Wind farm power prediction method and device based on multi-dimensional redundancy and cross-domain collaboration

Abstract

This specification provides a wind farm power prediction method and device based on multi-dimensional redundancy and cross-domain collaboration. The method ensures the continuity and reliability of data acquisition under extreme environments by constructing an active redundancy sensing and emergency back-propagation mechanism, providing a solid data foundation for prediction. It employs a scenario-triggered dynamic coupling prediction model, achieving adaptive fusion of physical laws and data-driven approaches, significantly improving prediction accuracy and stability under different extreme weather conditions. By establishing a cross-domain linkage correction and reliable collaborative architecture between power, operation and maintenance, and the power grid, a decision optimization chain from accurate prediction to closed-loop execution is formed, improving overall operational efficiency and grid adaptability. It innovatively proposes a regional cluster collaborative prediction and dynamic power allocation method under privacy protection, realizing large-scale precise power control from single wind farms to regional clusters, providing key technical support for the reliable consumption of wind power under the new power system.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of power technology, and in particular to a wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration. Background Technology

[0002] Current methods for predicting wind farm power in extreme weather lack effective active redundancy mechanisms at the perception level to address the collective failure of multiple sensors, resulting in unreliable data sources. At the model level, the use of fixed-weight fusion of physical and artificial intelligence models makes it difficult to dynamically adapt to the dynamic characteristics of different extreme scenarios such as typhoons, icing, and blizzards with snowmelt. At the application level, prediction, operation and maintenance, and grid dispatch are isolated from each other, failing to form effective cross-domain collaboration and closed-loop correction. At the same time, there is a lack of systematic solutions that can achieve centralized power prediction and optimized allocation of regional wind farm clusters while protecting the data privacy of each site, which restricts the safe and economical operation of the power grid under large-scale wind power grid connection.

[0003] Therefore, a better solution is urgently needed. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration. One or more embodiments of this specification also relate to a wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration is provided, including: Deploy heterogeneous sensing devices to collect data and activate an emergency reverse mechanism to obtain key data when sensing fails. Preprocess and label the collected and reversed data to distinguish extreme scenarios. The preprocessed data is input into the scenario-triggered physical-AI dynamic coupling prediction model. The model automatically adjusts the weight ratio of the physical model and the AI ​​model according to the working condition label. Through physical feature extraction, time series prediction and physical constraint loss, the predicted power value of a single wind farm is output. Based on the cross-domain linkage correction process triggered by power deviation and operation and maintenance needs, the power prediction value of a single wind farm is corrected, and operation and maintenance adjustments and grid dispatch are linked. The correction process iterates the model parameters through a federated learning framework and stores the evidence through a blockchain architecture. Multiple wind farms are divided into clusters. A regional centralized prediction model is constructed by coordinating data within the clusters through federated learning. The regional centralized power prediction value is output, and a dynamic allocation strategy is used to distribute the regional centralized power prediction value to each wind farm within the cluster. The scene-triggered physics-AI dynamic coupling prediction model and the regional centralized prediction model were trained and validated using meta-learning algorithms and a dataset containing extreme scenarios.

[0006] In one possible implementation, heterogeneous sensing devices include millimeter-wave radar, fiber optic strain sensors, ultrasonic snow depth sensors, fixed-wing weather stations, and wind profiler radar. The ultrasonic snow depth sensor integrates a heating-ventilation dual-control anti-condensation module, the millimeter-wave radar is equipped with a self-heating insulation sleeve, and the fixed-wing meteorological station is deployed around the wind farm and distributed at preset intervals to collect the wake wind speed between the wind turbines.

[0007] In one possible implementation, the emergency reverse-engineering mechanism includes: The wind speed is calculated by inversely based on the real-time torque and speed of the wind turbine generator; The real-time current of the blade pitch motor is used to inversely calculate the blade ice thickness or snow depth through a preset current-load mapping model. Spatial consistency verification is performed between the ice thickness or snow depth on the blades and the wind speed and the sensing data of adjacent wind turbines. If the deviation exceeds the threshold, a weighted voting mechanism is activated to determine the final data.

[0008] In one possible implementation, the weight adjustment rules in the scene-triggered physics-AI dynamic coupling prediction model are driven by the scene feature vector; The scene feature vector includes wind speed fluctuation rate, ice thickness growth rate, and temperature change rate. In the typhoon scene, the physical model has a weight of 70% and the AI ​​model has a weight of 30%. In the ice accumulation scene, the physical model has a weight of 40% and the AI ​​model has a weight of 60%. In the blizzard-snowmelt transition scene, the physical model has a weight of 60% and the AI ​​model has a weight of 40%.

[0009] In one possible implementation, physical feature extraction includes: Theoretical power correction and wake-terrain coupling calculations are performed. Theoretical power correction is based on the wind turbine aerodynamic formula and introduces correction terms for air density in cases of icing, blizzards, and high altitudes. Wake-terrain coupling calculations correct the basic wake attenuation coefficient by introducing a terrain slope factor. The physical constraint loss is calculated based on a hybrid loss function that includes prediction error terms, physical theoretical upper limit deviation terms, and wind energy capture coefficient deviation terms. The weights of each term in the hybrid loss function are dynamically allocated according to the extreme scenario type.

