Two-way communication method for twinborn scene

Abstract

The invention relates to the technical field of communication, in particular to a twinborn scene two-way communication method. Collecting multi-source real-time operation data and state data of the physical entity of the equipment, and uploading the multi-source real-time operation data and the state data to the corresponding digital twin; calculating a relative deviation based on the predicted equipment parameter of the digital twin and the actual equipment parameter of the equipment physical entity; judging whether the relative deviation exceeds a preset error threshold value or not, if yes, starting difference analysis on the relative deviation, and determining a deviation root cause causing the prediction deviation; based on the deviation root cause, combining with the performance attenuation state data of the physical entity of the equipment, and generating a deviation correction coefficient through a quantitative deviation correction model; the predicted equipment parameters of the digital twin are corrected according to the correction coefficient, and corrected equipment parameters are generated; and issuing the corrected equipment parameters to a corresponding monitoring system to complete a two-way communication closed loop. According to the method, the prediction precision and accuracy of the digital twin are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a two-way communication method for twin scenarios. Background Technology

[0002] Digital twin technology is widely used in the field of wind power generation prediction. However, traditional wind power digital twin models, once built, typically run continuously based on initially set parameters and algorithms, relying on historical data and fixed model logic to output prediction results. In existing solutions, prediction accuracy decreases over time due to environmental factors, equipment aging, and other factors, causing model drift. Moreover, deviation handling relies on human experience, making it difficult to quantify errors and pinpoint root causes, resulting in blind correction and insufficient prediction accuracy and poor performance of the digital twin model. Summary of the Invention

[0003] This invention addresses the technical problems of insufficient prediction accuracy and poor accuracy of existing digital twin models by providing a two-way communication method for twin scenarios.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: This invention provides a bidirectional communication method for a digital twin scenario, comprising: collecting multi-source real-time operating data and status data of a physical device entity, and uploading the multi-source real-time operating data and status data to a corresponding digital twin; calculating a relative deviation based on the predicted device parameters of the digital twin and the actual device parameters of the physical device entity; determining whether the relative deviation exceeds a preset error threshold, and if so, initiating a difference analysis of the relative deviation to determine the root cause of the prediction deviation; generating a correction coefficient based on the root cause of the deviation and the performance degradation status data of the physical device entity through a quantization correction model; and correcting the predicted device parameters of the digital twin according to the correction coefficient to generate corrected device parameters. The corrected device parameters are sent to the corresponding monitoring system to complete the two-way communication loop.

[0005] Optionally, the physical entity of the equipment includes a wind power generation equipment physical entity, the equipment parameters include power generation capacity, the multi-source real-time operating data includes environmental data and operating parameters, the status data includes equipment status data and historical operation and maintenance data, wherein the environmental data includes at least wind speed, wind direction, air temperature, turbulence intensity and air density, the operating parameters include at least generator output power, generator speed and pitch angle, the equipment status data includes at least main shaft bearing temperature, gearbox oil temperature, generator winding temperature and nacelle vibration spectrum data, and the historical operation and maintenance data includes at least cumulative operating time and historical maintenance frequency.

[0006] The process of initiating a difference analysis of the relative deviation to determine the root cause of the power prediction deviation includes: performing correlation analysis based on the multi-source real-time operating data and status data to identify a set of key influencing factors; inputting the set of key influencing factors into a pre-trained contribution calculation model to calculate the contribution of each key influencing factor to the current power prediction deviation; sorting the contribution in descending order and selecting the top N key influencing factors as the root cause of the power prediction deviation.

[0007] Among them, based on the multi-source real-time operation data and status data, correlation analysis is performed to identify a set of key influencing factors, including: calculating the correlation coefficient between the relative deviation and multiple monitoring variables, wherein the monitoring variables include at least wind speed, wind direction, turbulence intensity, blade pitch angle, gearbox oil temperature, generator winding temperature and cumulative operating time; and screening out monitoring variables whose absolute value of correlation coefficient is greater than a first preset threshold to form a set of key influencing factors.

[0008] The training steps of the contribution calculation model include: constructing a base model based on a gradient boosting tree; collecting historical multi-source real-time running data and status data as a historical data sample set, and identifying historical key influencing factors of the historical data samples; obtaining the relative deviation between the historical actual power and the historical predicted power corresponding to the historical data samples; using the historical key influencing factors as input features and the corresponding relative deviation values ​​as prediction targets, supervising the training of the base model; and integrating the SHAP or LIME interpretation framework on the trained base model to construct the contribution calculation model.

