Wind turbine blade predictive anti-icing system and method based on data driving

Abstract

The invention belongs to the technical field of intelligent operation and maintenance of wind power generation equipment, and particularly relates to a wind turbine blade predictive anti-icing system and method based on data driving, and the system comprises a data collection and preprocessing module which is used for collecting multi-source heterogeneous data and carrying out the cleaning, fusion and feature extraction; the preprocessing comprises abnormal value elimination, space-time alignment and feature engineering; the data driving prediction model module is used for constructing and training a prediction model based on the preprocessed data and outputting a blade surface icing risk prediction result; the intelligent anti-icing execution module comprises a distributed electric heating unit and a carbon fiber self-adaptive airflow regulation and control unit; the distributed electric heating unit is embedded into the front edge of the blade and an easy-to-freeze area; according to the invention, through collection, fusion and modeling of multi-source heterogeneous data, a high-precision icing prediction model is constructed, an anti-icing execution mechanism is driven to realize dynamic cooperative control, the anti-icing efficiency is significantly improved, and the energy consumption is reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent operation and maintenance of wind power generation equipment, and particularly relates to a wind turbine blade predictive anti-icing system and method based on data driving. BACKGROUND

[0002] With the global wind energy development expanding to low-temperature complex environments such as high latitudes and high altitudes, the power loss and structural safety hazards caused by icing of wind turbine blades have become industry pain points. Traditional anti-icing technologies mainly rely on passive deicing, which has problems such as high energy consumption and response lag, and the false alarm rate of the early warning mechanism relying on a single sensor is as high as 35%, which is difficult to cope with the risk of hidden icing, such as supercooled water droplet adhesion when humidity is greater than 85% and wind speed is less than 2 m / s.

[0003] Problems of the prior art: In the prior art, although some schemes introduce data modeling ideas, there are the following limitations: the data source is single, and it mainly relies on meteorological station data, ignoring the correlation between blade surface micro-environment and historical operation and maintenance data; the prediction model has weak generalization ability, and is based on a fixed threshold or a simple linear model, which cannot adapt to the differences in icing characteristics of different wind farms and different blade models; the anti-icing execution is disconnected from the prediction model, and lacks a dynamic adjustment mechanism based on real-time data feedback, resulting in "excessive anti-icing" or "insufficient anti-icing"; therefore, there is an urgent need for a driving mechanism based on deep fusion of multi-source data to realize accurate prediction of icing risk and intelligent optimization of anti-icing strategies, and to upgrade from "passive response" to "active prevention". SUMMARY

[0004] The purpose of the present application is to provide a wind turbine blade predictive anti-icing system and method based on data driving, which can collect, fuse and model multi-source heterogeneous data, build a high-precision icing prediction model, drive the anti-icing execution mechanism to realize dynamic collaborative control, and significantly improve the anti-icing efficiency and reduce the energy consumption.

[0005] The technical solutions adopted by the present application are as follows: A wind turbine blade predictive anti-icing system based on data driving, comprising: a data acquisition and preprocessing module for acquiring multi-source heterogeneous data and performing cleaning, fusion and feature extraction; the multi-source heterogeneous data includes real-time meteorological data, blade state data, wind farm SCADA system data and third-party data; a data-driven prediction model module for building and training a prediction model based on the preprocessed data, and outputting a blade surface icing risk prediction result; The intelligent anti-icing execution module comprises a distributed electric heating unit and a carbon fiber adaptive airflow regulation unit; the distributed electric heating unit is embedded in the leading edge of the blade and the icing prone area, and the carbon fiber adaptive airflow regulation unit changes the surface airflow speed through blade micro-variable pitch to suppress water droplet adhesion. The data closed-loop control module is used for comparing the predicted icing state with the actual icing state data in real time, and the data is obtained by an infrared sensor or image recognition, and dynamically optimizing the anti-icing strategy and triggering the model parameter self-updating.

[0006] The preprocessing includes outlier removal, spatiotemporal alignment and feature engineering.

[0007] The third-party data includes satellite cloud image and regional climate model data.

[0008] The preprocessing module identifies and removes the data deviating from the normal range caused by sensor failure and transmission interference based on the combination of 3σ criterion and isolation forest algorithm; the data of different collection frequencies are time calibrated based on high-precision timestamps; and icing-related derived features are generated from the original data.

