A method for active power distribution of actual wind farm considering wind speed prediction and fatigue damage

Abstract

This invention relates to a method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage. The method includes: selecting several model turbines in the target wind farm as benchmarks; clustering non-model turbines to obtain several turbine clusters; predicting the wind speed of the model turbines based on a prediction model using spatiotemporal feature extraction; obtaining the wind speed prediction results for the model turbines; inputting the wind speed prediction results of the model turbines into an intra-cluster spatial mapping model to obtain the wind speed prediction results for the non-model turbines; and, based on the target active power command issued by the power grid, ambient temperature, real-time operating status, and wind speed prediction results, using an average power allocation mechanism when the ambient temperature is below a preset temperature, and using a multi-dimensional optimization mechanism when the ambient temperature is not below the preset temperature. This invention maximizes the overall wind energy utilization rate of the wind farm and effectively extends the global mechanical life of the turbine cluster.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and in particular to a method for active power allocation in a real wind farm that takes into account wind speed prediction and fatigue damage. Background Technology

[0002] As wind power penetration in new power systems continues to increase, wind farms need to strictly respond to active power dispatch commands issued by the power grid and rationally allocate them to each wind turbine unit within the farm. However, in the actual operation and dispatch of wind farms, existing active power allocation strategies still have the following significant shortcomings: Under the influence of complex wind farm environments, the effective wind speed of wind turbines exhibits significant spatiotemporal heterogeneity. Although advanced machine learning models can provide excellent wind speed prediction accuracy, their complex network structures and lengthy training processes result in extremely high computational costs. In practical applications, directly deploying these complex models in parallel across all wind turbines within a wind farm would create a massive computational burden, severely hindering their engineering application in real-time dynamic allocation of active power.

[0003] Power dispatching in actual wind farms must balance meteorological conditions and equipment physical constraints. In severe winter environments, the kinematic viscosity of the gearbox lubricating oil increases dramatically, severely reducing its fluidity. Existing allocation strategies often ignore dynamic changes in ambient temperature. If start-stop commands are frequently issued or turbine power is drastically adjusted under extremely low oil temperatures, it introduces significant mechanical resistance, leading to a sudden increase in starting torque and potentially causing generator overload or even catastrophic mechanical failure. Under normal operating conditions such as medium and high temperatures, existing power allocation often excessively pursues immediate power generation efficiency. The drawbacks of this strategy are twofold: firstly, when local wind resources are favorable, the algorithm forces specific turbines to operate at full load for extended periods, rapidly exacerbating the cumulative fatigue damage of these units; secondly, after high-scoring turbines are prioritized for full-load operation, the remaining "tail-end adjustment" (fine-tuning) quota is often fixedly allocated to a specific subset of low-scoring turbines, causing these turbines to bear excessive mechanical stress due to frequent operational adjustments. This allocation method, which ignores long-term potential and adjustment fatigue, significantly shortens the overall mechanical lifespan of the wind farm. Summary of the Invention

[0004] The purpose of this invention is to provide a method for allocating active power in a real wind farm that takes into account wind speed prediction and fatigue damage, thereby maximizing the overall wind energy utilization rate of the wind farm and effectively extending the overall mechanical life of the unit group.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for active power allocation in a real wind farm that considers wind speed prediction and fatigue damage includes: Several prototype turbines were selected in the target wind farm as benchmarks, and non-prototype turbines were clustered to obtain several turbine clusters; The wind speed of the prototype machine is predicted by a prediction model based on spatiotemporal feature extraction, and the wind speed prediction result of the prototype machine is obtained. The wind speed prediction results of the prototype machine are input into the intra-cluster spatial mapping model to obtain the wind speed prediction results of the non-prototype machine. The intra-cluster spatial mapping model constructs the feature mapping relationship between the prototype machine and the non-prototype machine through a ridge regression model and is obtained by training with the historical wind speed of the prototype machine. Based on the target active power command issued by the power grid, ambient temperature, real-time operating status, and wind speed prediction results, the average power allocation mechanism is used for power allocation when the ambient temperature is lower than the preset temperature, and the multi-dimensional optimization mechanism is used for power allocation when the ambient temperature is not lower than the preset temperature.

[0006] Optionally, clustering non-sample machines includes: The Pearson correlation coefficient is used to quantify the synchronicity of wind speed evolution trends between the prototype machine and the non-prototype machine. The correlation coefficient is calculated, and the maximum correlation criterion is used to perform clustering based on the correlation coefficient. Alternatively, the standardized time series of the prototype machine can be used as the initial cluster center set of the K-Means algorithm. By minimizing the intra-cluster squared error and assigning the time series feature vector of each non-prototype machine to the centroid with the shortest Euclidean distance, the centroid position is dynamically updated according to the current cluster members until the objective function converges, and the final wind turbine clustering result is obtained.

