A data inversion and completion system and method based on a cabin wind measurement radar

Abstract

The application provides a system and method for data inversion and completion based on a cabin wind measurement radar, which has high adaptability and precision, and specifically relates to processing structured industrial data by using a machine learning model to achieve high-precision data repair. The system comprises a data acquisition module (1), a data preprocessing module (2), a model construction module (3), a model training module (4) and a data inversion module (5). The method comprises the following steps: data preparation, beam model construction by the model construction module (3), model training by the model training module (4), and data inversion and completion by the data inversion module (5). By constructing a machine learning model with XGBoost as the core, the inherent spatio-temporal structured features of wind speed data are fully utilized to perform high-precision inversion and completion on invalid or missing radial wind speed data. The application is applied to the technical field of wind power generation.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and more particularly to a data inversion and completion system and method based on nacelle wind-measuring radar. Specifically, this invention falls under the technical category of using machine learning models to process structured data for data repair and prediction, and is applicable to data completion of nacelle four-beam laser wind-measuring radar in wind turbine blade obstruction scenarios, providing reliable wind speed data for the safe and efficient operation of wind turbine generator sets. Background Technology

[0002] As a core development direction of renewable and clean energy, wind power generation has seen a continuous increase in single-unit capacity and blade length. While this has significantly improved power generation efficiency, it has also brought about the technical challenge of beam obstruction for nacelle wind-measuring radar. Nacelle laser wind-measuring radar is a key device for ensuring the safe operation of wind turbines. It collects second-level wind speed data in real time by emitting four sets of beams, providing important data for core operations such as wind turbine start-up and shutdown control and power regulation.

[0003] However, in actual operation, radar often fails to acquire complete wind measurement data due to obstruction by wind turbine blades. This problem is particularly prominent in two typical scenarios: first, when waiting for wind to pick up while the blades are unlocked, the dynamically rotating blades randomly obstruct part of the beam; second, in the operating scenario where the blades are locked, blades in fixed positions continuously obstruct specific beams, preventing the corresponding beams from detecting wind speed information normally. This obstruction not only affects the accuracy of wind measurement but may also reduce the operational safety and power generation efficiency of the wind turbine.

[0004] Existing technologies mainly use methods such as linear interpolation and moving average for data completion. These methods can only play a limited role when the beam is regularly blocked, and cannot obtain effective data in non-regular blocking scenarios such as wind turbine shutdown. At the same time, existing methods do not fully consider the spatiotemporal correlation of wind speed, making it difficult to capture the nonlinear relationship between each beam data and the time dimension, resulting in insufficient data completion accuracy and failing to provide reliable protection for wind turbine operation.

[0005] Against the backdrop of the rapid development of the next-generation information network industry, utilizing advanced machine learning algorithms to process industrial big data with inherent structure has become crucial for improving the intelligence level of traditional industries. Wind speed data, as a typical example of multidimensional structured time-series data, exhibits spatial correlations between different beams and strong autocorrelation within the same beam over time. However, existing technologies have failed to effectively utilize this structured data characteristic to build intelligent predictive models to address the challenge of data gaps under complex operating conditions.

[0006] Therefore, there is an urgent need for a technical method that can deeply mine and utilize the spatiotemporal structured characteristics of wind speed data, adapt to different operating states of wind turbines (start-up, operation, shutdown), and achieve high-precision data completion in scenarios with blade shading. This method should be based on advanced machine learning models to form an intelligent data repair and self-healing system, providing core technical support for the digital and intelligent operation and maintenance of wind power generation. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a highly adaptable and accurate data inversion and completion system and method based on cabin wind radar. This method belongs to the category of processing structured data using machine learning models.

[0008] The technical solution adopted in this invention is a data inversion and completion system based on cabin wind measurement radar. The system includes a data acquisition module, a data preprocessing module, a model building module, a model training module, and a data inversion module.

[0009] The data acquisition module is used to collect radial wind speed data from the cabin four-beam laser wind radar.

[0010] The data preprocessing module is used to perform data cleaning and variable extraction operations;

[0011] The model building module is used to build four models of the XGBoost model according to the four groups of beams respectively;

[0012] The model training module is used to perform dataset partitioning, model training, and parameter optimization operations.

[0013] The data inversion module is used to call the optimal model to invert and complete the missing data.

[0014] Furthermore, the data acquisition module is connected to the cabin laser wind radar via wired or wireless communication, supporting real-time data reception and historical data retrieval.