[0009] The prediction model comprises a basic model layer and a deep optimization layer, which learn the time sequence correlation between meteorological parameters and icing state through a sliding window mechanism, output the icing risk value, high-risk area coordinates and ice thickness prediction in the future 1-6 hours; the basic model layer adopts random forest or gradient boosting tree, and the deep optimization layer adopts improved LSTM or Transformer model.

[0010] The data closed-loop control module calculates the deviation rate by comparing the predicted value with the actual icing state in real time, the deviation rate threshold is ≤5%, the model parameter self-updating is triggered when the deviation is out of limit, the optimization period is ≤15 minutes, and the heating power and airflow regulation amplitude are dynamically adjusted according to the historical energy consumption data.

[0011] A data-driven predictive anti-icing method for wind turbine blades, comprising the following steps: S1, data system construction: deploying a sensor network, collecting nearly 3 years of historical data, completing data preprocessing through an edge computing node, and forming a standardized data set; S2, model training and deployment: building a hybrid prediction model of “basic model + deep optimization layer” on a cloud server, and deploying it to a wind farm edge node after training; S3, real-time risk prediction: the edge node receives real-time data input into the prediction model to generate an icing risk heat map, and sends an anti-icing instruction to the data closed-loop control module when the risk value of any area is ≥60%; S4, dynamic anti-icing execution: the data closed-loop control module starts the execution mechanism according to the risk level; S5, model iteration optimization: collect actual icing state data every hour to calculate prediction deviation and update model weight.

[0012] In S2, the mixed prediction model takes the "weather features + blade state features" in the historical data as input and the "icing risk and thickness" as output, optimizes the hyperparameters through 5-fold cross-validation, and is deployed to the edge node of the wind field to support offline prediction.

[0013] The dynamic anti-icing execution of S4 meets the following control logic: The first threshold value of the risk value is 60%, and the second threshold value is 70%; Low risk, when the risk value is 60%-70%, only start airflow regulation; High risk, when the risk value is greater than 70%, link distributed heating and micro-variable pitch, target to maintain surface temperature > 2℃ and surface airflow speed > 8m / s.

[0014] In S5, the model weight is updated through the reinforcement learning algorithm; the energy consumption data is summarized daily, and the "heating power-airflow regulation" combination strategy is optimized based on the Pareto optimization principle.

[0015] The technical effects achieved by the present application are: The present application breaks through the limitations of traditional single threshold method through multi-source data fusion and deep time series model, and the accuracy of icing risk prediction is improved, and the early warning time is effectively shortened compared with the digital twin scheme; The closed-loop control mechanism based on real-time feedback data reduces the anti-icing energy consumption, and avoids the blade thermal damage caused by "over-icing"; The present application migrates learning through historical data, and the model can quickly adapt to different types of blades without the need to rebuild a three-dimensional model, effectively shortening the deployment period; The present application significantly reduces the power loss caused by icing, reduces the operation and maintenance cost, and is suitable for complex weather wind fields such as high altitude and coastal areas.

[0016] The present application can effectively reduce the downtime of wind turbines in the icing season, effectively control the aerodynamic power loss, significantly improve the power supply stability of the power grid, save the operation and maintenance cost of high-latitude wind fields, and show significant engineering application value and economic benefits. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a system structure diagram of the present application; Figure 2 is a method flowchart of the present application. DETAILED DESCRIPTION

[0018] In order to make the purpose and advantages of the present application more clear and apparent, the present application will be specifically described below in conjunction with embodiments. It should be understood that the following description is merely used to describe one or several specific embodiments of the present application and does not strictly limit the scope of protection specifically requested by the present application.