[0009] The steps for calculating the N value include: obtaining the average relative prediction deviation of the digital twin within a preset historical window; calculating the prediction confidence of the digital twin based on the average relative prediction deviation; and calculating the product of the prediction confidence and the total number of key influencing factors of a set of key influencing factors, and rounding down to obtain the N value.

[0010] The process involves generating correction coefficients based on the root causes of the deviation and the performance degradation status data of the physical entity of the equipment, using a quantization correction model. This includes: calling a pre-trained quantization correction model; obtaining monitoring variable data corresponding to the root causes of the deviation and combining it with the performance degradation status data to form an input feature vector, wherein the performance degradation status data includes cumulative running time and historical maintenance frequency; inputting the input feature vector into the quantization correction model and outputting the correction coefficients.

[0011] The training steps of the quantitative correction model include: constructing a quantitative correction model; collecting historical monitoring variable data and historical performance degradation status data at several historical moments as sample training datasets; obtaining the actual power generation and predicted power generation at the corresponding historical moments, calculating several historical correction coefficients as sample label datasets; and performing supervised training based on the sample training dataset and sample label datasets, obtaining the quantitative correction model after meeting the convergence conditions.

[0012] This also includes: If the relative deviation does not exceed the preset error threshold, the prediction output of the digital twin is determined to be reliable, and the predicted power generation is sent to the wind farm monitoring system as the final output value to complete the two-way communication closed loop.

[0013] By implementing this invention, it is possible to collect multi-source real-time operating data and status data of the physical entity of the device, and upload the multi-source real-time operating data and status data to the corresponding digital twin, so as to avoid the digital twin from becoming disconnected from the physical entity due to data loss or lag, and to lay an accurate data foundation for subsequent prediction, deviation analysis and other links, so that the operation of the digital twin always conforms to the actual situation of the physical device. By implementing this invention, it is possible to calculate the relative deviation between the predicted device parameters of the digital twin and the actual device parameters of the physical entity of the device, providing an objective and quantifiable basis for subsequent judgment on the reliability of the prediction results, avoiding the neglect of prediction deviation due to subjective judgment errors, and facilitating the accurate positioning of the severity of prediction errors. By implementing this invention, it is possible to determine whether the relative deviation exceeds a preset error threshold. If it does, a difference analysis of the relative deviation is initiated to determine the root cause of the prediction deviation. This avoids over-processing of small deviations within an acceptable range, which would otherwise waste resources. At the same time, it accurately locates the root cause, providing a clear direction for subsequent correction and preventing blind adjustments from exacerbating the problem. By implementing this invention, it is possible to generate correction coefficients through a quantitative correction model based on the root cause of the deviation and the performance degradation status data of the physical entity of the equipment. This avoids over- or under-adjustment caused by empirical correction based solely on the root cause, and ensures that the correction coefficients match the current deviation problem and conform to the actual performance of the equipment after aging, thereby improving the scientific nature and accuracy of the correction. By implementing this invention, the prediction device parameters of the digital twin can be corrected according to the correction coefficient, corrected device parameters can be generated, the prediction accuracy of the digital twin can be quickly restored, the model drift problem caused by prediction deviation can be solved, and the digital twin can maintain a high-reliability prediction capability. By implementing this invention, the corrected equipment parameters can be sent to the corresponding monitoring system to complete a two-way communication loop, ensuring that accurate predictions can directly guide actual work such as power grid dispatching and equipment operation and maintenance, while forming closed-loop management to provide feedback references for the next data collection and prediction process.

[0014] In summary, by implementing this invention, the prediction accuracy and precision of digital twins can be significantly improved, ensuring that the predicted power output always matches the actual operating status of physical equipment, providing a reliable basis for grid dispatch, and thus improving grid dispatch efficiency and the economic benefits of wind power projects. Attached Figure Description

[0015] Figure 1 A flowchart illustrating a bidirectional communication method for twin scenarios provided by the present invention; Figure 2 This is a schematic diagram illustrating the construction process of the contribution calculation model in a twin scenario bidirectional communication method provided by the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0018] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0019] Example 1, as Figure 1 As shown, this embodiment of the invention provides a bidirectional communication method for twin scenarios, including: S100 collects multi-source real-time operating data and status data of the physical entity of the device, and uploads the multi-source real-time operating data and status data to the corresponding digital twin; S200 calculates the relative deviation based on the predicted device parameters of the digital twin and the actual device parameters of the physical entity of the device; S300 determines whether the relative deviation exceeds a preset error threshold. If it does, it initiates a difference analysis of the relative deviation to determine the root cause of the prediction deviation. S400 generates correction coefficients based on the root cause of the deviation and the performance degradation status data of the physical entity of the equipment through a quantitative correction model. S500 corrects the prediction device parameters of the digital twin based on the correction coefficient, and generates corrected device parameters; S600 sends the corrected device parameters to the corresponding monitoring system, completing the two-way communication loop.