[0007] Optionally, the prediction model based on spatiotemporal feature extraction for predicting the wind speed of the prototype includes: For the wind turbine cluster, a low-rank parameterized adjacency matrix is ​​used to represent the spatial topology; Based on the spatial topology and the historical wind speed observation feature matrix of the prototype wind turbine in the wind turbine cluster, a graph convolutional network is used to aggregate the node features within the cluster to obtain spatial enhancement features. The spatial enhancement features are input into the time series decomposition module to extract time features, and the time features are decomposed into long-term trend components and seasonal fluctuation components using a moving average mechanism. The decomposed components are projected onto the future preset step prediction time domain of the prototype machine through a linear layer, and the projected components are superimposed to obtain the wind speed prediction result of the prototype machine.

[0008] Optionally, training the ridge regression model using the historical wind speeds of the prototype machine includes: solving for the optimal mapping weight vector by minimizing the penalized sum of squared residuals. β : ; in, Let X be the objective function, representing the historical wind speed observation feature matrix obtained from the prototype machine, and y be the actual wind speed vector of the target non-prototype machine. This represents the squared L2 norm of the weight vector β. It is the regularization coefficient that controls the complexity of the model.

[0009] Optionally, when the ambient temperature is lower than a preset temperature, an average power distribution mechanism may be used for power distribution, including: when At this time, it is the zero-dispatch operating condition, and the active power allocation of all wind turbines is 0. Target active power command: when At that time, under ultra-low load conditions, only the pilot unit with the lowest available capacity was dispatched to complete the task, while the remaining wind turbines maintained zero active power output. This represents the minimum non-zero available power among all the prototype machines. when At that time, under low-load conditions, a dynamic uniform distribution mechanism with saturation constraints is triggered among the prototype machines, wherein... The available power for polymerization of the prototype machine; when At that time, under medium to high load conditions, the prototype machine is prioritized to operate at full load, outputting aggregated usable power, and the dynamic power residual is calculated. And evenly distributed to non-sample machines, among which, This represents the theoretical total available power of the entire wind farm. when During this period, which is a capacity-constrained operating condition, all wind turbines are dispatched to generate electricity at their current maximum available capacity.

[0010] Optionally, triggering a dynamic uniform allocation mechanism with saturation constraints among prototypes includes: Calculate the maximum aerodynamic power of the wind turbine at the current wind speed; Calculate the average baseline quota for the current set based on the target power to be allocated and the number of wind turbines participating in operation; If the maximum aerodynamic power is less than the average baseline quota, the fan corresponding to the maximum aerodynamic power is recorded as a saturated fan, the maximum dynamic power is deducted from the target active power, and the saturated fan is removed from the active set. The remaining unallocated power continues to be evenly distributed and iterated in the updated unsaturated fan subset until all power quotas are fully allocated.

[0011] Optionally, when the ambient temperature is not lower than a preset temperature, a multi-dimensional optimization mechanism is used for power allocation, including: when At that time, a dynamic comprehensive evaluation index under multi-state constraints is constructed for non-template machines, and greedy active power allocation is performed based on the dynamic comprehensive evaluation index, wherein, The target active power command, The available power for polymerization of the prototype machine. This represents the theoretical total available power of the entire wind farm.

[0012] Optionally, the dynamic comprehensive evaluation index The calculation method is as follows: ; ; in, Based on the predicted score, , and Let represent the hysteresis reward, tail adjustment penalty, and macroscopic energy fatigue penalty of wind turbine i at time t, respectively. The available power at the current moment. To achieve the minimum grid-connected power at the cut-off wind speed, This represents the theoretically available power generation potential.

[0013] Optionally, the basic prediction score is calculated as follows: ; in, The weighting factor for the current wind speed. and Let these represent the available power at the current moment and at the τ-th future step, respectively. Calculated based on wind speed prediction results Here, H is the confidence decay coefficient, and H is the total number of future steps.

[0014] Optionally, greedy active power allocation based on the dynamic comprehensive evaluation index includes: The dynamic comprehensive evaluation indicators are sorted in descending order, and the dynamic power residual is allocated to the non-sample unit with the highest score. During the allocation process, the active power allocation of the wind turbines is as follows: , For dynamic power residual, Theoretically available power generation potential; The remaining unallocated power is dynamically updated, and active power instructions are issued sequentially according to the sorted list until the dynamic power residual is completely exhausted.