[0019] As shown in Figure 1 A data-driven predictive anti-icing system for wind turbine blades comprises: A data acquisition and preprocessing module is configured to acquire multi-source heterogeneous data and perform cleaning, fusion and feature extraction. Preprocessing includes outlier removal, time-space alignment and feature engineering, such as extracting derived features such as "supercooled water droplet concentration" and "wind speed shear". Real-time weather data includes temperature, humidity, wind speed and precipitation data. Blade state data includes surface temperature distribution, vibration signal and strain data. Wind farm SCADA system data includes unit operating parameters and historical icing records. The multi-source heterogeneous data includes real-time weather data, blade state data, wind farm SCADA system data and third-party data, and the third-party data includes satellite cloud images and regional climate model data. A data-driven predictive model module is configured to build and train a predictive model based on preprocessed data, and output blade surface icing risk prediction results. The preprocessing module identifies and removes data deviating from the normal range caused by sensor failure and transmission interference based on a combination of the 3σ criterion and the isolation forest algorithm. Time calibration is performed on data with different collection frequencies based on high-precision timestamps. Icing-related derived features are generated from the original data. The predictive model includes a basic model layer and a deep optimization layer. The sliding window mechanism is used to learn the time sequence correlation between weather parameters and icing states, and to output icing risk values, high-risk area coordinates and ice thickness predictions for the next 1-6 hours. The basic model layer uses random forest or gradient boosting tree, and the deep optimization layer uses improved LSTM or Transformer model. An intelligent anti-icing execution module includes a distributed electric heating unit and a carbon fiber adaptive airflow regulation unit. The distributed electric heating unit is embedded in the leading edge of the blade and the icing-prone area. The carbon fiber adaptive airflow regulation unit changes the surface airflow speed by adjusting the blade pitch, thereby inhibiting water droplet adhesion. A data closed-loop control module is configured to compare the predicted icing state with the actual icing state data in real time, and the data is obtained by infrared sensors or image recognition. The data closed-loop control module calculates the deviation rate by comparing the predicted value with the actual icing state in real time. When the deviation rate threshold is less than or equal to 5%, the model parameter self-update is triggered when the deviation exceeds the limit. The optimization period is less than or equal to 15 minutes. The heating power and airflow regulation amplitude are dynamically adjusted according to the historical energy consumption data.

[0020] As shown in Figure 2As shown, a data-driven wind turbine blade predictive deicing method includes the following steps: S1, data system construction: deploy a sensor network including a weather station, blade surface temperature sensor, fiber optic strain gauge, and high-definition camera, collect nearly 3 years of historical data, and the historical data contains at least 3 complete icing seasons, complete data preprocessing through edge computing nodes to form a standardized data set, and the sample size of the standardized data set is ≥5000 groups; S2, model training and deployment: build a hybrid prediction model of "base model + deep optimization layer" on a cloud server, deploy to a wind farm edge node after training; the hybrid prediction model takes "meteorological features + blade state features" in historical data as input and "icing risk and thickness" as output, optimizes hyperparameters through 5-fold cross-validation, the model accuracy is ≥93%, and the trained model is deployed to a wind farm edge node to support offline prediction with a delay of ≤1 second; S3, real-time risk prediction: the edge node receives real-time data every 5 minutes, and the meteorological data is updated at a frequency of 1 Hz and the blade state data is updated at a frequency of 10 Hz, which is input into the prediction model to generate a 5 cm x 5 cm resolution icing risk heat map. When the risk value of any region is ≥60%, send a deicing instruction to the data closed-loop control module; S4, dynamic deicing execution: the data closed-loop control module starts the actuator according to the risk level; The dynamic deicing execution satisfies the following control logic: Define the first threshold of risk value as 60% and the second threshold as 70%; Low risk, when the risk value is between 60%-70%, only start airflow regulation; High risk, when the risk value is >70%, link distributed heating and micro-variable pitch, target to maintain surface temperature >2℃ and surface airflow speed >8m / s; S5, model iteration optimization: collect actual icing state data every hour to calculate prediction bias and update model weights; update model weights through reinforcement learning algorithm; summarize energy consumption data daily, optimize "heating power-airflow regulation" combination strategy based on Pareto optimization principle; reduce unit deicing energy consumption by ≥20%.

[0021] Example 1 Application scenario: high-humidity mountain wind farm deicing system, annual humidity >85%, winter rain and snow, blade icing leads to an annual power generation loss of 12%; Data collection and preprocessing Deploy microwave icing sensors, fiber optic strain gauges, and infrared thermometers with a resolution of 0.01 mm, collect blade surface temperature, vibration signal, and ice thickness data every 10 minutes; Satellite cloud images and regional climate model data are fused to extract derived features such as "supercooled water droplet concentration" and "wind speed shear". Sensor fault data is removed through the 3σ criterion and the Isolation Forest algorithm. Prediction model construction The base model layer uses gradient boosting trees to learn the nonlinear relationship between historical meteorological data and icing conditions. The deep optimization layer uses an improved LSTM to capture the temporal variation of temperature and humidity. The model outputs a 6-hour icing risk thermal map with an accuracy of 95%. Anti-icing execution and closed-loop control Low risk: Start the carbon fiber adaptive airflow regulation unit. Increase the surface airflow speed to 8-10 m / s through micro-variable pitch to suppress water droplet adhesion. Medium to high risk: Link the distributed electric heating unit. The electric heating unit is embedded in the blade leading edge to maintain a surface temperature of >2°C. At the same time, start the air-heating deicing system to use the blade web cavity to circulate hot air. The data closed-loop module compares the predicted value with the infrared measured data every 15 minutes. If the deviation is more than 5%, trigger the model parameter self-update.