[0020] In embodiment S100 of this application, it is necessary to collect multi-source real-time operating data and status data of the physical entity of the equipment, and upload the multi-source real-time operating data and status data to the corresponding digital twin. The physical entity of the equipment includes a wind power generation equipment physical entity. The multi-source real-time operating data includes environmental data and operating parameters, and the status data includes equipment status data and historical operation and maintenance data.

[0021] The environmental data includes at least wind speed, wind direction, air temperature, turbulence intensity, and air density; the operating parameters include at least generator output power, generator speed, and blade pitch angle; the equipment status data includes at least main shaft bearing temperature, gearbox oil temperature, generator winding temperature, and nacelle vibration spectrum data; and the historical maintenance data includes at least cumulative operating time and historical maintenance frequency.

[0022] In this embodiment of the application, the purpose of step S100 is to build a data foundation for the digital twin in the wind power generation scenario that is synchronized with the physical entity in real time and fully mapped, so as to ensure that the digital twin can accurately replicate the actual operating status of the wind power generation equipment, and provide reliable data support for the entire process of subsequent prediction, deviation analysis, and correction, so as to avoid the disconnect between the digital twin and the physical equipment due to missing, delayed or one-sided data, which could lead to problems such as prediction deviation and scheduling error.

[0023] First, it is necessary to collect the multi-source real-time operating data, namely the environmental data and operating parameters.

[0024] Environmental data can be collected in real time by meteorological monitoring equipment such as anemometers, wind vanes, temperature sensors, turbulence monitors, and densitometers deployed at wind farms. This data includes wind speed, wind direction, air temperature, turbulence intensity, and air density, capturing the dynamic impact of the external environment on the operation of wind power generation equipment. Operating parameters can be obtained in real time from the equipment's own sensors and data acquisition modules, acquiring core operating indicators such as generator output power, generator speed, and blade pitch angle, directly reflecting the equipment's current power generation efficiency and operating status.

[0025] Next, it is necessary to collect the status data, namely the device status data and historical operation and maintenance data. The equipment status data includes at least the main shaft bearing temperature, gearbox oil temperature, generator winding temperature, and nacelle vibration spectrum data. This data can be collected through temperature and vibration sensors on the equipment to monitor the health status of key components in real time and identify potential problems such as overheating or abnormal vibration. Historical operation and maintenance data is retrieved from the wind farm's operation and maintenance management system, including long-term operating records such as cumulative equipment operating time and historical maintenance frequency, providing a basis for subsequent analysis of equipment performance degradation.

[0026] Then, through the Industrial Internet communication protocol, the collected multi-source real-time operation data and status data are transmitted in real time to the corresponding digital twin system of the wind power generation equipment.

[0027] The aforementioned digital twin system for wind power generation equipment is a digital mirror system that precisely corresponds to the physical equipment. Using the physical equipment as a prototype, the digital twin system integrates multi-source data such as wind speed, turbulence, equipment operating parameters (e.g., power generation and rotational speed), status data (e.g., component temperature and vibration spectrum), and historical maintenance data. Through modeling, it reconstructs the equipment's structure, operating logic, and performance characteristics. This digital twin system can simulate the equipment's operating status and predict power generation. It can also compare predicted and actual data to calculate relative deviations and identify the root causes of these deviations. The specific construction method for the digital twin system is existing technology and will not be elaborated here.

[0028] In step S200 of this embodiment, it is necessary to calculate the relative deviation based on the predicted device parameters of the digital twin and the actual device parameters of the physical device entity. The device parameters include power generation capacity.

[0029] In this embodiment of the application, the purpose of step S200 is to establish an objective and quantifiable accuracy measurement standard for the prediction results of the digital twin. By calculating the relative deviation between the predicted power generation and the actual power generation, the prediction effect of the digital twin is transformed from a subjective and vague judgment into an objective numerical evaluation. This provides a key basis for subsequent judgment on whether the prediction results are credible and whether a correction process needs to be initiated, while accurately reflecting the severity of the prediction error.

[0030] First, based on the multi-source real-time operation data and status data uploaded in step S100, the digital twin needs to calculate and output the predicted power generation value, denoted as P1, as the parameter of the prediction device.

[0031] Next, the actual power output of the wind turbine is collected in real time by the power monitoring sensors on the wind turbine itself, and recorded as P2. This power output data falls under the category of operating parameters in the multi-source real-time operating data of step S100, and has been collected and uploaded to the system.