[0015] Furthermore, the system also includes a result output module, which outputs the complete wind speed data after inversion and completion to the wind turbine control system, providing data support for wind turbine start-up and shutdown control and power regulation.

[0016] A method for data inversion and completion using the aforementioned data inversion and completion system based on cabin wind radar, comprising the following steps:

[0017] S1. Data Preparation: Radial wind speed data from the cabin four-beam laser wind radar is collected through the data acquisition module. The laser wind radar collects data with a time resolution of seconds and a spatial resolution of 30m when there is no obstruction. The initial distance layer is 50m, and the maximum effective distance is 400m. The radial data of the four beams at the current time t are used as target variables in sequence. The input variables of the XGBoost model include the historical instantaneous data of the target beam at times t-1, t-5, and t-10, the historical average data of the target beam at the previous 5, 10, and 20 times, and the radial data of the other three beams at time t. The data is cleaned using the data preprocessing module to remove outliers, and filled and invalid values ​​are set to NaN.

[0018] S2. Beam model construction through the model construction module: four models are constructed according to the four groups of beams. Taking advantage of the characteristic that at least one beam is not blocked in the scenario of wind turbine blades not locked and waiting for wind or blades locked, the unblocked beam data is used as the input basis.

[0019] S3. Use the model training module to train the model: Divide the dataset processed in step S1 into a 50% training set and a 50% test set. Input the training set into the XGBoost model for training. At the same time, optimize the model parameters and verify the accuracy through the validation set to obtain the optimal inversion model for each beam.

[0020] S4. Use the data inversion module to complete the data inversion: Substitute the input variables at the time to be processed into the optimal inversion model of the corresponding beam, output the inversion wind speed data of the beam, complete the missing data of the obstructed beam, and form complete cabin wind measurement radar data.

[0021] Furthermore, in step S1, the collection period for the radial wind speed data covers at least two complete wind turbine operating cycle.

[0022] Furthermore, in step S1, data cleaning includes removing data that exceeds the reasonable wind speed range and abnormal data with continuous sudden changes greater than 5 m / s, wherein the reasonable wind speed range is 0 to 30 m / s.

[0023] Furthermore, in step S2, when constructing the model, the radial data of a single beam is used as the target variable, and the input variables are specifically: three historical instantaneous data, three historical average data, and synchronous radial data of the other three beams.

[0024] Furthermore, in step S3, the parameter optimization of the XGBoost model includes iterative adjustment of the learning rate, maximum tree depth, minimum sample weights, and subsampling rate.

[0025] Furthermore, in step S3, the evaluation indicators for accuracy verification include the coefficient of determination R², mean square error MSE, and root mean square error RMSE.

[0026] Furthermore, in step S4, when all four beams simultaneously produce invalid or filled values, the data at that moment is discarded.

[0027] The beneficial effects of this invention are:

[0028] This invention constructs XGBoost models for four beams respectively, eliminating the need for separate modeling for each distance layer. It utilizes the model's adaptability to wind speed values ​​to cover wind speed differences at different distance layers, thereby simplifying model construction complexity, reducing computational costs, and ensuring the relevance and effectiveness of data inversion for each beam.

[0029] The input variables of this invention simultaneously include the historical instantaneous data and historical average data of the target beam, as well as the synchronous data of other beams. This fully exploits the temporal correlation (historical data correlation) and spatial correlation (inter-beam data correlation) of wind speed, solving the defect of existing technologies that ignore spatiotemporal correlation and significantly improving the accuracy of data completion in complex scenarios.

[0030] This invention uses the XGBoost algorithm to build the model. This algorithm has a strong ability to capture nonlinear relationships and process NaN values. It can effectively adapt to the scenario of irregular data loss caused by wind turbine blade shading. It overcomes the problem that traditional linear interpolation and moving average methods are not effective in scenarios such as wind turbine shutdown, and realizes data completion throughout the entire operation cycle.

[0031] This invention cleans and removes outliers, iteratively optimizes model parameters, and uses the coefficient of determination, mean square error, and root mean square error as quantitative evaluation indicators to ensure the accuracy of model inversion. In practical applications, the coefficients of determination of each beam inversion model can provide reliable data support for wind turbine operation control.