[0010] In one possible implementation, the triggering condition for the cross-domain collaborative correction process is: In typhoon or blizzard scenarios, the power deviation rate exceeds the first threshold and the predicted future power fluctuation rate exceeds the second threshold. In ice accumulation or snow melting scenarios, the power deviation rate exceeds the third threshold and the blade fatigue damage value exceeds the preset threshold. The correction process in the blizzard-snowmelt scenario also activates the secondary icing prediction model, which predicts the secondary icing thickness based on ambient temperature and relative humidity.

[0011] In one possible implementation, the dynamic allocation strategy includes: The base power quota is allocated according to the prediction accuracy of each wind farm's single station. The allocation ratio will be adjusted according to the equipment health score of each wind farm. Verification and correction are performed based on power grid line capacity constraints and power grid stability constraints.

[0012] According to a second aspect of the embodiments of this specification, a wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration is provided, comprising: The data acquisition module is configured to deploy heterogeneous sensing devices for data acquisition and to activate an emergency reverse mechanism to obtain key data when sensing fails. The acquired and reversed data are preprocessed and labeled with operating conditions to distinguish extreme scenarios. The coupled prediction module is configured to input preprocessed data into a scenario-triggered physical-AI dynamic coupled prediction model. The model automatically adjusts the weight ratio of the physical model and the AI ​​model according to the working condition label, and outputs the predicted power value of a single wind farm through physical feature extraction, time series prediction and physical constraint loss. The linkage correction module is configured to trigger a cross-domain linkage correction process based on power deviation and operation and maintenance requirements, correct the power prediction value of a single wind farm, and link operation and maintenance adjustments and grid dispatch. The correction process iterates model parameters through a federated learning framework and stores evidence through a blockchain architecture. The collaborative optimization module is configured to divide multiple wind farms into clusters, build a regional centralized prediction model by coordinating data within the clusters through federated learning, output the regional centralized power prediction value, and use a dynamic allocation strategy to distribute the regional centralized power prediction value to each wind farm within the cluster. The model training module is configured to train and validate scene-triggered physics-AI dynamic coupling prediction models and regional centralized prediction models using meta-learning algorithms and datasets containing extreme scenarios.

[0013] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration.

[0014] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above.

[0015] According to a fifth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, the computer is instructed to perform the steps of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above.

[0016] This specification provides a wind farm power prediction method and device based on multi-dimensional redundancy and cross-domain collaboration. The method ensures the continuity and reliability of data acquisition under extreme environments by constructing an active redundancy sensing and emergency back-propagation mechanism, providing a solid data foundation for prediction. It employs a scenario-triggered dynamic coupling prediction model, achieving adaptive fusion of physical laws and data-driven approaches, significantly improving prediction accuracy and stability under different extreme weather conditions. By establishing a cross-domain linkage correction and reliable collaborative architecture between power, operation and maintenance, and the power grid, a decision optimization chain from accurate prediction to closed-loop execution is formed, improving overall operational efficiency and grid adaptability. It innovatively proposes a privacy-protected regional cluster collaborative prediction and dynamic power allocation method, realizing large-scale precise power control from single wind farms to regional clusters, providing key technical support for the reliable absorption of wind power under the new power system. Attached Figure Description

[0017] Figure 1 This is a flowchart of a wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration, provided in one embodiment of this specification. Figure 2 This is a comparison chart of prediction results for a wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration, provided in one embodiment of this specification. Figure 3 This is a schematic diagram of the structure of a wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration, provided in one embodiment of this specification. Figure 4 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0018] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0019] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0020] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0021] This specification provides a wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration. This specification also relates to a wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0022] See Figure 1 , Figure 1 A flowchart is shown of a wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration according to an embodiment of this specification, which specifically includes the following steps.

[0023] Step 101: Deploy heterogeneous sensing devices to collect data and activate the emergency reverse mechanism to obtain key data when sensing fails. Preprocess and label the collected and reversed data to distinguish extreme scenarios.

[0024] In one possible implementation, the heterogeneous sensing devices include millimeter-wave radar, fiber optic strain sensor, ultrasonic snow depth sensor, fixed-wing weather station and wind profiler radar; the ultrasonic snow depth sensor integrates a heating-ventilation dual-control anti-condensation module, the millimeter-wave radar is equipped with a self-heating insulation sleeve, and the fixed-wing weather station is deployed around the wind farm and distributed at preset intervals to collect the wake wind speed between the wind turbines.

[0025] In practical applications, five types of heterogeneous sensing devices across physical quantities are deployed to address extreme environmental characteristics. Specific parameters and adaptation functions are shown in the table below: Table 1

[0026] In one possible implementation, the emergency reverse propagation mechanism includes: reverse propagating the wind speed based on the real-time torque and speed of the wind turbine generator; reverse propagating the blade ice thickness or snow depth based on the real-time current of the blade pitch motor and through a preset current-load mapping model; and performing spatial consistency verification between the blade ice thickness or snow depth and the wind speed and the sensing data of adjacent wind turbines. If the deviation exceeds a threshold, a weighted voting mechanism is initiated to determine the final data.