[0015] The beneficial effects of this invention are as follows: This invention adopts a computationally efficient two-stage prediction strategy. First, it uses an adaptive spatial feature extractor and a time series decomposition (GCN+DLinear) module to perform high-precision prediction on the "sample turbine unit (sample turbine)". Then, it uses a ridge regression model to construct a spatial mapping function, and directly extrapolates the predicted wind speed of non-sample turbine units based on the prediction results of the sample turbine unit. This scheme avoids applying complex machine learning models to all wind turbines simultaneously. While ensuring the inherent spatiotemporal correlation characteristics of the wind farm from both mathematical and physical perspectives, it greatly reduces the overall computational complexity and computing power overhead of the algorithm, thereby meeting the real-time engineering application requirements of dynamic power scheduling in wind farms.

[0016] This invention introduces ambient temperature as a strategy branch condition after obtaining scheduling instructions. In low-temperature (winter) scenarios, a dynamic uniform distribution mechanism with saturation constraints is triggered first to reduce frequent start-stop of individual units. By maintaining stable gearbox oil temperature and uniform load distribution within the cluster in frigid environments, it effectively avoids the huge mechanical resistance and sudden increase in starting torque caused by frequent adjustments due to excessively high lubricating oil viscosity. This fundamentally reduces the risk of wind turbine overload and mechanical failure in winter, ensuring the safe operation of the unit.

[0017] This invention constructs a dynamic comprehensive evaluation index, Jfinal, under multi-state constraints for non-model wind turbines in medium- and high-temperature scenarios. This index not only includes a "basic prediction reward" that integrates current and future predicted power, but also innovatively introduces a "tail-end adjustment fatigue penalty" to punish frequent fine-tuning tasks, and a "macro-energy fatigue penalty" to limit long-term full-load operation of a single turbine. Greedy active power allocation is then performed based on this dynamic evaluation index. This evaluation mechanism, incorporating state feedback, effectively breaks the deadlock of traditional greedy algorithms where a few turbines operate at full load for extended periods or undertake long-term tail-end fine-tuning tasks. While strictly meeting the overall grid dispatch instructions at the macro level, it avoids excessive mechanical stress and fatigue damage to individual turbines caused by frequent adjustments or continuous heavy loads at the micro level, thereby maximizing the overall wind energy utilization rate of the wind farm and effectively extending the overall mechanical life of the turbine group. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of an active power allocation method for a real wind farm that considers wind speed prediction and fatigue damage, according to an embodiment of the present invention. Figure 2This is a flowchart illustrating wind speed prediction according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the power allocation strategy of an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, this embodiment proposes a method for active power allocation in a real wind farm that considers wind speed prediction and fatigue damage, including: Several prototype turbines were selected in the target wind farm as benchmarks, and non-prototype turbines were clustered to obtain several turbine clusters; The wind speed of the prototype machine is predicted by a prediction model based on spatiotemporal feature extraction, and the wind speed prediction result of the prototype machine is obtained. The wind speed prediction results of the prototype machine are input into the intra-cluster spatial mapping model to obtain the wind speed prediction results of the non-prototype machine. The intra-cluster spatial mapping model constructs the feature mapping relationship between the prototype machine and the non-prototype machine through a ridge regression model and is obtained by training with the historical wind speed of the prototype machine. Based on the target active power command issued by the power grid, ambient temperature, real-time operating status, and wind speed prediction results, the average power allocation mechanism is used for power allocation when the ambient temperature is lower than the preset temperature, and the multi-dimensional optimization mechanism is used for power allocation when the ambient temperature is not lower than the preset temperature.

[0023] Furthermore, clustering of non-sample machines includes: The Pearson correlation coefficient is used to quantify the synchronicity of wind speed evolution trends between the prototype machine and the non-prototype machine. The correlation coefficient is calculated, and the maximum correlation criterion is used to perform clustering based on the correlation coefficient. Alternatively, the standardized time series of the prototype machine can be used as the initial cluster center set of the K-Means algorithm. By minimizing the intra-cluster squared error and assigning the time series feature vector of each non-prototype machine to the centroid with the shortest Euclidean distance, the centroid position is dynamically updated according to the current cluster members until the objective function converges, and the final wind turbine clustering result is obtained.

[0024] Specifically, in wind farms, a set of model turbines is typically evenly distributed across different geographical locations and terrain areas to accurately reflect the region's power generation potential. Therefore, this embodiment uses the spatial distribution of these model turbines as a benchmark to divide the wind farm into different clusters. Furthermore, although wind speed fundamentally determines the output power of wind turbines, in actual operation, due to curtailment and manual dispatch intervention, the recorded actual active power is often lower than the theoretical available power corresponding to the current wind speed. This discrepancy makes historical power records unreliable in representing the true wind energy potential. Therefore, this embodiment selects wind speed as a key feature for wind turbine clustering and active power allocation strategies.