[0032] Then, the relative deviation is calculated using a standardized formula: Relative deviation = |(P1-P2) / P2|×100%.

[0033] Finally, the calculated relative deviation value is stored in the system database, serving as the core basis for determining whether to initiate root cause analysis. For example, the calculated relative deviation is 8%.

[0034] Furthermore, it is necessary to determine whether the relative deviation exceeds a preset error threshold. If it does, a difference analysis of the relative deviation is initiated. The preset error threshold can be manually set according to the tolerance level for prediction error, such as 5%.

[0035] In step S300 of this application embodiment, the difference analysis of the relative deviation is initiated to determine the root cause of the power prediction deviation, including: Based on the multi-source real-time operation data and status data, a correlation analysis was performed to identify a set of key influencing factors; The set of key influencing factors is input into a pre-trained contribution calculation model to calculate the contribution of each key influencing factor to the current power prediction deviation. The contribution levels are sorted in descending order, and the top N key influencing factors are selected as the root causes of power prediction deviations.

[0036] In this embodiment of the application, the purpose of step S300 is to accurately pinpoint the core cause of the prediction deviation from the complex factors affecting wind power generation prediction, rather than attributing it to vague categories such as environmental changes or equipment problems.

[0037] First, a correlation analysis needs to be performed based on the multi-source real-time operating data and status data to identify a set of key influencing factors.

[0038] In step S300 of this application embodiment, a correlation analysis is performed based on the multi-source real-time operating data and status data to identify a set of key influencing factors, including: Calculate the correlation coefficient between the relative deviation and multiple monitored variables, wherein the monitored variables include at least wind speed, wind direction, turbulence intensity, blade pitch angle, gearbox oil temperature, generator winding temperature and cumulative operating time; Monitoring variables with absolute values ​​of correlation coefficients greater than the first preset threshold are selected to form a set of key influencing factors.

[0039] The first step is to select monitoring variables that are closely related to wind power generation, including environmental factors such as wind speed, wind direction, and turbulence intensity; operating parameters such as pitch angle; equipment status data such as gearbox oil temperature and generator winding temperature; and historical operation and maintenance data such as cumulative operating time.

[0040] The second step requires using statistical methods such as the Pearson correlation coefficient to calculate the correlation between the relative deviation and each monitored variable, and to obtain the correlation coefficient. The correlation coefficient ranges from -1 to 1, and the larger the absolute value of the correlation coefficient, the stronger the correlation.

[0041] Suppose that the wind power generation equipment in a wind farm experiences a power prediction deviation of 12%, exceeding the preset error threshold of 5%, and the S300 step analysis is initiated. The correlation coefficients between the relative deviation and multiple monitored variables are calculated. For example, the correlation coefficient between the relative deviation and turbulence intensity is 0.82, with gearbox oil temperature is 0.75, with cumulative operating time is 0.68, and with wind speed is 0.45, etc.

[0042] Next, based on the required level of precision in screening key influencing factors, a first preset threshold needs to be set empirically. Monitoring variables with correlation coefficients greater than this threshold are included in the set of key influencing factors, while variables with weak correlations are excluded, focusing on potential major influencing factors. For example, the first preset threshold could be 0.6. In the example above, according to the first preset threshold, turbulence intensity, gearbox oil temperature, and cumulative operating time are selected as key influencing factors and included in the set of key influencing factors.

[0043] Furthermore, the set of key influencing factors needs to be input into the pre-trained contribution calculation model to calculate the contribution of each key influencing factor to the current power prediction deviation.

[0044] Specifically, the selected key influencing factors need to be input into a pre-trained contribution calculation model. This model, trained with historical data, is capable of quantifying the impact of each factor on the prediction deviation and can output the specific contribution of each key influencing factor to the current power prediction deviation. For example, inputting the above three key influencing factors into the contribution calculation model yields the following contributions: turbulence intensity 45%, gearbox oil temperature 30%, and cumulative operating time 20%.

[0045] The following explains how to obtain the contribution calculation model.

[0046] like Figure 2 As shown, in step S300 of this application embodiment, the training step of the contribution calculation model includes: Construct a base model based on gradient boosting trees; The collected historical multi-source real-time operational data and status data are used as a historical data sample set, and the key historical influencing factors of the historical data samples are identified. Obtain the relative deviation between the historical actual power and the historical predicted power corresponding to the historical data sample; Using the aforementioned key historical influencing factors as input features and the corresponding relative deviation values ​​as prediction targets, the base model is trained under supervision. The contribution calculation model is constructed by integrating the SHAP or LIME interpretation framework onto the trained base model.