[0027] In practical applications, when three or more types of equipment fail simultaneously (e.g., freezing at -35℃), the system automatically activates the emergency reverse propagation module. This module indirectly acquires key environmental data through fan operating parameters to ensure continuous sensing. Specifically, this includes: Wind speed inverse propagation: based on real-time generator torque T (in N) The calculation is based on the relationship between m and rotational speed n (in rpm), combined with the core aerodynamic formulas: in Real-time atmospheric density (unit: kg / m³) collected by wind profiler radar. R is the swept area of ​​the fan (R is the radius of the fan, in meters). The wind energy capture coefficient is corrected by the snow-ice thickness coupling model, and the back-calculated wind speed error is ≤0.8m / s; Ice / snow load inversion: Based on the real-time current I (unit A) of the blade pitch motor, the "current-load" mapping model trained through 1000+ extreme working condition experiments calculates that: in the icing scenario, every 10A increase in current corresponds to an increase of 1.2mm in blade ice thickness (inversion accuracy ±0.1mm); in the blizzard scenario, every 10A increase in current corresponds to an increase of 5mm in blade snow depth (inversion accuracy ±0.5mm). Spatial consistency verification: The back-inferred data is compared with the normal perception data of the three adjacent wind turbines. If the deviation is >5%, the "weighted voting mechanism" is activated (weight = wind turbine health score × distance attenuation coefficient, and the health score is a quantitative index of 0-100 points generated based on historical fault data) to ensure that the reliability of emergency perception is ≥95%.

[0028] Furthermore, this solution also includes data preprocessing and working condition labeling, as detailed below.

[0029] Outlier Identification: Abandoning the traditional 3σ criterion, based on the wind turbine aerodynamic equations To determine the reasonableness of the data, and to avoid the wind speed jumps caused by typhoon turbulence being misjudged as abnormal data; Time synchronization: Through the digital twin time synchronization module, multi-source heterogeneous data such as millimeter-wave radar (5Hz), fiber optic sensor (10Hz), and meteorological satellite (10 minutes) are uniformly aligned to a 1-minute time scale; Operating conditions labeling: Construct a three-dimensional quantitative labeling system of "wind speed level - ice thickness - snow depth" to clearly distinguish three types of extreme scenarios: typhoon (wind speed ≥ level 12), ice accumulation (ice thickness ≥ 1 mm), and blizzard-snow melt transition (temperature 0℃±2℃ and relative humidity ≥ 85%).

[0030] Step 102: Input the preprocessed data into the scenario-triggered physical-AI dynamic coupling prediction model. The model automatically adjusts the weight ratio of the physical model and the AI ​​model according to the working condition label. It outputs the predicted power value of a single wind farm through physical feature extraction, time series prediction and physical constraint loss.

[0031] In one possible implementation, the weight adjustment rule in the scene-triggered physics-AI dynamic coupling prediction model is driven by the scene feature vector; the scene feature vector includes wind speed fluctuation rate, ice thickness growth rate and temperature change rate; in the typhoon scenario, the weight of the physical model is 70% and the weight of the AI ​​model is 30%; in the ice accumulation scenario, the weight of the physical model is 40% and the weight of the AI ​​model is 60%; in the blizzard-snowmelt transition scenario, the weight of the physical model is 60% and the weight of the AI ​​model is 40%.

[0032] In practical applications, this scheme automatically adjusts the model weight ratio based on the dynamic characteristics of extreme weather, as detailed in the table below: Table 2

[0033] Weight adjustment is driven by "scene feature vectors", which include three types of quantitative indicators: wind speed fluctuation rate (level 10 threshold), ice thickness growth rate (level 10 threshold), and temperature change rate (level 10 threshold). Once the threshold is triggered, the preset weight matrix is ​​automatically called, with a response time of ≤1 second, ensuring that the model can quickly adapt to scene changes.

[0034] In one possible implementation, physical feature extraction includes: performing theoretical power correction and wake-terrain coupling calculation. The theoretical power correction is based on the wind turbine aerodynamic formula and introduces correction terms for icing, blizzards, and high-altitude air density. The wake-terrain coupling calculation corrects the basic wake attenuation coefficient by introducing a terrain slope factor. The physical constraint loss is calculated based on a hybrid loss function that includes a prediction error term, a physical theoretical upper limit deviation term, and a wind energy capture coefficient deviation term. The weights of each term in the hybrid loss function are dynamically allocated according to the extreme scenario type.

[0035] In practical applications, the physical feature extraction layer includes: a theoretical power correction model: a multi-dimensional correction formula is constructed to address the impact of different extreme scenarios on the wind turbine's wind energy capture coefficient. Basic formulas for common scenarios: ,in ; Ice accumulation correction formula: ( (Ice thickness of the blade, in mm). Blizzard corrected the formula: ( Snow depth on the leaves (in mm). High-altitude air density correction: Real-time atmospheric density values ​​(in a blizzard scenario at an altitude of 3500m) acquired using wind profiler radar =0.72kg / m³). Wake-Terrain Coupled Model: First Introduction of Terrain Slope Factor (Unit: °) Corrected wake attenuation coefficient, the formula is: in Basic wake attenuation coefficient (typhoon / icing scenario) =0.2, Blizzard scenario =0.3), which solves the technical defect of existing technologies that ignore the impact of terrain on wake.

[0036] Furthermore, the Transformer time series prediction layer includes: Input features: 31-dimensional multi-source heterogeneous features (18 conventional perception features + 5 emergency backpropagation features + 8 scene label features); Model parameters: 8 layers each for encoder and decoder, 384 hidden layer dimensions, learning rate 0.0003 (pre-training stage) / 0.0001 (fine-tuning stage), batch size = 48; Attention mechanism optimization: Spatial attention is based on the "wind turbine spatial correlation matrix" generated by the wake-terrain coupling model to assign weights (the weight values ​​are negatively correlated with the distance between the two wind turbines and the terrain slope); Temporal attention adopts a dynamic window design: a 30-minute time window is used in typhoon / blizzard scenarios (to adapt to rapid power fluctuations), and a 1-hour time window is used in icing scenarios (to adapt to steady-state changes).