[0025] In wind farms, the effective wind speed of a single wind turbine exhibits significant spatial heterogeneity and temporal dynamics due to the coupling effect of micro-topography and meteorological randomness. Therefore, this embodiment selects K geographically representative sample turbines as benchmarks and extracts the wind speed time series of all turbines to construct a feature space. Based on this, two different wind turbine clustering strategies are proposed and compared.

[0026] (1) Deterministic similarity matching based on Pearson correlation coefficient: Assume that all wind turbines in the wind farm are divided into a set of prototype turbines. and unassigned non-template collection During the observation period T, the wind speed time series vector of any wind turbine i can be represented as: .

[0027] Pearson correlation coefficient was used to quantify the synchronicity of wind speed evolution trends between the prototype and non-prototype machines. For any non-prototype machine... and prototype machine The correlation coefficient is calculated as follows: (1); in, and These represent the non-sample machines during the observation period T. With the prototype machine The average wind speed.

[0028] After obtaining the complete correlation coefficient matrix, the maximum correlation criterion was used for clustering. For each non-sample machine... , recognition Template index for obtaining the maximum value and will Assigned to the cluster corresponding to this prototype machine : (2); This process ensures that each wind turbine is assigned to a sample cluster that most closely resembles its wind speed fluctuation trend.

[0029] (2) Seed-constrained K-Means dynamic clustering: The K-Means dynamic clustering method is characterized by its simplicity, high computational efficiency, and robustness in handling large-scale data. Before clustering, the wind speed time series of all wind turbines are standardized (Z-score normalization) to eliminate the influence of differences in units and baselines. The standardization formula is as follows: (3); in, This represents the standardized wind speed characteristics. This represents the average of the original wind speed data for the fan. It represents the standard deviation.

[0030] This embodiment proposes using the standardized time series of K prototype machines as the initial cluster center set for the K-Means algorithm. : (4); Under this prior constraint, the algorithm divides the wind turbines into K clusters S={S1,S2,…,S...} by minimizing the sum of squared errors within each cluster. K}, whose corresponding centroid is The objective function J is defined as follows: (5); In each iteration, the time-series feature vector of each wind turbine is assigned to the centroid with the shortest Euclidean distance; then the centroid position is dynamically updated according to the current cluster members until the objective function J converges, thus obtaining the final wind turbine clustering result.

[0031] Furthermore, the prediction model based on spatiotemporal feature extraction predicts the wind speed of the prototype machine, including: For the wind turbine cluster, a low-rank parameterized adjacency matrix is ​​used to represent the spatial topology; Based on the spatial topology and the historical wind speed observation feature matrix of the prototype wind turbine in the wind turbine cluster, a graph convolutional network is used to aggregate the node features within the cluster to obtain spatial enhancement features. The spatial enhancement features are input into the time series decomposition module to extract time features, and the time features are decomposed into long-term trend components and seasonal fluctuation components using a moving average mechanism. The decomposed components are projected onto the future preset step prediction time domain of the prototype machine through a linear layer, and the projected components are superimposed to obtain the wind speed prediction result of the prototype machine.

[0032] Specifically, after completing the aforementioned wind turbine clustering, the future wind speed of all individual wind turbines within the wind farm is predicted. Although advanced machine learning models have excellent prediction accuracy, their complex structure and lengthy training process lead to significant computational costs. Directly applying these models to predict the wind speed of all wind turbines simultaneously would result in excessive computational burden, hindering their real-time engineering applications. To overcome this limitation, this embodiment proposes a computationally efficient two-stage prediction strategy. First, the machine learning model is used only for predicting the wind speed of a prototype turbine; then, an intra-cluster spatial mapping relationship is established between the prototype and non-prototype turbines, allowing the wind speed of non-prototype turbines to be extrapolated based on the prediction results of the prototype turbines. The execution framework of this wind speed prediction process is shown in Figure 2.

[0033] Sample machine wind speed prediction: To achieve high-precision wind speed prediction for the prototype, this embodiment proposes a prediction framework based on spatiotemporal feature extraction. This framework seamlessly integrates an adaptive spatial feature extractor with a time prediction module based on linear decomposition, ensuring high computational efficiency while maintaining physical interpretability.