[0047] In this embodiment of the application, the core purpose of training the contribution calculation model is to build a model that can accurately quantify the contribution of each key influencing factor to the power prediction deviation and clearly explain the influence logic of each factor, so as to provide a reliable quantitative analysis tool for locating the root cause of the deviation in step S300.

[0048] Considering the task type of the contribution calculation model, in this embodiment of the application, a gradient boosting tree combined with the SHAP interpretation framework is used to build the contribution calculation model, and the gradient boosting tree is used as the base model.

[0049] Next, historical multi-source real-time operational and status data needs to be collected as a historical data sample set. Specifically, this requires collecting historical multi-source real-time operational and status data from the same type of wind power equipment over a historical period, such as three years. This includes historical environmental data, historical operating parameters, historical equipment status data, and historical maintenance data, and identifying key historical influencing factors. These key historical influencing factors are monitoring variables selected from the historical multi-source real-time operational and status data of the wind power equipment that have a significant impact on the historical power prediction deviation of the digital twin at a certain point in the past. Their acquisition method and data composition are consistent with the key influencing factors in the real-time scenario.

[0050] Finally, no less than 10,000 historical data samples were collected. Each sample contained the historical key influencing factors at the corresponding time and the relative deviation between the historical actual power and the historical predicted power. The samples were divided into a training set and a validation set in an 8:2 ratio as training data for the contribution calculation model.

[0051] In the parameter settings of the base model, the number of trees is 100; the maximum depth of each tree is 6; the learning rate is 0.1; the minimum number of sample splits per tree is 20; the minimum number of leaf nodes per tree is 5; the feature sampling ratio is 0.8; the subsampling ratio is 0.8; the maximum number of iterations is 500; the number of early stopping rounds is 50; the regularization coefficient is 0.1; the number of leaf nodes is limited to 60; and the split threshold is 0.001. In the training of the base model, an early stopping mechanism is adopted. The initial maximum number of training rounds is set to 500. Training is stopped when the loss function value on the validation set does not decrease for 50 consecutive rounds. The actual number of training rounds is determined based on the model convergence. When the mean square error between the predicted and actual relative deviation values ​​on the validation set is less than 0.001 and remains stable for 10 consecutive rounds, the model is considered to have converged, and the base model is obtained.

[0052] Furthermore, in order to address the issue of poor interpretability in gradient boosting tree-based models, and to transform the model's "prediction results" into "the contribution degree of each factor," allowing users to clearly understand the specific impact percentage of each key influencing factor on the relative deviation, it is necessary to integrate the SHAP or LIME interpretation framework onto the trained base model to construct a contribution calculation model.

[0053] First, it is necessary to obtain an existing SHAP or LIME interpretation framework. SHAP is suitable for global interpretation, which can quantify the average contribution of each feature to all samples; while LIME is suitable for local interpretation, which can focus on the feature influence of a single sample.

[0054] Then, the trained gradient boosting tree base model is connected to the selected interpretation framework. Using the framework's built-in algorithms, such as SHAP value calculation and local linear regression, the "contribution value" of each input feature (historical key influencing factor) to the prediction result is analyzed when the base model predicts relative bias. A positive contribution indicates that the factor aggravates the bias, while a negative contribution indicates that it suppresses the bias.

[0055] Finally, the predictive power of the base model and the contribution analysis capability of the explanatory framework are integrated to form the final contribution calculation model. After inputting key influencing factors, this model can output both the predicted value of the relative deviation and the specific contribution of each key influencing factor to the deviation. For example, when the selected key influencing factors are input into the pre-trained contribution calculation model, the output relative deviation prediction value is 11.5%, and the contribution values ​​are turbulence intensity 45%, gearbox oil temperature 30%, and cumulative running time 20%.

[0056] Furthermore, the contributions need to be sorted in descending order, and the top N key influencing factors should be selected as the root causes of power prediction deviations. For example, the contributions of each key influencing factor can be sorted from high to low, and the top N key influencing factors can be selected as the final root causes of deviations. These factors are the main reasons for the current power prediction deviations.

[0057] In step S300 of this application embodiment, the calculation step of the N value includes: Obtain the average relative prediction deviation of the digital twin within a preset historical window; The prediction confidence of the digital twin is calculated based on the average relative prediction deviation. The prediction confidence level is calculated by multiplying it by the total number of key influence factors in a set of key influence factors and then rounding down to obtain the N value.