[0037] Furthermore, the dynamic weighted physical constraint loss layer includes: A dynamic weighted hybrid loss function is constructed to prevent the predicted value from exceeding the physical limits of the wind turbine. The formula is as follows:

[0038] in Dynamically assigned based on scenario (Typhoon / Ice Accumulation Scenario) Blizzard - Snowmelt Scene =0.55, =0.3, =0.15), ensuring that the prediction results conform to historical patterns and do not violate physical principles.

[0039] Step 103: Based on power deviation and operation and maintenance requirements, trigger a cross-domain linkage correction process to correct the power prediction value of a single wind farm, and link operation and maintenance adjustments and grid dispatch. The correction process iterates model parameters through a federated learning framework and stores evidence through a blockchain architecture.

[0040] In one possible implementation, the triggering conditions for the cross-domain linkage correction process are as follows: in typhoon or blizzard scenarios, the power deviation rate exceeds the first threshold and the predicted future power fluctuation rate exceeds the second threshold; in icing or snow melting scenarios, the power deviation rate exceeds the third threshold and the blade fatigue damage value exceeds the preset threshold; in blizzard-snow melting scenarios, the correction process also activates the secondary icing prediction model, which predicts the secondary icing thickness based on ambient temperature and relative humidity.

[0041] In practical applications, the cross-domain linkage correction of power, operation and maintenance, and power grid is described in detail below.

[0042] Dual-drive correction trigger conditions: The correction process is triggered based on a dual-dimensional approach of "technical deviation + operational needs," ensuring the necessity and timeliness of the correction. Typhoon / blizzard scenario: Power deviation rate Furthermore, the predicted power volatility for the next 2 hours is >15%; Ice / snow melting scenario: Power deviation rate Furthermore, the blade fatigue damage value inferred by the fiber optic strain sensor is greater than the preset threshold (based on the quantified setting of the wind turbine's 20-year design life).

[0043] The cross-domain collaborative correction process is described in detail below.

[0044] The technical parameter revisions include: Standard correction: Recalculate the wake wind speed using the wake-terrain coupling model and update the spatial attention weights of the Transformer time series prediction layer; Blizzard-Snowmelt Specific Correction: The world's first secondary icing prediction model, fitted based on 1200 hours of measured data from a high-altitude wind farm:

[0045] in The prediction error for the secondary icing thickness after snow melting (in mm) is ≤0.3 mm, solving the problem of sudden increase in power deviation in high-altitude composite scenarios.

[0046] Operation and maintenance resource coordination includes: Send an "advance on-site command" to the explosion-proof inspection robot within 3km upstream of the wind farm, and adjust the inspection route based on the "power fluctuation-terrain risk" model (avoiding high-risk areas with wind speed > 20m / s). Push proactive protection suggestions for equipment to the operation and maintenance center (such as activating the blade de-icing device in advance during blizzard-snow melting scenarios), with an operation and maintenance response delay of ≤5 minutes.

[0047] Power grid dispatch adaptation includes: The system pushes "Operation and Maintenance Window Suggestions" to the power grid dispatching side, recommending that power grid load adjustments be carried out during periods when power fluctuation rate is less than 8%. The prediction correction report (including the cause of the deviation, the correction logic, and the data source) is uploaded synchronously, and the data interaction complies with the IEC61850 international communication protocol.

[0048] Trusted evidence storage and parameter iteration include: Blockchain-based evidence storage: Adopting a consortium blockchain architecture, the nodes include wind farms, power grid companies, and operation and maintenance enterprises, generating a full-link associated certificate of "prediction correction report - operation and maintenance execution record - power grid adjustment feedback", ensuring that the data is tamper-proof and traceable; Federated learning: It adopts an architecture of "edge as the main and center as the auxiliary". The local edge computing server of the wind farm is based on the parameter of the data iteration model, with each iteration ≤100 times. The aggregated parameters are synchronized to the regional center server monthly through homomorphic encryption technology, reducing the amount of data transmission by 70% and ensuring business privacy.

[0049] Step 104: Divide multiple wind farms into clusters, construct a regional centralized prediction model by coordinating data within the clusters through federated learning, output the regional centralized power prediction value, and use a dynamic allocation strategy to distribute the regional centralized power prediction value to each wind farm within the cluster.

[0050] In one possible implementation, the dynamic allocation strategy includes: allocating base power quotas based on the prediction accuracy of each wind farm's single site; adjusting the allocation ratio based on the equipment health score of each wind farm; and performing verification and correction based on grid line capacity constraints and grid stability constraints.

[0051] In practical applications, the regional cluster division is specifically described as follows.

[0052] Clustering is based on four core principles to ensure the feasibility of collaboration within the cluster and the consistency of predictions: Consistent grid access point: All wind farms within the cluster are connected to the same 220kV / 500kV step-up substation to avoid cross-grid dispatch conflicts; Geographical distance adaptation: The maximum distance between stations is ≤500km to ensure the spatial correlation of meteorological fields (correlation coefficient ≥0.8). Meteorological fields share the same origin: the prevailing wind direction and topographic type are consistent within the cluster, and the patterns of extreme weather impacts are similar; Operation and maintenance management adaptation: Total installed capacity of the cluster is 200-2000MW (including 5-15 wind farms), adapting to the existing operation and maintenance management radius and resource configuration.