[0034] In the spatial domain, an adaptive graph learning mechanism is introduced to construct intra-cluster spatial mapping relationships. Assuming the wind farm is divided into K clusters, when predicting the wind speed of the i-th cluster consisting of N wind turbines, let the historical lookback window be L. The historical wind speed observation feature matrix of this cluster is defined as... Where B represents the batch size. To limit the computational complexity of the prediction model, a low-rank parameterized adjacency matrix A is used to represent the spatial topology, and its calculation form is as follows: (6); in, and Represents the learnable node embedding matrix, and r is a low-rank hyperparameter controlling the learning complexity of the graph; This represents the Sigmoid activation function, used to ensure that edge weights are confined within (0,1). Furthermore, L1 sparse regularization is used to prevent overfitting and eliminate redundant spatial connections.

[0035] Based on the learned spatial topology, graph convolutional networks (GCNs) are used to aggregate node features within clusters. Furthermore, to alleviate the vanishing gradient problem and stabilize the training process, a residual connection mechanism is introduced into the structure. The spatially augmented feature representation is as follows: (7); Where W is the learnable weight matrix of the graph convolutional layer.

[0036] Wind speed dynamics exhibit not only spatial correlation but also significant temporal dependence. Therefore, after obtaining the spatial enhancement features, they are input into the time series decomposition module. This embodiment uses the DLinear model to extract the temporal features of wind speed. To accurately capture the temporal evolution pattern of wind speed, a moving average mechanism is used to decompose the time series into long-term trend components and seasonal fluctuation components. The decomposition form is as follows: (8); Subsequently, the decomposed historical components are projected onto the prototype's future H-step prediction time domain using two independent linear layers. The final wind speed prediction is obtained by superimposing these two projected components, calculated as follows: (9); in, and These represent the learnable linear transformation weight matrices for the trend component and the seasonal component, respectively.

[0037] During the model optimization phase, the total loss function consists of the sum of the prediction error and the graph sparsity regularization term. To strictly prevent temporal data leakage and ensure the model's generalization ability, a time-series cross-validation strategy is employed during training. The optimal model, which achieves the lowest average validation loss across all trade-offs, is selected and further evaluated on the independent hold-out dataset. This comprehensive structure mathematically guarantees that the model can systematically utilize the temporal evolution trends and inherent spatial correlations within the wind farm, thereby significantly improving the accuracy and robustness of wind speed prediction.

[0038] Non-sample machine wind speed prediction: Wind turbines within the same cluster exhibit highly similar micro-topographic features and aerodynamic wake effects, resulting in strong linear or near-linear correlations in their wind speed variations. Therefore, after obtaining the wind speed prediction results for the prototype turbine, this section aims to establish the spatial mapping relationship within the cluster. To balance large-scale computational efficiency and prevent model overfitting, a ridge regression model with L2 regularization penalty is used to construct the feature mapping function between the prototype and non-prototype turbines.

[0039] Let X represent the historical wind speed observation feature matrix obtained from the prototype machine, and y represent the actual wind speed vector of the target non-prototype machine. Ridge regression solves for the optimal mapping weight vector β by minimizing the penalized sum of squared residuals. The objective function is defined as follows: (10); in, It is a regularization coefficient that controls the complexity of the model. Let β represent the squared L2 norm of the weight vector β. During the training phase, the model parameters are fitted using historical data; during the prediction phase, the future predicted wind speed of the prototype machine is input into the trained model to quickly derive the prediction sequence for non-prototype machines.

[0040] Furthermore, to accurately assess the robustness of the mapping model and prevent overfitting, this embodiment introduces a cross-validation mechanism. This strategy ensures the overall accuracy and generalization ability of wind speed prediction across the entire wind farm area.

[0041] Furthermore, after obtaining the wind speed forecast results for all wind turbines, the active power allocation strategy system integrates the overall grid dispatch command, ambient temperature, real-time operating status, and wind speed forecast results to achieve optimal power allocation within the wind farm. The overall execution framework of this strategy is shown in Figure 3.

[0042] In actual operation, wind farm power dispatch must continuously balance meteorological constraints and grid demand. Specifically, in severe winter environments, when gearbox oil temperature drops significantly, the kinematic viscosity of the lubricating oil increases sharply, leading to a severe reduction in fluidity. Under these conditions, starting or frequently adjusting the wind turbine will introduce excessive mechanical resistance, causing a sudden increase in starting torque, which may lead to generator overload or even catastrophic mechanical failure. Therefore, the proposed strategy prioritizes the average power distribution mechanism in winter (low-temperature scenario: when the ambient temperature is below the preset temperature of 0 degrees Celsius) to maintain stable gearbox oil temperature and ensure equipment safety. Conversely, under medium- and high-temperature conditions (when the ambient temperature is not lower than the preset temperature of 0 degrees Celsius), a multi-dimensional optimization mechanism is implemented to achieve a more refined optimal power distribution.