[0058] First, it is necessary to obtain the average prediction relative deviation of the digital twin within a preset historical window. For example, if the preset historical window is set to the past 30 days, all historical predicted power of the digital twin and the historical actual power of the corresponding physical device within this window are retrieved. The relative deviation of each historical data is calculated according to the relative deviation formula in step S200, and then the arithmetic mean of all relative deviations is calculated to obtain the average prediction relative deviation.

[0059] Next, the prediction confidence level of the digital twin needs to be calculated based on the average relative prediction deviation. Specifically, a reverse mapping relationship can be used to calculate the prediction confidence level. For example, the formula can be "Prediction Confidence Level = 1 - Average Relative Prediction Deviation," and the calculation result should be controlled between 0 and 1. When the average relative prediction deviation ≥ 1, the confidence level is 0; when the average relative prediction deviation ≤ 0, the confidence level is 1. For example, if the average relative prediction deviation is 8% (0.08), then the prediction confidence level is 0.92.

[0060] Then, the prediction confidence level needs to be calculated by multiplying it by the total number of key influence factors in a set of key influence factors and rounding down to obtain the N value.

[0061] Specifically, it is necessary to count the total number of key influencing factors screened through correlation analysis in step S300, such as 5. Multiply the total number of key influencing factors by the prediction confidence level, and round the result to the nearest integer, either by rounding to the nearest whole number or by rounding up, to ensure that N is a positive integer and does not exceed the total number of key influencing factors. For example, if the prediction confidence level is 0.92 and the total number of key influencing factors is 5, the product is 4.6, and the rounded N is 5; if the prediction confidence level is 0.5 and the total number of key influencing factors is 5, the product is 2.5, and the rounded N is 3.

[0062] In step S400 of this application embodiment, based on the root cause of the deviation and combined with the performance degradation status data of the physical entity of the device, a correction coefficient is generated through a quantization correction model, including: Invoke the pre-trained quantization correction model; Obtain the monitoring variable data corresponding to the root cause of the deviation, and combine it with the performance degradation status data to form an input feature vector, wherein the performance degradation status data includes cumulative running time and historical maintenance frequency; The input feature vector is input into the quantization correction model, and the correction coefficient is output.

[0063] In this embodiment, the core objective of step S400 is to generate quantization correction coefficients that can accurately compensate for power prediction deviations. To achieve this step, a quantization correction model must first be trained and obtained.

[0064] In step S400 of this application embodiment, the training step of the quantization correction model includes: Construct a quantitative correction model; Collect historical monitoring variable data and historical performance degradation status data at several historical moments as a sample training dataset; Obtain the actual power generation and predicted power generation at the corresponding historical moment, calculate several historical correction coefficients, and use them as a sample label dataset. Supervised training is performed using the sample training dataset and the sample label dataset. Once the convergence condition is met, a quantized correction model is obtained.

[0065] First, a quantitative correction model needs to be constructed. Considering the task type of the quantitative correction model, the random forest regression algorithm can be used to build the quantitative correction model.

[0066] Optionally, in the parameter settings of the quantization correction model, the number of decision trees is 150. The maximum depth of each decision tree is 8 layers. The minimum number of splits per decision tree is 25, meaning that splitting stops when the number of node samples is less than 25, reducing the model's sensitivity to noisy data. The minimum number of leaf nodes per decision tree is 5. The feature subset selection method is to randomly select 70% of all input features for splitting during each tree construction, enhancing model diversity. The random seed is set to 123 to ensure the reproducibility of model training results. The bootstrap sampling ratio uses an 80% sample sampling ratio. The regularization parameter is set to 0.05 to further suppress overfitting by slightly penalizing complex models.

[0067] Furthermore, it is necessary to prepare sample data for training the quantitative correction model, namely, to collect historical monitoring variable data and historical performance degradation status data at several historical moments as sample training datasets; to obtain the actual power generation and predicted power generation at the corresponding historical moments, and to calculate several historical correction coefficients as sample label datasets.

[0068] Specifically, it is necessary to collect relevant data from the equipment's historical database at several historical moments, including: historical monitoring variable data, i.e., the values ​​of monitoring variables corresponding to the root causes of historical deviations, such as historical turbulence intensity and historical gearbox oil temperature; and historical performance degradation status data, i.e., the cumulative operating time and historical maintenance frequency corresponding to the historical moment. This data is then organized into a structured sample training dataset along the time dimension, with each sample corresponding to a complete input feature for a historical moment.

[0069] Then, for each historical sample in the sample training dataset, the actual power generation at the corresponding historical moment and the predicted power generation of the digital twin are retrieved, and the historical correction coefficient is calculated based on the difference between the two. For example, the calculation method can be: historical correction coefficient = historical actual power generation / historical predicted power generation). All historical correction coefficients are organized into a sample label dataset, which corresponds one-to-one with the sample training dataset.