[0053] Furthermore, privacy-preserving data collaboration includes: We adopt a federated learning horizontal federated architecture to build a collaborative mechanism of "data not leaving the factory, models being trained together": Each site trains its own model based on local data and only uploads intermediate model parameters (weights, gradients) to the regional central server, without disclosing raw data (such as turbine efficiency and operation and maintenance costs). Intermediate parameters are transmitted using homomorphic encryption technology. The regional central server aggregates the parameters to generate a cluster-shared model, which is then distributed to each site to update the local model. A data contribution incentive mechanism is established, which allocates model usage weights based on the data quality (completeness ≥95%, accuracy ≥98%) and coverage of extreme scenarios at each site to ensure collaborative fairness.

[0054] Furthermore, it also includes the construction of regional centralized prediction models, as detailed below.

[0055] Based on the single-site model, a regional-level Transformer centralized prediction model is built to enhance the cluster collaboration characteristics: Input feature expansion: Added cluster-level features (meteorological correlation between stations, regional load forecast, power grid line capacity constraints, and average health of cluster equipment), increasing the total feature dimensions to 42. Model structure optimization: The hidden layer dimension was increased from 384 to 512, and the number of encoder / decoder layers was increased to 10 each. Attention mechanism upgrade: A new "cluster spatial correlation matrix" has been added, which allocates spatial attention weights based on the distance between stations, terrain connectivity, and meteorological transmission paths to enhance collaborative prediction between stations; Extreme scenario adaptation: To address the propagation characteristics of cluster-level extreme weather (such as the gradient attenuation of typhoons from the coast to the inland areas and the continuous coverage of blizzards in mountainous areas), a "scenario propagation coefficient" correction term is added to improve the prediction accuracy of cluster extreme scenarios.

[0056] Furthermore, the dynamic power allocation strategy employs a hierarchical allocation approach, precisely decomposing the concentrated power in the region to each power station, balancing efficiency and safety. First layer: Basic power allocation (accounting for 60%), the quota is allocated according to the prediction accuracy of each station (MAPE≤8.7%), the higher the prediction accuracy, the greater the allocation weight; The second layer: equipment load capacity adaptation (accounting for 20%), the allocation ratio is adjusted based on the equipment health score (≥80 points are given priority) and remaining lifespan (>10 years are given priority) to avoid overloading of heavily loaded equipment; The third layer: power grid safety verification (accounting for 20%), which makes the final correction based on the power grid line capacity (not exceeding 90% of the rated capacity), frequency stability constraints (fluctuation ≤ ±0.2Hz), and voltage deviation (≤ ±5%). Dynamic adjustment mechanism: Every 15 minutes, the allocation scheme is iterated based on real-time power deviation (≤2%), weather changes (triggered by wind speed fluctuation rate >10%), and equipment status updates (the allocation quota is reduced when a fault warning is issued) to ensure that the allocation dynamically adapts to the actual working conditions.

[0057] Step 105: Train and validate the scene-triggered physics-AI dynamic coupling prediction model and the regional centralized prediction model using a meta-learning algorithm and a dataset containing extreme scenarios.

[0058] In practical applications, model training and validation are described in detail below.

[0059] Construction of a dedicated dataset for extreme weather: Two years of historical data were collected from three types of wind farms in typical complex environments and their corresponding regional clusters to construct a special dataset for extreme weather events, as detailed in the table below: Table 3

[0060] Dataset partitioning: Training set: Validation set: Test set = 7:2:1, and the data format strictly follows the three-dimensional labeling system in step 1.3.

[0061] Furthermore, this scheme employs a three-stage training strategy: Pre-training phase: Train the base model on regular weather data, with a convergence criterion of MAPE ≤ 8% for power prediction in regular scenarios from 0 to 4 hours; Extreme Condition Enhancement Training Phase: The model is fine-tuned using a special dataset for extreme weather conditions, with a focus on optimizing physical constraint losses. Transfer learning phase: The MAML meta-learning algorithm is used to transfer the model parameters trained in high-altitude wind farms and regional clusters to new wind farms / clusters of the same type, reducing the training data requirements of new wind farms by 60% and new clusters by 75%, and shortening the deployment cycle by 20 days.

[0062] To further illustrate this scheme, the verification indicators and experimental results are given below.

[0063] On-site verification was conducted at a 3500m high-altitude blizzard-melting wind farm (99MW installed capacity, 33 3MW permanent magnet direct-drive wind turbines) and the corresponding regional cluster (including 8 wind farms with a total installed capacity of 1000MW). The results are shown in the table below: Table 4

[0064] In summary, this application pioneers a multi-sensor failure proactive emergency reverse mechanism, reducing the probability of perception interruption in extreme environments from 18% in existing technologies to 0.3% (0.2% for clusters); the sensor's "self-heating + anti-condensation" proactive adaptation design reduces the sensor condensation rate in snow melting scenarios from 20% to 0, improves the accuracy of snow depth / ice thickness measurement by 40%, and completely solves the problem of "perception paralysis" in extreme low temperatures.

[0065] The scenario-triggered dynamic coupling weight design ensures that the prediction error fluctuation is ≤5% in different extreme scenarios; the wake-terrain coupling model fills the gap in existing technologies that ignore the influence of terrain, reducing the power prediction error of wind farm clusters by 35%; the MAPE of power prediction for a single site from 0 to 4 hours is 8.2%, and the MAPE of the cluster is 7.3%, which are 60% and 65% lower than the existing fixed fusion schemes, respectively.