[0032] The method of this invention has a clear process, is highly operable, and is compatible with the general parameters of the four-beam laser wind measurement radar in the nacelle. It has good practicality and compatibility and can be directly applied to the wind measurement data processing of existing wind power generation systems without the need to modify the radar hardware, thus reducing the cost of engineering applications. Attached Figure Description

[0033] Figure 1 This is a simplified structural diagram of the system of the present invention;

[0034] Figure 2 This is a simplified flowchart of the method of the present invention;

[0035] Figure 3 This is a roadmap of radar data inversion technology for this invention;

[0036] Figure 4 This is a scatter plot of wind speed retrieved from the radial data of the cabin radar based on the XGBoost model and the wind speed measured by the radar. Detailed Implementation

[0037] like Figures 1-4 As shown, the present invention provides a data inversion and completion system based on a cabin wind measurement radar. The system includes a data acquisition module 1, a data preprocessing module 2, a model building module 3, a model training module 4, and a data inversion module 5.

[0038] The data acquisition module 1 is used to collect radial wind speed data from the cabin four-beam laser wind radar.

[0039] The data preprocessing module 2 is used to perform data cleaning and variable extraction operations;

[0040] The model building module 3 is used to build four models of the XGBoost model according to the four groups of beams. The core of the model is to use the structured features such as the synchronous data of other beams and the historical data of this beam as input to learn the spatiotemporal correlation pattern of the wind field.

[0041] The model training module 4 is used to perform dataset partitioning, model training, and parameter optimization operations.

[0042] The data inversion module 5 is used to call the optimal model to invert and complete the missing data.

[0043] Furthermore, the data acquisition module 1 is connected to the nacelle laser wind radar via wired or wireless communication, supporting real-time data reception and historical data retrieval. The system also includes a result output module 6, used to output the inverted and completed complete wind speed data to the wind turbine control system, providing data support for wind turbine start-up and shutdown control and power regulation.

[0044] A method for data inversion and completion using the data inversion and completion system based on cabin wind radar as described above, the method comprising the following steps:

[0045] S1. Data Preparation: Radial wind speed data from the cabin four-beam laser wind radar is collected through the data acquisition module 1. The laser wind radar collects data with a time resolution of seconds and a spatial resolution of 30m when there is no obstruction. The initial distance layer is 50m, and the maximum effective distance is 400m. The radial data of the four beams at the current time t are used as target variables in sequence. The input variables of the XGBoost model include the historical instantaneous data of the target beam at times t-1, t-5, and t-10, the historical average data of the target beam at the previous 5, 10, and 20 times, and the radial data of the other three beams at time t. The data is cleaned using the data preprocessing module 2 to remove outliers, and the filled and invalid values ​​are set to NaN.

[0046] S2. Construct the 3-beam model using the model construction module: Construct four models according to the four beams respectively, taking advantage of the characteristic that at least one beam is not blocked in the scenario of wind turbine blades not locked and waiting for wind or blades locked, and using the unblocked beam data as the input basis.

[0047] S3. Use the model training module 4 to train the model: Divide the dataset processed in step S1 into a 50% training set and a 50% test set. Input the training set into the XGBoost model for training. At the same time, optimize the model parameters and verify the accuracy through the validation set to obtain the optimal inversion model for each beam.

[0048] S4. Use the data inversion module 5 to complete the data inversion: Substitute the input variables at the time to be processed into the optimal inversion model of the corresponding beam, output the inversion wind speed data of the beam, complete the missing data of the obstructed beam, and form complete cabin wind measurement radar data.

[0049] Specifically, in step S1, the collection period for the radial wind speed data covers at least two complete wind turbine operating cycle periods (including start-up, operation, and shutdown). Data cleaning includes removing data that exceeds the reasonable wind speed range and abnormal data with continuous abrupt changes greater than 5 m / s, wherein the reasonable wind speed range is 0–30 m / s.

[0050] In step S2, when constructing the model, the radial data of a single beam is used as the target variable, and the input variables are: three historical instantaneous data, three historical average data, and synchronous radial data of the other three beams.

[0051] In step S3, the parameter optimization of the XGBoost model includes iterative adjustments to the learning rate, maximum tree depth, minimum sample weights, and subsampling rate. The evaluation metrics for accuracy verification include the coefficient of determination (R²), mean squared error (MSE), and root mean square error (RMSE).

[0052] In step S4, when all four beams simultaneously produce invalid or filled values, the data at that moment is discarded and no inversion completion is performed.

[0053] The present invention will now be further described with reference to specific embodiments.