[0070] A total of no less than 12,000 sets of sample data were collected to form a sample training dataset and a sample label dataset, which were then divided into a validation set and a training set in a 2:8 ratio.

[0071] Finally, the sample training dataset is used as the model input, and the sample label dataset (historical correction coefficients) is used as the training target to conduct supervised training on the quantization correction model.

[0072] When the "mean absolute error between the predicted correction coefficient and the actual historical correction coefficient" of the model on the validation set is less than 0.02, and after five consecutive fine-tunings of the model parameters, the error value remains stable within 0.02, the model is considered to have reached convergence. Training is then stopped and the final model parameters are saved to obtain the quantized correction model.

[0073] Furthermore, the monitoring variable data corresponding to the root cause of the deviation are obtained, and the input feature vector is constructed by combining it with the performance degradation status data, wherein the performance degradation status data includes cumulative running time and historical maintenance frequency; the input feature vector is input into the quantization correction model, and the correction coefficient is output.

[0074] Specifically, based on the root cause of the deviation determined in step S300, the current value of the corresponding monitored variable needs to be retrieved from the equipment's real-time monitoring system. For example, if the root cause of the deviation is "turbulence intensity + cumulative running time", then the current turbulence intensity monitoring value and cumulative running time data should be extracted.

[0075] Then, from the historical database of the wind farm operation and maintenance management system, retrieve the current equipment's "cumulative running time," which is the total running time from the equipment's commissioning to the current moment; and the "historical maintenance frequency," which is the total number of maintenance operations since the equipment's commissioning. These two data directly reflect the performance degradation of the equipment due to aging and wear.

[0076] The aforementioned monitoring variable data and performance degradation status data are integrated into a one-dimensional structured data vector in a preset order. For example, if the root cause of the deviation is "turbulence intensity (current value 1.2m)", then... 2 / s 2 If the gearbox oil temperature is 58℃, then the performance degradation status data is "cumulative running time 8760 hours" and "historical maintenance frequency 3 times". The input feature vector is [1.2,58,8760,3], ensuring that the data format meets the input requirements of the quantization correction model.

[0077] Finally, the input feature vector is fed into the quantization correction model, which outputs correction coefficients, such as 1.08 or 0.95. These correction coefficients directly correspond to the correction magnitude of the digital twin's predicted power generation.

[0078] In step S500 of this embodiment, the predicted equipment parameters of the digital twin need to be corrected according to the correction coefficient to generate corrected equipment parameters. For example, if the correction coefficient output by the quantization correction model is 1.08, it means that the current predicted power generation of the digital twin needs to be increased by 8%; if the correction coefficient is 0.95, then the current predicted power generation of the digital twin needs to be decreased by 5%. The adjusted predicted power generation is used as the corrected equipment parameter to achieve accurate correction of the prediction deviation.

[0079] Finally, as described in step S600 of the embodiment of this application, the modified device parameters are sent to the corresponding monitoring system to complete the two-way communication closed loop.

[0080] Specifically, the modified device parameters are encrypted and encapsulated and sent to the monitoring system via industrial Ethernet or a 5G private network using the MQTT protocol. Upon receiving the parameters, the monitoring system parses them, updates the display interface, and synchronizes them with the control module. Simultaneously, the monitoring system sends a confirmation signal to the digital twin, forming a closed-loop two-way data flow between the physical device, the digital twin, and the monitoring system, ensuring seamless integration of command execution and status feedback.

[0081] In step S300 of the embodiments of this application, the following is also included: If the relative deviation does not exceed a preset error threshold, the predicted output of the digital twin is determined to be reliable, and the predicted power generation is sent to the wind farm monitoring system as the final output value, completing the two-way communication loop. When the relative deviation does not exceed the preset error threshold, the system automatically determines the prediction to be reliable. The predicted power generation of the digital twin is directly extracted and transmitted to the wind farm monitoring system through a secure communication protocol. After receiving the data, the monitoring system updates the power prediction dashboard and sends a successful reception signal back to the digital twin, completing the two-way communication loop.

[0082] In summary, by providing embodiments of this application, at least the following can be achieved: 1. Significantly improves the prediction accuracy and precision of digital twins; 2. By quantifying prediction bias, accurately locating the root cause of the bias, and combining the equipment performance degradation status to generate scientific correction coefficients and correct the digital twin, the problem of model drift is effectively solved, and the high prediction accuracy of the digital twin is maintained continuously. 3. By using error threshold judgment, correlation analysis, and the construction of contribution calculation models, the root causes of prediction deviations can be accurately located, avoiding excessive processing of small deviations and wasting resources, while also preventing blind corrections that could lead to the expansion of problems.