[0066] By constructing a cross-domain collaborative closed loop of "prediction-operation and maintenance-grid", operation and maintenance response delays were reduced by 30%, and equipment failure rate was reduced from 12% to 3.5% (3.2% for the cluster). The federated learning + blockchain architecture not only protects business privacy but also achieves trusted evidence storage, increasing grid absorption rate by 10%. A single 99MW wind farm generates an additional 6 million kWh of electricity annually, and a 2000MW regional cluster generates an additional 40 million kWh of electricity annually, directly increasing revenue by 20 million yuan and reducing annual operation and maintenance costs by 1.8 million yuan.

[0067] The pioneering high-altitude blizzard-melt secondary icing prediction model reduces the power deviation caused by secondary icing from 18.5% to 8.2%. The sensor's "hardware protection-software correction" linkage design enables the first practical application in wind farms above 3000m altitude, reducing the deployment cost of a single site by 150,000 yuan and the deployment cost of a regional cluster by 800,000 yuan.

[0068] An innovative four-step approach—"cluster partitioning, privacy collaboration, centralized forecasting, and dynamic allocation"—solves the core pain points of large power fluctuations and grid connection difficulties in regional wind farms. Cluster power fluctuation rate is reduced from 20% to 7.8%, and grid acceptance capacity is improved by 15%. The privacy-protected collaboration mechanism avoids the leakage of commercial data, and the dynamic allocation strategy adapts to equipment and grid constraints under extreme weather conditions, providing key technical support for the large-scale grid connection of regional wind farms.

[0069] Next, in an overall embodiment, taking into account a high-altitude snowstorm-melt type wind farm example in a certain province and the corresponding regional cluster, the implementation details of this application will be further explained: 1. Basic parameters.

[0070] Single site: A high-altitude wind farm in a certain province, with an installed capacity of 60MW (20 3MW permanent magnet direct-drive wind turbines, radius R=55m, extreme minimum temperature in winter -35℃, and an average of 4-6 blizzard-melt events per year). Regional cluster: includes 8 similar high-altitude wind farms with a total installed capacity of 400MW, connected to the same 220kV substation, with a maximum distance of 45km between the stations, a meteorological field correlation coefficient of 0.85, and a terrain slope of 5°-8°.

[0071] 2. Hardware deployment plan.

[0072] Single wind farm hardware: Each wind turbine is equipped with a 24GHz millimeter-wave radar, fiber optic strain sensor, and ultrasonic snow depth sensor; two fixed-wing meteorological stations are deployed around the wind farm, and one wind profiler radar is deployed at the farm boundary; the central control room is equipped with an edge computing server (CPU i7-13700K + GPU RTX4090). Cluster hardware: Added regional center server (CPU i9-13900K+GPU A100), consortium blockchain node server (including 8 sites, a provincial power grid company, and maintenance companies, totaling 10 nodes); configured 5 explosion-proof inspection robots, and optimized the layout of on-site points according to the cluster maintenance radius.

[0073] 3. Model training and field validation.

[0074] 3.1 Training parameter settings.

[0075] Dataset: Historical data from 2021 to 2023 from 8 sites within the cluster (including 12 cluster-level blizzard-melt events); MAML meta-learning: 1000 iterations, pre-training learning rate 0.0003, fine-tuning learning rate 0.0001, batch size = 48; Cluster-based model: hidden layer dimension 512, encoder / decoder each 10 layers, spatial attention weights are introduced into the cluster spatial correlation matrix.

[0076] 3.2 See also Figure 2 On-site verification of a blizzard-melt event in a certain month of a certain year.

[0077] Event parameters: Blizzard across the cluster, snow depth 15-18mm, ambient temperature -12℃→2℃, relative humidity 85%-90%, secondary icing thickness 4-5mm; (1) Sensing layer performance: In a low temperature environment of -30℃, the sensors of 5 wind turbines in 3 stations froze and failed. The system automatically activated the emergency reverse thrust module, the reverse thrust wind speed error was 0.5-0.7m / s, and the probability of cluster sensing interruption was 0.2%; (2) Single-site forecast performance: 0-4 hour power forecast MAPE=8.2%, of which 0-2 hour short-term forecast MAPE=6.5%; (3) Cluster forecast performance: 0-4 hour concentrated power forecast MAPE=7.3%, cluster power volatility=7.8%; (4) Correction and allocation performance: Triggered 3 cluster-level cross-domain corrections, the operation and maintenance robot completed the inspection of 3 sites according to the optimized route, the accuracy of the power grid window period suggestion was 93%, and the deviation of each site after power allocation was ≤2%; (5) Operation and maintenance effect: The failure rate of cluster equipment is 3.2%, which is 73.3% lower than the traditional solution of 12%.