[0054] (a) Prerequisites for implementation

[0055] This embodiment uses measured data from a four-beam laser wind measuring radar in a wind farm nacelle. The radar parameters are: time resolution 1 second, spatial resolution 30m, initial distance layer 50m, and maximum effective distance 400m. The wind turbine model is 1.5MW, and the blade length is 42m. It covers three typical scenarios: blades not locked while waiting for wind to pick up, blades locked while running, and shutdown.

[0056] (II) Data Preparation Implementation

[0057] Data collection: Continuous radar radial wind speed data were collected from August 23 to 25, 2025, totaling 259,200 data sets (3 days × 24 hours × 3,600 seconds), including various scenarios such as startup, operation, and shutdown.

[0058] Variable definition: The radial data of beam 1 (RWS-1) is used as the target variable. The input variables include: historical instantaneous data of RWS-1 at t-1, t-5, and t-10, historical average data of RWS-1 at the previous 5, 10, and 20 times, and synchronous data of RWS-2, RWS-3, and RWS-4 at time t. At the same time, time features such as hours and minutes are extracted.

[0059] Data cleaning: Remove data with wind speed greater than 30m / s (128 sets removed in total), remove data with continuous abrupt changes greater than 5m / s (86 sets removed in total); set invalid marker values ​​in the original data to NaN, and fill in missing values ​​for historical moments using the forward filling method.

[0060] (III) Model Construction and Training Implementation

[0061] Model construction: Four XGBoost models are constructed according to the four beam groups, corresponding to the inversion completion of RWS-1 to RWS-4 respectively.

[0062] Dataset partitioning: The processed RWS dataset will be divided into a 50% training set and a 50% test set. The training set contains 86,293 data sets, and the test set contains 86,293 data sets.

[0063] Model training: The training set was input into the XGBoost model, and the initial parameters were set as follows: learning rate 0.1, maximum tree depth 6, minimum sample weight sum 1, and subsampling rate 0.8. The parameters were optimized through 5-fold cross-validation, and the optimal parameters were finally determined as follows: learning rate 0.08, maximum tree depth 5, minimum sample weight sum 1, and subsampling rate 0.9.

[0064] Accuracy Verification: The test set was input into the optimal model, and the inversion data of RWS-1 was output. The evaluation index was calculated by comparing it with the measured data. The results are: coefficient of determination (R²) 0.9539, mean square error (MSE) 0.1827, root mean square error (RMSE) 0.4274. Similarly, the accuracy verification results of the RWS-2 to RWS-4 models are shown in the table below, and the scatter plot of the inverted wind speed and the measured wind speed is shown in the figure below. Figure 4 As shown.

[0065] XGBoost model inversion radial wind speed and measured wind speed performance indicators

[0066]

[0067] (iv) Implementation of data inversion and completion

[0068] Shading Scenario Simulation: Complete daily data from 00:00 to 24:00 on August 25, 2025 (86,400 sets of data sampled at a frequency of 1 second per set) was selected. Combined with the wind turbine blade rotation cycle and wind direction variation patterns, four typical shading scenarios were simulated:

[0069] 1. 08:15:30~08:16:45 (75 seconds, 75 sets of data): Blades block beam 2 ’ (During the morning when wind speeds are low);

[0070] 2. 11:32:10~11:33:50 (100 seconds, 100 sets of data): Blades block beam 3 ’ (Medium wind speed period at noon);

[0071] 3. 16:48:20~16:49:35 (75 seconds, 75 sets of data): Blade blocking beam 1 ’ (Afternoon high wind speed period);

[0072] 4. 22:10:05~22:11:20 (75 seconds, 75 sets of data): Blades block beam 4 ’ (During periods of stable wind speed at night);

[0073] The duration and timing of various occlusion scenarios closely match the probability distribution characteristics of blades occluding radar beams in actual wind farm operation, ensuring the authenticity and representativeness of the simulated scenarios.

[0074] Model invocation: For the corresponding beams during the four types of occlusion periods mentioned above (blade occlusion beam 1) ’ 2 ’ 3 ’ 4 ’ The data includes the optimal model trained for each beam, the 24-hour historical valid data of the beam, and the synchronous measured data of the other three unobstructed beams.

[0075] Output results: Wind speed data for the corresponding shading period was obtained by inverting the optimal model of each beam and integrated in real time with the measured data of other unshading beams to form 86,400 complete sets of four-beam wind measurement data. Verification of the inversion results shows that the trend of the inverted data during the shading period is consistent with that of the measured data during the adjacent unshading period, with no obvious abrupt changes. The absolute value error of the wind speed is extremely small, and the root mean square error is also extremely small, which fully conforms to the continuous change law of the actual wind speed in the wind farm. Moreover, the complete data after completion has no missing values ​​or outliers, and the data is complete.

