Wind power plant environment adaptive power generation capacity dynamic evaluation method and system

Abstract

The embodiment of the invention relates to the technical field of power generation capacity management, and provides a power generation capacity dynamic evaluation method and system adaptive to a wind power plant environment. According to the method, a multi-modal guiding and coupling mechanism is used for dynamically guiding and analyzing time sequence relevance hidden in other rich feature data, a unified state representation vector deeply fused with a working condition context can be generated, and the vector not only improves the precision of power prediction, enables anomaly detection based on power deviation to be more reliable, but also improves the accuracy of power prediction. The method also becomes a data packet with highly concentrated information, and provides high-quality input for a subsequent attribution classifier, thereby realizing closed-loop monitoring from accurate state evaluation to accurate fault traceability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power generation capacity management, and particularly relates to a wind farm environment adaptive power generation capacity dynamic evaluation method and system. BACKGROUND

[0002] As a clean and renewable energy, wind energy plays an increasingly important role in the transformation of global energy structure. Wind turbine, as the core equipment of wind energy conversion, is usually installed in remote and harsh environment. Its structure is complex, value is high and it is in harsh operating conditions for a long time. Therefore, in order to ensure the safe and stable operation of wind turbine, maximize the power generation efficiency and reduce the operation and maintenance cost, it is very important to develop an efficient and reliable remote monitoring and fault diagnosis system. Programmable logic controller (PLC) is the core of the main control system of wind turbine, which collects and controls the key parameters of wind turbine operation in real time, and constitutes the source of massive supervisory control and data acquisition (SCADA) data. Therefore, it is an inevitable choice and mainstream direction of technical development to build a remote monitoring scheme based on PLC in the industry.

[0003] In the existing practice of wind farm environment adaptive generation capacity dynamic assessment, a mode of transmitting SCADA data collected by PLC to a remote monitoring center through a network, and then performing centralized processing and analysis by a server of the remote monitoring center is adopted. However, this traditional generation capacity assessment mode has technical bottlenecks. First, the judgment of faults depends on setting static thresholds for a single parameter, and an alarm is triggered when a parameter (such as cabin temperature) exceeds the preset limit. This way ignores the deep coupling relationship between parameters in a complex dynamic system, cannot effectively identify early performance degradation or potential faults caused by multi-variable coordination, and is prone to a large number of false positives and false negatives. Secondly, in the data processing layer, the existing generation capacity assessment mode often simply splices or directly inputs sensors from different wind farm environments and data of different natures (such as unit operating parameters and external meteorological data) into the model, and cannot fully utilize the internal correlation between the data. This processing method makes it difficult for the model to discover the fine running mode of the wind turbine under specific working conditions, thereby affecting the accuracy of the health status assessment of the wind turbine. The power generation of the wind turbine is the most direct manifestation of its comprehensive health status, but the accuracy of its prediction model depends heavily on a deep understanding of multi-dimensional input features. However, the existing prediction model cannot effectively fuse multi-source heterogeneous data, which will make the power prediction result not accurate enough, and make it difficult for the abnormality detection method based on power deviation to accurately identify whether the power generation capacity loss is caused by normal working condition fluctuation or specific faults. Therefore, an optimized generation capacity dynamic assessment scheme is expected. SUMMARY

[0004] The present application aims to at least solve one of the problems existing in the prior art, and provide a wind farm environment adaptive generation capacity dynamic assessment method and system.

[0005] In one aspect of the present application, a wind farm environment adaptive generation capacity dynamic assessment method is provided, comprising: obtaining a real-time SCADA data stream of a wind turbine and a real-time meteorological data stream of a wind farm environment, wherein the real-time SCADA data stream comprises wind speed, active power, pitch angle, rotor speed and cabin temperature, and the real-time meteorological data stream comprises environmental temperature, air pressure and humidity; performing data preprocessing and feature enrichment on the real-time SCADA data stream and the real-time meteorological data stream to obtain rich feature data; calculating power deviation data based on the rich feature data and a trained power prediction model; starting an attribution classifier to perform fault diagnosis based on the power deviation data to obtain a diagnosis result; performing quantitative assessment based on generation capacity loss based on the diagnosis result and the power deviation data to generate alarm information and an assessment report.