[0083] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0084] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0085] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0089] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method of twin scene bidirectional communication, characterized in that, The method comprises the following steps: Collecting multi-source real-time operation data and state data of a physical entity of a device, and uploading the multi-source real-time operation data and state data to a corresponding digital twin; Calculating a relative deviation based on predicted device parameters of the digital twin and actual device parameters of the physical entity of the device; Judging whether the relative deviation exceeds a preset error threshold, and if so, starting a difference analysis of the relative deviation to determine a deviation root cause leading to the prediction deviation; Based on the deviation root cause, combining performance degradation state data of the physical entity of the device, generating a correction coefficient through a quantitative correction model; Correcting the predicted device parameters of the digital twin according to the correction coefficient to generate corrected device parameters; Downloading the corrected device parameters to a corresponding monitoring system to complete a two-way communication closed loop.

2. The method of claim 1, wherein, The physical entity of the device includes a physical entity of a wind power generation device, the device parameters include power generation power, the multi-source real-time operation data include environmental data and operation parameters, and the state data include device state data and historical operation and maintenance data, wherein the environmental data at least includes wind speed, wind direction, air temperature, turbulence intensity and air density, the operation parameters at least include generator output power, generator speed and pitch angle, and the device state data at least includes main shaft bearing temperature, gearbox oil temperature, generator winding temperature and cabin vibration spectrum data, and the historical operation and maintenance data at least includes cumulative running time and historical maintenance frequency.

3. The method of claim 2, wherein, Starting the difference analysis of the relative deviation to determine the deviation root cause leading to the power prediction deviation comprises: Performing correlation analysis according to the multi-source real-time operation data and state data to identify a group of key influence factors; Inputting the group of key influence factors into a pre-trained contribution degree calculation model to calculate the contribution degree of each key influence factor to the current power prediction deviation; Arranging the contribution degrees in descending order, and selecting the top N key influence factors as the deviation root cause leading to the power prediction deviation.

4. The method of claim 3, wherein, Performing correlation analysis according to the multi-source real-time operation data and state data to identify a group of key influence factors comprises: Calculating correlation coefficients between the relative deviation and a plurality of monitoring variables, wherein the monitoring variables at least include wind speed, wind direction, turbulence intensity, pitch angle, gearbox oil temperature, generator winding temperature and cumulative running time; Screening monitoring variables with correlation coefficients greater than a first preset threshold to form a group of key influence factors.

5. The method of claim 3, wherein, The training steps of the contribution degree calculation model comprise: Building a base model based on gradient boosting trees; Collecting historical multi-source real-time operation data and state data as a historical data sample set, and identifying historical key influence factors of the historical data samples; Obtaining relative deviation values of historical actual power and historical predicted power corresponding to the historical data samples; Supervised training of the base model with the historical key influence factors as input features and the corresponding relative deviation values as prediction targets; Integrating SHAP or LIME interpretation framework on the trained base model to build the contribution degree calculation model.

6. The method of claim 3, wherein, The calculation steps of the N value comprise: acquiring an average prediction relative deviation of the digital twin within a preset historical window; calculating a prediction confidence of the digital twin according to the average prediction relative deviation; calculating a product of the prediction confidence and a total number of key influence factors, and rounding to obtain an N value.

7. The method of claim 2, wherein, Based on the deviation root cause, combined with the performance degradation state data of the equipment physical entity, a quantitative correction model is used to generate a correction coefficient, including: calling a pre-trained quantitative correction model; acquiring monitoring variable data corresponding to the deviation root cause, and combining the performance degradation state data to form an input feature vector, wherein the performance degradation state data includes cumulative running time and historical maintenance frequency; inputting the input feature vector into the quantitative correction model to output a correction coefficient.

8. The method of claim 7, wherein, The training steps of the quantitative correction model include: constructing a quantitative correction model; collecting historical monitoring variable data and historical performance degradation state data at several historical moments as sample training data sets; acquiring actual and predicted power generation at corresponding historical moments to calculate several historical correction coefficients as sample label data sets; based on the sample training data sets and the sample label data sets, performing supervised training, and obtaining the quantitative correction model when the convergence condition is met.

9. The method of claim 2, wherein, Further comprising: if the relative deviation does not exceed the preset error threshold, it is determined that the prediction output of the digital twin is reliable, the predicted power generation is taken as the final output value and is issued to the wind farm monitoring system to complete the two-way communication closed loop.

Images