[0043] Furthermore, in low-temperature scenarios, power allocation using an average power distribution mechanism includes: To mitigate the risk of mechanical damage induced by extremely low gearbox oil temperatures in cold weather, the winter active power distribution strategy aims to minimize frequent start-stop cycles of individual units while ensuring uniform load distribution within the cluster. Under this premise, the target active power command issued by the grid is defined as... The theoretical total available power of the entire wind farm is denoted as: The available polymerization power of the prototype machine is The minimum non-zero available power among all prototype machines is According to the target instructions Relative to the size of these capacity boundaries, the control logic is divided into five operating scenarios by the system, and corresponding differentiated power allocation instructions are triggered.

[0044] (1) Zero-schedule operating condition: : When the grid-side dispatch command is zero or wind power grid connection is suspended, all wind turbines in the wind farm are ordered to shut down or enter an idling state. Therefore, the active power distribution of all wind turbines is as follows: .

[0045] (2) Ultra-low load condition: : When the grid demand is extremely low and below the minimum real-time available capacity of any single prototype unit, in order to avoid frequent start-ups and shutdowns, only the prototype unit with the lowest available capacity is scheduled to complete the task, while the remaining wind turbines maintain zero active power output.

[0046] (3) Low load condition: : In this scenario, power generation is entirely handled by the prototype turbines. If absolute equal distribution were implemented, commands exceeding the aerodynamic limits would be issued to some low-wind-speed turbines. To ensure uniform mechanical wear and prevent physical limit exceedances, a dynamic equal distribution mechanism with saturation constraints is triggered only between the prototype turbines. At the current wind speed, the turbine's maximum aerodynamic power (theoretical available power generation potential) is... The calculation is as follows: (11); Where R is the impeller radius, V is the wind speed, β is the blade pitch angle, and λ represents the tip speed ratio (λ=ω*R / V, where ω is the rotor speed). This is the wind energy utilization coefficient.

[0047] set up Let M represent the target power to be allocated, and M represent the number of wind turbines participating in operation. In each iteration, the algorithm calculates the average baseline quota for the current set. For any wind turbine j, its output is limited by a lower limit (the minimum grid-connected power at the cut-in wind speed). ) and upper limit (its available potential) Constraints. If the theoretical available power generation potential of a certain wind turbine satisfies... If the value is less than or equal to saturation, then the active power command is considered to be constrained to the physical limit (the power value that can be allocated to the j-th wind turbine). The saturated power is then deducted from the total target, and the saturated turbine is removed from the active set. The remaining unallocated power continues to be iteratively and evenly distributed across the updated subset of unsaturated turbines until all power quotas are fully allocated.

[0048] (4) Medium to high load conditions: : Within this operating range, the prototype machine is prioritized to operate at full load, outputting its theoretical maximum available power for polymerization. Subsequently, the system calculates the dynamic power residual. And use the above-mentioned dynamic uniform distribution mechanism to distribute it evenly to non-sample machines.

[0049] (5) Capacity-constrained operating conditions: : When the grid-side dispatch instructions exceed the total available power generation potential of the wind farm, all wind turbines are dispatched to generate power at their current maximum available capacity.

[0050] Furthermore, in medium- and high-temperature scenarios, a multi-dimensional optimization mechanism for power allocation includes the aforementioned dynamic uniform allocation strategy, which can effectively maintain stable gearbox oil temperature and reduce mechanical wear under winter conditions. However, at medium- and high ambient temperatures, strictly adopting this uniform method ignores the spatial heterogeneity of wind speeds and differences in aerodynamic efficiency across different regions, inevitably leading to suboptimal wind energy utilization. Therefore, in higher-temperature scenarios, a more refined and differentiated active power allocation strategy must be implemented.

[0051] The high-temperature strategy retains the principle of prioritizing prototype wind turbines under low grid dispatch commands. However, under high dispatch commands, a more complex optimization mechanism is activated. Specifically, when the grid target command falls between the aggregated available capacity of the prototype wind turbines and the total available potential of the wind farm (…), the optimization mechanism will be more complex. When the power distribution of non-model turbines is not uniform, it strictly follows the goal of maximizing the overall operating efficiency and comprehensive wind energy capture capability of the entire wind farm.

[0052] Under this premise, this embodiment constructs a dynamic comprehensive evaluation index. This is used to guide the greedy sorting and power allocation of non-sample machines, thereby establishing an optimal active power scheduling strategy. The final score of the non-sample machine at time t. It can be broken down into four core mathematical components: (12).