[0022] Implementation effect: Anti-icing energy consumption is reduced by 25%, icing-induced downtime is reduced by 80%, and annual average power generation is increased by 1.48 million kilowatt-hours.

[0023] Example 2 Application scenario: Wind field anti-icing system in coastal foggy areas with >150 days of annual average salt fog erosion. Blade surface icing and salt erosion synergistically reduce the service life by 30%. Data collection and preprocessing Use corrosion-resistant microwave sensors to monitor salt fog concentration and ice thickness in real time. Integrate SCADA system unit vibration data to identify changes in blade aerodynamic performance. Introduce ship meteorological data and calibrate multi-source data timestamps through spatiotemporal alignment algorithms. Prediction model construction The base model layer uses random forests to learn the relationship between salt fog concentration, wind speed, and icing probability. The deep optimization layer uses a Transformer model to process unstructured satellite cloud data. The model outputs a salt fog-icing coupling risk assessment to predict the impact of ice layer salt content on adhesion. Anti-icing execution and closed-loop control Low risk: Start the super-hydrophobic coating to reduce ice adhesion. At the same time, adjust the micro-variable pitch angle to make the surface airflow speed >8 meters / second. Medium to high risk: Link the distributed electric heating unit to maintain a surface temperature of >3°C. Optimize the heating power through reinforcement learning to reduce salt fog condensation. The data closed loop module summarizes energy consumption data daily, optimizes the combination of "heating power-air flow regulation" based on the Pareto optimization principle, and reduces the unit energy consumption by 22%.

[0024] Implementation effect: The sensor failure rate caused by salt spray is reduced by 50%, the blade coating life is extended to 5 years, and the anti-icing energy consumption is reduced by 30% compared with traditional electric heating wires.

[0025] Example 3 Application scenario: Anti-icing system for wind field in extremely cold regions, winter extreme low temperature up to-35℃, blade icing leads to annual average downtime exceeding 100 hours; Data acquisition and preprocessing Deploy low-temperature infrared sensors to collect blade surface temperature field every 5 minutes; fuse the Arctic Oscillation Index to predict the duration of extremely cold weather; The preprocessing module calibrates meteorological data and blade state data through high-precision timestamps to generate features such as "low-temperature duration" and "ice layer growth rate"; Prediction model construction The base model layer uses random forest to learn the icing threshold under extreme low temperature; the deep optimization layer uses an attention mechanism improved Transformer model to capture the mutation characteristics of temperature drop; The model outputs 12-hour ice layer thickness prediction, supporting offline prediction delay ≤1 second; Anti-icing execution and closed loop control Low risk: Start carbon fiber adaptive air flow regulation unit, form turbulence through micro-variable pitch, reduce ice layer adhesion; Medium and high risk: Link distributed electric heating unit, maintain surface temperature >2℃, and enable intelligent infrared ice melting equipment to melt 3cm thick ice layer within 20 seconds; The data closed loop module updates model weights every hour through reinforcement learning algorithm to optimize heating power distribution, reducing unit anti-icing energy consumption by 22%.

[0026] Implementation effect: The unit can operate continuously under extremely cold weather, the downtime caused by icing is reduced by 90%, and the blade surface temperature control accuracy is ±0.5℃.

[0027] In summary, the present application realizes precise anti-icing under different climate conditions through multi-source data fusion, dynamic risk prediction and adaptive anti-icing execution.

[0028] The above merely describes the preferred embodiments of the present application, and it should be pointed out that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application. The structures, devices and operation methods not specifically described and explained in the present application are implemented according to the conventional means in the art, unless specifically described and limited.