[0076] Application of Results: The complete 24-hour wind measurement data is continuously output to the wind turbine control system of the wind farm. Based on the continuous and accurate wind measurement data, the system dynamically adjusts the turbine pitch angle, yaw angle and power output threshold, which significantly improves the power generation efficiency of the turbine at different times of the day and under different shading scenarios. At the same time, it reduces the number of turbine shutdowns caused by missing or abnormal data to 0 times / day, which significantly improves the stability and economy of turbine operation.

[0077] Finally, it should be emphasized that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data inversion and completion method based on cabin wind radar, the method is implemented based on a data inversion and completion system of cabin wind radar, the system includes a data acquisition module (1), a data preprocessing module (2), a model building module (3), a model training module (4) and a data inversion module (5). The data acquisition module (1) is used to collect radial wind speed data from the cabin four-beam laser wind radar; The data preprocessing module (2) is used to perform data cleaning and variable extraction operations; The model building module (3) is used to build four models of the XGBoost model according to the four groups of beams respectively; The model training module (4) is used to perform dataset partitioning, model training and parameter optimization operations; The data inversion module (5) is used to call the optimal model to invert and complete the missing data; Its features are, The method includes the following steps: S1. Data preparation: The radial wind speed data of the cabin four-beam laser wind radar is collected by the data acquisition module (1). The laser wind radar collects data with a time resolution of seconds and a spatial resolution of 30m when there is no obstruction. The initial distance layer is 50m and the maximum effective distance is 400m. The radial data of the four beams at the current time t are used as target variables in sequence. The input variables of the XGBoost model include the historical instantaneous data of the target beam at time t-1, time t-5, and time t-10, the historical average data of the target beam at the previous 5, 10, and 20 times, and the radial data of the other three beams at time t. The data is cleaned by the data preprocessing module (2), outliers are removed, and the filled values ​​and invalid values ​​are set to NaN. S2. Beam model construction through the model construction module (3): Construct four models according to the four groups of beams respectively. Utilize the characteristic that at least one beam is not blocked in the scenario of wind turbine blades not locked and waiting for wind or blades locked, and take the unblocked beam data as the input basis. S3. Use the model training module (4) to train the model: Divide the dataset processed in step S1 into a training set of 50% and a test set of 50%, input the training set into the XGBoost model for training, optimize the model parameters, and verify the accuracy through the validation set to obtain the optimal inversion model for each beam. S4. Use the data inversion module (5) to complete the data inversion: Substitute the input variables at the time to be processed into the optimal inversion model of the corresponding beam, output the inversion wind speed data of the beam, complete the missing data of the obstructed beam, and form complete cabin wind measurement radar data.

2. The method according to claim 1, characterized in that, The data acquisition module (1) is connected to the cabin laser wind radar via wired or wireless communication, supporting real-time data reception and historical data retrieval.

3. The method according to claim 1, characterized in that, The system also includes a result output module (6), which outputs the complete wind speed data after inversion to the wind turbine control system to provide data support for wind turbine start-up and shutdown control and power regulation.

4. The method according to claim 1, characterized in that, In step S1, the collection period for the radial wind speed data covers at least two complete wind turbine operating cycles.

5. The method according to claim 1, characterized in that, In step S1, data cleaning includes removing data that exceeds the reasonable wind speed range and abnormal data with continuous sudden changes greater than 5 m / s. The reasonable wind speed range is 0 to 30 m / s.

6. The method according to claim 1, characterized in that, In step S2, when constructing the model, the radial data of a single beam is used as the target variable, and the input variables are: three historical instantaneous data, three historical average data, and synchronous radial data of the other three beams.

7. The method according to claim 1, characterized in that, In step S3, the parameter optimization of the XGBoost model includes iterative adjustment of the learning rate, maximum tree depth, minimum sample weights, and subsampling rate.

8. The method according to claim 1, characterized in that, In step S3, the evaluation indicators for accuracy verification include the coefficient of determination R², mean square error MSE, and root mean square error RMSE.

9. The method according to claim 1, characterized in that, In step S4, when all four beams simultaneously produce invalid or filled values, the data at that moment is discarded.

Images