[0078] Corresponding to the above method embodiments, this specification also provides an embodiment of a wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration. Figure 3 This specification illustrates a schematic diagram of a wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration, according to one embodiment of this specification. Figure 3 As shown, the device includes: The data acquisition module 301 is configured to deploy heterogeneous sensing devices for data acquisition and to activate an emergency backtracking mechanism to obtain key data when sensing fails. It also preprocesses and labels the acquired and backtracked data to distinguish extreme scenarios. The coupling prediction module 302 is configured to input the preprocessed data into the scenario-triggered physical-AI dynamic coupling prediction model. The model automatically adjusts the weight ratio of the physical model and the AI ​​model according to the working condition label, and outputs the predicted power value of a single wind farm through physical feature extraction, time series prediction and physical constraint loss. The linkage correction module 303 is configured to trigger a cross-domain linkage correction process based on power deviation and operation and maintenance requirements, correct the power prediction value of a single wind farm, and link operation and maintenance adjustments and grid dispatch. The correction process iterates model parameters through a federated learning framework and stores evidence through a blockchain architecture. The collaborative optimization module 304 is configured to divide multiple wind farms into clusters, construct a regional centralized prediction model by coordinating data within the clusters through federated learning, output regional centralized power prediction values, and use a dynamic allocation strategy to distribute the regional centralized power prediction values ​​to each wind farm within the cluster. Model training module 305 is configured to train and validate scene-triggered physics-AI dynamic coupling prediction models and regional centralized prediction models using meta-learning algorithms and datasets containing extreme scenarios.

[0079] In one possible implementation, heterogeneous sensing devices include millimeter-wave radar, fiber optic strain sensors, ultrasonic snow depth sensors, fixed-wing weather stations, and wind profiler radar. The ultrasonic snow depth sensor integrates a heating-ventilation dual-control anti-condensation module, the millimeter-wave radar is equipped with a self-heating insulation sleeve, and the fixed-wing meteorological station is deployed around the wind farm and distributed at preset intervals to collect the wake wind speed between the wind turbines.

[0080] In one possible implementation, the emergency reverse-engineering mechanism includes: The wind speed is calculated by inversely based on the real-time torque and speed of the wind turbine generator; The real-time current of the blade pitch motor is used to inversely calculate the blade ice thickness or snow depth through a preset current-load mapping model. Spatial consistency verification is performed between the ice thickness or snow depth on the blades and the wind speed and the sensing data of adjacent wind turbines. If the deviation exceeds the threshold, a weighted voting mechanism is activated to determine the final data.

[0081] In one possible implementation, the weight adjustment rules in the scene-triggered physics-AI dynamic coupling prediction model are driven by the scene feature vector; The scene feature vector includes wind speed fluctuation rate, ice thickness growth rate, and temperature change rate. In the typhoon scene, the physical model has a weight of 70% and the AI ​​model has a weight of 30%. In the ice accumulation scene, the physical model has a weight of 40% and the AI ​​model has a weight of 60%. In the blizzard-snowmelt transition scene, the physical model has a weight of 60% and the AI ​​model has a weight of 40%.

[0082] In one possible implementation, physical feature extraction includes: Theoretical power correction and wake-terrain coupling calculations are performed. Theoretical power correction is based on the wind turbine aerodynamic formula and introduces correction terms for air density in cases of icing, blizzards, and high altitudes. Wake-terrain coupling calculations correct the basic wake attenuation coefficient by introducing a terrain slope factor. The physical constraint loss is calculated based on a hybrid loss function that includes prediction error terms, physical theoretical upper limit deviation terms, and wind energy capture coefficient deviation terms. The weights of each term in the hybrid loss function are dynamically allocated according to the extreme scenario type.

[0083] In one possible implementation, the triggering condition for the cross-domain collaborative correction process is: In typhoon or blizzard scenarios, the power deviation rate exceeds the first threshold and the predicted future power fluctuation rate exceeds the second threshold. In ice accumulation or snow melting scenarios, the power deviation rate exceeds the third threshold and the blade fatigue damage value exceeds the preset threshold. The correction process in the blizzard-snowmelt scenario also activates the secondary icing prediction model, which predicts the secondary icing thickness based on ambient temperature and relative humidity.

[0084] In one possible implementation, the dynamic allocation strategy includes: The base power quota is allocated according to the prediction accuracy of each wind farm's single station. The allocation ratio will be adjusted according to the equipment health score of each wind farm. Verification and correction are performed based on power grid line capacity constraints and power grid stability constraints.

[0085] The above is a schematic scheme of a wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration according to this embodiment. It should be noted that the technical solution of this wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration belongs to the same concept as the technical solution of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above. Details not described in detail in the technical solution of the wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration can be found in the description of the technical solution of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above.

[0086] Figure 4 A structural block diagram of a computing device 400 according to one embodiment of this specification is shown. The components of the computing device 400 include, but are not limited to, a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and a database 450 is used to store data.

[0087] The computing device 400 also includes an access device 440, which enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 440 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0088] In one embodiment of this specification, the aforementioned components of the computing device 400 and Figure 4 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 4The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0089] The computing device 400 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 400 can also be a mobile or stationary server.

[0090] The processor 420 executes the following computer-executable instructions, which, when executed by the processor, implement the steps of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above. The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above belong to the same concept. Details not described in detail in the technical solution of the computing device can be found in the description of the technical solution of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above.

[0091] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above.

[0092] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above.

[0093] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, the computer is instructed to perform the steps of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above.

[0094] The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the technical solution of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above. For details not described in detail in the technical solution of the computer program, please refer to the description of the technical solution of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration described above.