[0006] Optionally, the real-time SCADA data stream and the real-time meteorological data stream are data pre-processed and feature enriched to obtain feature-rich data, including: time aligning, cleaning and invalid data filtering the real-time SCADA data stream and the real-time meteorological data stream to obtain cleaned joint data; calculating real-time air density and normalized power based on the cleaned joint data, and integrating them into the cleaned joint data to obtain the feature-rich data.

[0007] Optionally, calculating real-time air density and normalized power based on the cleaned joint data includes: calculating real-time air density based on cleaned ambient temperature, cleaned air pressure and cleaned humidity, according to the following air density formula: ; wherein, is real-time air density, is cleaned air pressure, is cleaned ambient temperature, is water vapor partial pressure, and , is cleaned humidity, is saturated water vapor pressure, is specific gas constant of dry air, is specific gas constant of water vapor.

[0008] Optionally, calculating real-time air density and normalized power based on the cleaned joint data further includes: normalizing and correcting active power in the cleaned joint data based on the real-time air density, according to the following correction formula to obtain the normalized power: ; wherein, is normalized power, is active power in the cleaned joint data, is standard air density, is real-time air density.

[0009] Optionally, calculating power deviation data based on the feature-rich data and the trained power prediction model includes: passing the feature-rich data through the trained power prediction model to obtain predicted power of the wind turbine under current working condition; calculating the difference between the predicted power and the normalized power in the feature-rich data as the power deviation data.

[0010] Optionally, the rich feature data is input into the trained power prediction model to obtain a predicted power of the wind turbine under the current working condition, including: extracting a time sequence of active power from the rich feature data; constructing other data in the rich feature data into a rich feature data matrix according to time dimension and data sample dimension, except for the time sequence of active power; inputting the time sequence of active power and the rich feature data matrix into the trained power prediction model to obtain a power time sequence representation vector under rich feature guidance; decoding the power time sequence representation vector under rich feature guidance to obtain the predicted power.

[0011] Optionally, inputting the time sequence of active power and the rich feature data matrix into the trained power prediction model to obtain a power time sequence representation vector under rich feature guidance, includes: vectorizing and encoding the time sequence of active power to extract active power time sequence features to obtain an active power time sequence mode encoding vector; performing implicit time sequence correlation coding on the rich feature data matrix to obtain a rich feature implicit correlation feature map; performing multi-modal guided time sequence correlation analysis on the active power time sequence mode encoding vector and the rich feature implicit correlation feature map to obtain the power time sequence representation vector under rich feature guidance.

[0012] Optionally, performing multi-modal guided time sequence correlation analysis on the active power time sequence mode encoding vector and the rich feature implicit correlation feature map to obtain the power time sequence representation vector under rich feature guidance, includes: vectorizing and reconstructing the rich feature implicit correlation feature map along the channel dimension to obtain a set of rich feature implicit correlation encoding vectors; taking the active power time sequence mode encoding vector as a context modulation signal, constructing a time-varying context correlation matrix of the set of rich feature implicit correlation encoding vectors to obtain a rich feature implicit correlation time-varying context correlation matrix; inputting the set of rich feature implicit correlation encoding vectors and the rich feature implicit correlation time-varying context correlation matrix into a graph convolutional neural network model to obtain a rich feature implicit correlation context enhanced representation matrix; performing multi-modal time sequence coupling on the rich feature implicit correlation context enhanced representation matrix and the active power time sequence mode encoding vector to obtain the power time sequence representation vector under rich feature guidance.

