redistribution scheme that meets fault constraints can be quickly formed.

[0318] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.

Claims

1. A multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction, characterized in that: The method includes the following steps: S1. Construct a wind farm output prediction model under normal conditions, and obtain the predicted wind farm output under normal conditions based on the wind farm output prediction model. S2. Based on the wind farm predicted output obtained from S1, the wind farm intensity distortion caused by extreme weather and the impact of unit cut-out are corrected to obtain the corrected prediction boundary. S3. Construct a multi-objective optimization scheduling model for offshore wind power clusters, including a system operation safety objective function aimed at minimizing power fluctuations, a unit life protection objective function aimed at minimizing the improved load rate of the units, a power generation fairness objective function aimed at minimizing the fatigue distribution coefficient, and a transmission channel utilization objective function aimed at minimizing the curtailment rate of the transmission channels. S4. Based on the multi-objective optimization scheduling model of offshore wind power clusters, an improved bi-objective optimization model for solving NSGA-II is constructed. The system operation safety objective function, unit life protection objective function, power generation fairness objective function and transmission channel utilization objective function established in S3 are normalized respectively. The normalized system operation safety objective value, normalized unit life protection objective value and normalized power generation fairness objective value are weighted and merged according to preset weights to construct the first optimization objective with the goal of minimizing the comprehensive internal operation quality value. The normalized transmission channel utilization objective value is constructed as the second optimization objective with the goal of minimizing the transmission channel abandonment rate. S5. Solve the improved bi-objective optimization model based on NSGA-II, taking the first optimization objective and the second optimization objective as the two optimization objectives of NSGA-II, and using an adaptive penalty function to handle constraint violations, search for cluster power allocation schemes within the feasible region formed by physical constraints, and output the cluster collaborative optimization scheduling results.

2. The multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 1, characterized in that: S1 specifically includes: S11. Acquire and process wind farm operation data, meteorological data, and equipment parameters, and establish a physical baseline model, represented as follows: ； ； ； In the formula, It contributes to the physical model of the wind farm; For single-machine physical output; This refers to the number of wind turbines. This is the overall efficiency coefficient; This refers to the rated power of a single unit. The equivalent wind speed at the hub height; , and These are the cut-in wind speed, rated wind speed, and cut-out wind speed, respectively. Wind shear index; For height Wind speed at the location; Wheel hub height; S12. Construct meteorological physical characteristics and time periodic characteristics, establish an XGBoost data-driven compensation model to compensate for the residuals of the physical baseline model, and obtain the predicted power under normal conditions.

3. The multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 1, characterized in that: The XGBoost data-driven compensation model is as follows: ； In the formula, For the number of regression trees, The output function for the r-th regression tree of the predicted object k; The predicted power of object k at sample time t or date d at time t; The input feature vector; The input to the XGBoost data-driven compensation model is: ； In the formula, The input feature vector; It is a set of meteorological and physical characteristics; It is a set of time and periodic features; It contributes to the physical model of the wind farm; The XGBoost data-driven compensation model is trained using loss function terms and model complexity regularization terms: ； In the formula, To train the objective function for the overall model built for the prediction object k; The prediction error loss function; This is the set of training samples corresponding to the predicted object; The actual power of sample d at time t; Let be the complexity regularization term for the r-th regression tree; In the current iteration, we introduce the first and second derivatives of the loss function with respect to the predicted values: ； In the formula, Let be the first derivative of the loss function with respect to the predicted value of the i-th sample; Let be the second derivative of the predicted value of the i-th sample; This represents the actual power of the i-th sample. t′−1 is the predicted value obtained in the model iteration round; t′ is the model iteration round.

4. The multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 1, characterized in that: S2 specifically includes: S21. Establish a typhoon wind field model based on the typhoon path and structural parameters, obtain the latitude and longitude of the typhoon center, the maximum wind speed of the typhoon center and the maximum wind speed radius, calculate the distance between the wind farm and the typhoon center, use the Rankine composite vortex model to approximate the radial wind speed of the typhoon, and obtain the effective wind speed of the typhoon at the location of the wind farm by combining the environmental wind field and the typhoon rotating wind field. ； In the formula, The effective typhoon wind speed at the location of the wind farm; This represents the distance between the wind farm and the center of the typhoon. This is a Rankine complex vortex model; S22. Based on the conventional predicted power of S1, and combined with the typhoon's impact range and the unit's disconnection protection logic, a revised prediction boundary is formed, specifically as follows: ； In the formula, This represents the revised predicted power boundary under the influence of the typhoon. This is the baseline predicted power obtained from numerical weather prediction; For numerical weather forecast wind speed; Cut off the wind speed for the unit; and This is a correction factor.

5. A multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 1, characterized in that: The objective function for system safety is: ； ； In the formula, The objective function for ensuring system operation safety; This represents the total number of wind turbines in the wind farm. Power fluctuation coefficient; , Based on the closest to the current time The average and standard deviation of the unit power calculated over a historical period; To prevent the introduction of positive numbers with a denominator of zero; The length of the scheduling time domain; The objective function for unit life protection is: ； ； In the formula, The objective function for protecting the unit's lifespan; For the first Taiwanese crew The instantaneous fatigue coefficient of the machine at a given moment; For the first Taiwanese crew The actual active power at any given moment; For the first Rated installed capacity of the unit; The objective function for power generation fairness is: ； ； In the formula, The objective function is the fairness of power generation. for Fatigue distribution coefficient of wind farm at any given time; The objective function for the utilization rate of the outgoing channel is: ； ； In the formula, The objective function is the output channel utilization rate; The abandonment rate of the unit's power supply channel; This represents the maximum active power that the unit can theoretically generate under actual wind conditions at the current moment.