[0053] (1) Construction of basic prediction reward: To ensure the forward-looking nature of the allocation strategy, the power dispatching of non-sample units should not only rely on their real-time available capacity but also incorporate future power generation potential predictions. Therefore, this embodiment introduces a basic prediction score. By combining the current available power with the predicted power over the subsequent t+H steps, this index effectively characterizes the spatiotemporal evolution of wind resources. Considering that prediction errors propagate and amplify over time, a confidence decay coefficient is introduced. This is used to quantify and mitigate the impact of long-term forecast uncertainty on current decision-making. The calculation formula is defined as follows: (13); in, This is the weighting coefficient for the current wind speed; and Let these represent the available power at the current moment and at the τ-th future step, respectively. The wind speed was calculated based on the predicted wind speed and formula 11.

[0054] (2) Dynamic comprehensive evaluation under multi-state constraints: Scoring systems relying solely on prediction rewards have significant limitations. When local wind resources remain consistently favorable, the algorithm forces corresponding wind turbines to operate at full capacity for extended periods. This uninterrupted power generation rapidly exacerbates accumulated fatigue damage, significantly increasing the risk of mechanical failure. Furthermore, under the constraint of meeting the overall grid dispatch instructions, high-scoring wind turbines will prioritize full-capacity tasks, while the remaining "tail-end regulation" quota will be continuously allocated to a specific subset of low-scoring wind turbines. If the prediction curve is relatively stable, wind turbines that continuously undertake these tail-end regulation tasks will experience excessive mechanical stress due to frequent operational adjustments. To address these vulnerabilities, this embodiment constructs a dynamic comprehensive evaluation index that incorporates a state feedback penalty mechanism on top of the prediction reward: (14); in , and These represent the hysteresis reward, tail adjustment penalty, and macroscopic energy fatigue penalty of wind turbine i at time t, respectively. , and This represents the corresponding weighting coefficient. Furthermore, Record the cumulative number of tail-end adjustment tasks performed by wind turbine i. This indicates the maximum number of regulation tasks recorded throughout the entire wind farm. Similarly, This indicates the cumulative power generation of a single wind turbine. Defined as the maximum cumulative power generation of a single wind turbine within a wind farm.

[0055] (3) Greedy active power allocation based on state update: After completing the multi-state constraint calculation and obtaining the comprehensive score matrix of the non-template machine, the control system proceeds according to... The values ​​are sorted in descending order. Subsequently, the control center deploys a greedy algorithm to prioritize the extracted power residuals. The highest-scoring non-model unit i is allocated power. The allocated power quota is strictly constrained within the physical limits of the wind turbine. (15); Subsequently, the system dynamically updates the remaining unallocated power. The system then issues active power commands sequentially according to the sorted list. This iterative process continues until the dynamic residual is completely exhausted.

[0056] In summary, the active power allocation strategy proposed in this embodiment, by integrating wind turbine clustering, high-fidelity spatiotemporal prediction, and dynamic allocation under multi-state constraints, not only strictly satisfies real-time grid dispatching instructions at the macro level, but also achieves a globally optimal trade-off between wind energy capture efficiency and mechanical fatigue life at the micro level.

[0057] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage, characterized in that, include: Several prototype turbines were selected in the target wind farm as benchmarks, and non-prototype turbines were clustered to obtain several turbine clusters; The wind speed of the prototype machine is predicted by a prediction model based on spatiotemporal feature extraction, and the wind speed prediction result of the prototype machine is obtained. The wind speed prediction results of the prototype machine are input into the intra-cluster spatial mapping model to obtain the wind speed prediction results of the non-prototype machine. The intra-cluster spatial mapping model constructs the feature mapping relationship between the prototype machine and the non-prototype machine through a ridge regression model and is obtained by training with the historical wind speed of the prototype machine. Based on the target active power command issued by the power grid, ambient temperature, real-time operating status, and wind speed prediction results, the average power allocation mechanism is used for power allocation when the ambient temperature is lower than the preset temperature, and the multi-dimensional optimization mechanism is used for power allocation when the ambient temperature is not lower than the preset temperature.

2. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 1, characterized in that, Clustering of non-sample machines includes: The Pearson correlation coefficient is used to quantify the synchronicity of wind speed evolution trends between the prototype machine and the non-prototype machine. The correlation coefficient is calculated, and the maximum correlation criterion is used to perform clustering based on the correlation coefficient. Alternatively, the standardized time series of the prototype machine can be used as the initial cluster center set of the K-Means algorithm. By minimizing the intra-cluster squared error and assigning the time series feature vector of each non-prototype machine to the centroid with the shortest Euclidean distance, the centroid position is dynamically updated according to the current cluster members until the objective function converges, and the final wind turbine clustering result is obtained.

3. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 1, characterized in that, The prediction model based on spatiotemporal feature extraction predicts the wind speed of the prototype machine, including: For the wind turbine cluster, a low-rank parameterized adjacency matrix is ​​used to represent the spatial topology; Based on the spatial topology and the historical wind speed observation feature matrix of the prototype wind turbine in the wind turbine cluster, a graph convolutional network is used to aggregate the node features within the cluster to obtain spatial enhancement features. The spatial enhancement features are input into the time series decomposition module to extract time features, and the time features are decomposed into long-term trend components and seasonal fluctuation components using a moving average mechanism. The decomposed components are projected onto the future preset step prediction time domain of the prototype machine through a linear layer, and the projected components are superimposed to obtain the wind speed prediction result of the prototype machine.

4. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 1, characterized in that, Training the ridge regression model using historical wind speeds from a prototype machine involves: solving for the optimal mapping weight vector by minimizing the penalized sum of squared residuals. β : ； in, Let X be the objective function, representing the historical wind speed observation feature matrix obtained from the prototype machine, and y be the actual wind speed vector of the target non-prototype machine. This represents the squared L2 norm of the weight vector β. It is the regularization coefficient that controls the complexity of the model.

5. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 1, characterized in that, When the ambient temperature is lower than the preset temperature, the average power distribution mechanism is used for power allocation, including: when At this time, it is the zero-dispatch operating condition, and the active power allocation of all wind turbines is 0. Target active power command: when At that time, under ultra-low load conditions, only the pilot unit with the lowest available capacity was dispatched to complete the task, while the remaining wind turbines maintained zero active power output. This represents the minimum non-zero available power among all the prototype machines. when At that time, under low-load conditions, a dynamic uniform distribution mechanism with saturation constraints is triggered among the prototype machines, wherein... The available power for polymerization of the prototype machine; when At that time, under medium to high load conditions, the prototype machine is prioritized to operate at full load, outputting aggregated usable power, and the dynamic power residual is calculated. And evenly distributed to non-sample machines, among which, This represents the theoretical total available power of the entire wind farm. when During this period, which is a capacity-limited operating condition, all wind turbines are dispatched to generate electricity at their current maximum available capacity.

6. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 5, characterized in that, The dynamic uniform allocation mechanism with saturation constraints triggered between prototypes includes: Calculate the maximum aerodynamic power of the wind turbine at the current wind speed; Calculate the average baseline quota for the current set based on the target power to be allocated and the number of wind turbines participating in operation; If the maximum aerodynamic power is less than the average baseline quota, the fan corresponding to the maximum aerodynamic power is recorded as a saturated fan, the maximum dynamic power is deducted from the target active power, and the saturated fan is removed from the active set. The remaining unallocated power continues to be evenly distributed and iterated in the updated unsaturated fan subset until all power quotas are fully allocated.

7. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 1, characterized in that, When the ambient temperature is not lower than the preset temperature, a multi-dimensional optimization mechanism is used for power allocation, including: when... At that time, a dynamic comprehensive evaluation index under multi-state constraints is constructed for non-template machines, and greedy active power allocation is performed based on the dynamic comprehensive evaluation index, wherein, The target active power command, The available power for polymerization of the prototype machine. This represents the theoretical total available power of the entire wind farm.

8. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 7, characterized in that, The dynamic comprehensive evaluation index The calculation method is as follows: ； ； in, Based on the predicted score, , and Let represent the hysteresis reward, tail adjustment penalty, and macroscopic energy fatigue penalty of wind turbine i at time t, respectively. The available power at the current moment. To achieve the minimum grid-connected power at the cut-off wind speed, This represents the theoretically available power generation potential.

9. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 8, characterized in that, The calculation method for the basic prediction score is as follows: ； in, The weighting factor for the current wind speed. and Let these represent the available power at the current moment and at the τ-th future step, respectively. Calculated based on wind speed prediction results Here, H is the confidence decay coefficient, and H is the total number of future steps.

10. The method for active power allocation in a real wind farm considering wind speed prediction and fatigue damage as described in claim 7, characterized in that, Greedy active power allocation based on the aforementioned dynamic comprehensive evaluation index includes: The dynamic comprehensive evaluation indicators are sorted in descending order, and the dynamic power residual is allocated to the non-sample unit with the highest score. During the allocation process, the active power allocation of the wind turbines is as follows: , For dynamic power residual, Theoretically available power generation potential; The remaining unallocated power is dynamically updated, and active power instructions are issued sequentially according to the sorted list until the dynamic power residual is completely exhausted.

Images