[0095] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0096] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0097] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0098] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0099] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration, characterized in that, include: Deploy heterogeneous sensing devices to collect data and activate an emergency reverse mechanism to obtain key data when sensing fails. Preprocess and label the collected and reversed data to distinguish extreme scenarios. The preprocessed data is input into a scenario-triggered physical-AI dynamic coupling prediction model. The model automatically adjusts the weight ratio of the physical model and the AI ​​model according to the working condition label, and outputs the predicted power value of a single wind farm through physical feature extraction, time series prediction and physical constraint loss. Based on the cross-domain linkage correction process triggered by power deviation and operation and maintenance requirements, the power prediction value of the single wind farm is corrected, and operation and maintenance adjustments and grid dispatch are linked. The correction process iterates the model parameters through a federated learning framework and stores the evidence through a blockchain architecture. Multiple wind farms are divided into clusters. A regional centralized prediction model is constructed by coordinating data within the clusters through federated learning. The regional centralized power prediction value is output, and a dynamic allocation strategy is used to distribute the regional centralized power prediction value to each wind farm within the cluster. The scenario-triggered physical-AI dynamic coupling prediction model and the regional centralized prediction model are trained and validated using a meta-learning algorithm and a dataset containing extreme scenarios.

2. The method according to claim 1, characterized in that, The heterogeneous sensing devices include millimeter-wave radar, fiber optic strain sensor, ultrasonic snow depth sensor, fixed-wing weather station, and wind profiler radar. The ultrasonic snow depth sensor integrates a heating-ventilation dual-control anti-condensation module, the millimeter-wave radar is equipped with a self-heating insulation sleeve, and the fixed-wing meteorological station is deployed around the wind farm and distributed at preset intervals to collect the wake wind speed between wind turbines.

3. The method according to claim 2, characterized in that, The emergency reverse propagation mechanism includes: The wind speed is calculated by inversely based on the real-time torque and speed of the wind turbine generator; The real-time current of the blade pitch motor is used to inversely calculate the blade ice thickness or snow depth through a preset current-load mapping model. The spatial consistency of the ice thickness or snow depth on the blades and the wind speed with the sensing data of adjacent wind turbines is checked. If the deviation exceeds the threshold, a weighted voting mechanism is activated to determine the final data.

4. The method according to claim 1, characterized in that, The weight adjustment rule in the scene-triggered physics-AI dynamic coupling prediction model is driven by the scene feature vector; The scene feature vector includes wind speed fluctuation rate, ice thickness growth rate, and temperature change rate; in the typhoon scene, the physical model weight is 70% and the AI ​​model weight is 30%; in the ice accumulation scene, the physical model weight is 40% and the AI ​​model weight is 60%; and in the blizzard-melt transition scene, the physical model weight is 60% and the AI ​​model weight is 40%.

5. The method according to claim 1, characterized in that, The physical feature extraction includes: The theoretical power correction and wake-terrain coupling calculation are performed. The theoretical power correction is based on the wind turbine aerodynamic formula and introduces correction terms for air density in cases of icing, blizzards, and high altitudes. The wake-terrain coupling calculation corrects the basic wake attenuation coefficient by introducing a terrain slope factor. The physical constraint loss is calculated based on a hybrid loss function that includes a prediction error term, a physical theoretical upper limit deviation term, and a wind energy capture coefficient deviation term. The weights of each term in the hybrid loss function are dynamically allocated according to the extreme scenario type.

6. The method according to claim 1, characterized in that, The triggering condition for the cross-domain linkage correction process is: In typhoon or blizzard scenarios, the power deviation rate exceeds the first threshold and the predicted future power fluctuation rate exceeds the second threshold. In ice accumulation or snow melting scenarios, the power deviation rate exceeds the third threshold and the blade fatigue damage value exceeds the preset threshold. The correction process in the blizzard-melt scenario also activates a secondary icing prediction model, which predicts the secondary icing thickness based on ambient temperature and relative humidity.

7. The method according to claim 1, characterized in that, The dynamic allocation strategy includes: The base power quota is allocated according to the prediction accuracy of each wind farm's single station. The allocation ratio will be adjusted according to the equipment health score of each wind farm. Verification and correction are performed based on power grid line capacity constraints and power grid stability constraints.

8. A wind farm power prediction device based on multi-dimensional redundancy and cross-domain collaboration, characterized in that, include: The data acquisition module is configured to deploy heterogeneous sensing devices for data acquisition and to activate an emergency reverse mechanism to obtain key data when sensing fails. The acquired and reversed data are preprocessed and labeled with operating conditions to distinguish extreme scenarios. The coupled prediction module is configured to input the preprocessed data into a scenario-triggered physical-AI dynamic coupled prediction model. The model automatically adjusts the weight ratio of the physical model and the AI ​​model according to the working condition label, and outputs the predicted power value of a single wind farm through physical feature extraction, time series prediction and physical constraint loss. The linkage correction module is configured to trigger a cross-domain linkage correction process based on power deviation and operation and maintenance requirements, correct the power prediction value of the single wind farm, and link operation and maintenance adjustments and grid dispatch. The correction process iterates the model parameters through a federated learning framework and stores the evidence through a blockchain architecture. The collaborative optimization module is configured to divide multiple wind farms into clusters, construct a regional centralized prediction model by coordinating data within the clusters through federated learning, output regional centralized power prediction values, and use a dynamic allocation strategy to distribute the regional centralized power prediction values ​​to each wind farm within the cluster. The model training module is configured to train and validate the scenario-triggered physical-AI dynamic coupling prediction model and the regional centralized prediction model using a meta-learning algorithm and a dataset containing extreme scenarios.

9. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the wind farm power prediction method based on multi-dimensional redundancy and cross-domain collaboration as described in any one of claims 1 to 7.

Images