6. A multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 5, characterized in that: The constraints of the multi-objective optimization scheduling model for offshore wind power clusters include: ； ； ； ； ； In the formula, The maximum permissible climbing rate; This is the minimum grid-connected power of the unit; This is the short-time overload factor; The upper limit of the generateable power at time t is determined by the prediction results; The reference value for the total power of the wind farm cluster at time t; This is the maximum transmission capacity limit for the outgoing channel.

7. A multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 5, characterized in that: The normalization in S4 is performed in the following manner: ； In the formula, For the first One objective function; To normalize the lower bound; The normalized upper bound; For the j-th original target f j The normalized target value; The first optimization objective and the second optimization objective are expressed as follows: ； In the formula, The primary optimization objective is to achieve this. The second optimization objective; , and These correspond to weighting coefficients for system operational safety, unit lifespan, and wind farm power generation fairness, respectively, and satisfy the following conditions: .

8. A multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 5, characterized in that: S5 specifically includes: S51. Encode the hourly output allocation ratio with real numbers and decode the chromosome into the actual active power of each scheduling unit. S52. The improved bi-objective optimization model is solved iteratively using the constraint-controlled NSGA-II; S53. Select a compromise scheduling scheme from the set of feasible Pareto optimal solutions.

9. A multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 8, characterized in that: The chromosome is decoded into the actual active power of each scheduling unit as follows: ； In the formula, This refers to the actual active power. The actual reference power allocated to the cluster level in the t-th time period; For the first The scheduling unit in the first... Output allocation code for each time period; For the first The set of operable scheduling units for each time period; To prevent encoding tolerance for division by zero.

10. A multi-objective cooperative scheduling method for wind farm clusters considering predicted power correction according to claim 8, characterized in that: S52 specifically includes: S521. Initialize the parent population and calculate the biobjective value and constraint violation degree: Write the inequality constraint as Write the equality constraints as and respectively according to the allowable scale and Dimensionless transformation: ； In the formula, , q and r are the normalized violation degrees of inequality and equality constraints, respectively; q and r are the constraint indices. , which is the q-th inequality constraint; , is the r-th equality constraint; For candidate scheduling solutions; , These are the allowable scale and reference scale used for dimensionless transformation of the corresponding constraint violation degree, respectively; Based on the current known operating condition s, the weights of various constraints are set to form the adaptive total constraint violation degree: ； In the formula, The degree of total constraint violation under adaptive operating conditions; and These are the violation weights of inequality constraints and equality constraints under operating condition s, respectively. , These represent the number of inequality constraints and the number of equality constraints, respectively. Current running status / scenario; S522. Compare candidate solutions based on constraint dominance relationships: ； In the formula, and Candidate solution; For candidate solutions The objective value on the m-th optimization objective; For candidate solutions The objective value on the m-th optimization objective; S523, Perform quick non-dominated sort: ； In the formula, To constrain and dominate relationships; Let p be the set of individuals that are governed by individual p. q represents the number of individuals that dominate individual p; p and q are the population individual indices; It is the first non-dominant frontier; S524. Calculate the crowding distance for individuals within each non-dominated front: 0； In the formula, For the target scale; To achieve target normalization tolerance; , These are the maximum and minimum values ​​of the m-th objective in the current non-dominated frontier, respectively; , These are the target values ​​of adjacent individuals after sorting by the m-th target; Let i be the crowding distance for individual i; S525, Selecting the parent generation through a binary tournament: ； In the formula, , These are two candidate individuals participating in the binary tournament; , It is a non-dominant class; , These are their respective crowding distances; To select priority relationships; S526. Use arithmetic crossover and Gaussian mutation to generate the offspring population: Perform crossover on individuals in the current parent population with crossover probability; for the two selected parent individuals, in the first... Arithmetic crossover is used to generate two crossover offspring for each variable: ； In the formula, The arithmetic cross coefficient follows the following rules: (0,1) Uniformly distributed; , Let j be the value of the j-th variable of the two parent individuals; j is the index of the decision variable. Let be the values ​​of the two offspring generated after arithmetic crossover on the j-th decision variable, i.e. and , (0) and (1) are used to distinguish two cross-off offspring, and do not represent powers or numerical ranges; Individuals to be mutated are selected based on the mutation probability, and mutation sites are selected according to the proportion of variable mutation in a single mutated individual, using a standard deviation of 1. Gaussian perturbation generates mutated offspring: ； In the formula, For the mutated first One variable value; The first before the mutation One variable value; The standard deviation of the Gaussian perturbation; Let be a random variable that follows a standard normal distribution; Perform truncation repair on components that exceed the upper and lower bounds of the variable after crossover or mutation: ； In the formula, , The first Lower and upper bounds of each decision variable; S527. Merge parent and offspring populations and implement elite preservation: ； In the formula, This is the parent population of the kth generation; , These represent the crossover offspring and the mutated offspring generated in the k-th generation, respectively. Population size; For the next generation of population; To truncate and retain N according to non-dominance level ascending order and congestion distance descending order pop Individual.

Images