Claims

1. A data-driven predictive anti-icing system for wind turbine blades, characterized in that: include: Data acquisition and preprocessing module, used to collect multi-source heterogeneous data and perform cleaning, fusion and feature extraction. The multi-source heterogeneous data includes real-time meteorological data, blade status data, wind farm SCADA system data and third-party data; The data-driven prediction model module builds and trains a prediction model based on preprocessed data and outputs the blade surface icing risk prediction results; The intelligent anti-icing execution module includes a distributed electric heating unit and a carbon fiber adaptive airflow control unit. The distributed electric heating unit is embedded in the leading edge of the blade and in areas prone to icing. The carbon fiber adaptive airflow control unit changes the surface airflow velocity through blade micro-pitch to inhibit water droplet adhesion. The data closed-loop control module is used to compare the predicted icing status with the actual icing status data in real time. The data is obtained by infrared sensors or image recognition, dynamically optimizes the anti-icing strategy and triggers self-update of model parameters.

2. The data-driven predictive anti-icing system for wind turbine blades according to claim 1, characterized in that: Preprocessing includes outlier removal, spatiotemporal alignment, and feature engineering.

3. The data-driven predictive anti-icing system for wind turbine blades according to claim 1, characterized in that: The third-party data includes satellite cloud images and regional climate model data.

4. The data-driven predictive anti-icing system for wind turbine blades according to claim 1, characterized in that: The pre-processing module identifies and removes data that deviates from the normal range due to sensor failure and transmission interference based on the combination of the 3σ criterion and the isolation forest algorithm; Data with different acquisition frequencies are time-calibrated based on high-precision timestamps; ice-related derivative features are derived from the raw data.

5. The data-driven predictive anti-icing system for wind turbine blades according to claim 1, characterized in that: The prediction model includes a basic model layer and a deep optimization layer. It uses a sliding window mechanism to learn the temporal correlation between meteorological parameters and icing conditions, and outputs icing risk values, high-risk area coordinates, and ice thickness forecasts for the next 1-6 hours. The basic model layer adopts random forest or gradient boosting tree, and the deep optimization layer adopts improved LSTM or Transformer model.

6. The data-driven predictive anti-icing system for wind turbine blades according to claim 1, characterized in that: The data closed-loop control module calculates the deviation rate by comparing the predicted value with the actual icing state in real time. The deviation rate threshold is ≤5%. When the deviation exceeds the limit, the model parameters are automatically updated. The optimization cycle is ≤15 minutes, and the heating power and airflow control range are dynamically adjusted according to historical energy consumption data.

7. A data-driven predictive anti-icing method for wind turbine blades, applied to the data-driven predictive anti-icing system for wind turbine blades according to any one of claims 1 to 6, characterized in that: The following steps are involved: S1, data system construction: deploy a sensor network, collect historical data for the past three years, complete data preprocessing through edge computing nodes, and form a standardized data set; S2, model training and deployment: Build a hybrid prediction model consisting of a "basic model + deep optimization layer" on a cloud server and deploy it to wind farm edge nodes after training; S3, real-time risk prediction: The edge node receives real-time data and inputs it into the prediction model to generate an icing risk heat map. When the risk value of any area is ≥60%, an anti-icing instruction is sent to the data closed-loop control module. S4, dynamic anti-icing execution: the data closed-loop control module activates the actuator according to the risk level; S5, model iterative optimization: collect actual icing status data every hour to calculate the prediction deviation and update the model weight.

8. The data-driven predictive anti-icing method for wind turbine blades according to claim 7, characterized in that: In S2, the hybrid prediction model uses "meteorological characteristics + blade status characteristics" in historical data as input and "icing risk and thickness" as output. It optimizes hyperparameters through 5-fold cross-validation and supports offline prediction after being deployed to the edge nodes of the wind farm.

9. The data-driven predictive anti-icing method for wind turbine blades according to claim 7, wherein the dynamic anti-icing execution in S4 satisfies the following control logic: The first threshold of risk value is defined as 60%, and the second threshold is 70%; Low risk: When the risk value is between 60% and 70%, only airflow control is activated; Medium to high risk: When the risk value is greater than 70%, distributed heating and micro-pitch are linked, with the goal of maintaining a surface temperature greater than 2°C and a surface airflow velocity greater than 8 m / s.

10. The data-driven predictive anti-icing method for wind turbine blades according to claim 7, characterized in that: In S5, the model weights are updated through a reinforcement learning algorithm; energy consumption data is summarized daily, and the "heating power-airflow control" combination strategy is optimized based on the Pareto optimal principle.

Images