[0013] Optionally, based on the power deviation data, starting an attribution classifier to perform fault diagnosis to obtain a diagnosis result, including: performing an attribution diagnosis to obtain a diagnosis result when an absolute value of the power deviation data is greater than a preset abnormality judgment threshold value; The attribution diagnosis comprises: adding the power deviation data to a tail end of a power time sequence representation vector under rich feature guidance to obtain an attribution feature vector; passing the attribution feature vector through an attribution classifier to obtain the diagnosis result, the diagnosis result comprising an attribution diagnosis type label and a confidence thereof.

[0014] In another aspect of the present application, a wind farm environment adaptive power generation capacity dynamic evaluation system is provided, comprising: a real-time data acquisition module, configured to acquire a real-time SCADA data stream of a wind turbine and a real-time meteorological data stream of a wind farm environment, the real-time SCADA data stream comprising wind speed, active power, pitch angle, rotor speed and nacelle temperature, and the real-time meteorological data stream comprising ambient temperature, air pressure and humidity; a rich feature data generation module, configured to perform data preprocessing and feature enrichment on the real-time SCADA data stream and the real-time meteorological data stream to obtain rich feature data; a power deviation calculation module, configured to calculate power deviation data based on the rich feature data and a trained power prediction model; a fault diagnosis module, configured to start an attribution classifier to perform fault diagnosis based on the power deviation data to obtain a diagnosis result; a quantitative evaluation module, configured to perform power generation capacity loss-based quantitative evaluation based on the diagnosis result and the power deviation data to generate alarm information and an evaluation report.

[0015] Compared with the prior art, the wind farm environment adaptive power generation capacity dynamic evaluation method and system provided by the present application breaks the limitation of regarding all monitoring data as equal input in the traditional method, no longer performs simple feature splicing, but uses active power time sequence data as the core mode. By constructing a new power prediction model, the internal mode of the power sequence is used as the context prior information to dynamically guide and analyze the hidden time sequence correlation in other rich feature data (such as wind speed, pitch angle, meteorological parameters, etc.). This multi-modal guidance and coupling mechanism can generate a unified state representation vector deeply fused with the working condition context. This vector not only improves the accuracy of power prediction, making the abnormality detection based on power deviation more reliable, but also becomes a highly condensed data packet itself, providing high-quality input for the subsequent attribution classifier, thereby realizing closed-loop monitoring from accurate state evaluation to accurate fault tracing. BRIEF DESCRIPTION OF DRAWINGS

[0016] One or more embodiments are illustrated by way of example in the drawings and specification hereof, which constitute part of this patent document. Illustrative implementations of one or more embodiments are described in detail herein. Those of ordinary skill in the art will appreciate that the implementations described herein are exemplary and not the only way(s) to implement one or more embodiments. Other ways to implement the one or more embodiments will be apparent to those of ordinary skill in the art upon review of the specification.

[0017] Figure 1 A flow chart of a wind farm environment adaptive power generation capacity dynamic evaluation method according to an embodiment of the present application; Figure 2 A data flow schematic diagram of a wind farm environment adaptive power generation capacity dynamic evaluation method according to an embodiment of the present application; Figure 3 A flow chart of calculating power deviation data based on the rich feature data and the trained power prediction model of a wind farm environment adaptive power generation capacity dynamic evaluation method according to an embodiment of the present application; Figure 4 A flow chart of passing the rich feature data through the trained power prediction model to obtain the predicted power of the wind turbine under the current working condition of a wind farm environment adaptive power generation capacity dynamic evaluation method according to an embodiment of the present application; Figure 5 A block diagram of a wind farm environment adaptive power generation capacity dynamic evaluation system according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] As shown in the specification and claims of the present application, unless otherwise specified, "a," "an," "the," and / or "at least one" are used interchangeably and mean one or more than one of something. In general, the term "comprises" or "comprising" is used throughout the specification and claims to mean that the method or device includes the recited steps or elements but not excluding others. The term "coupled" is used herein to express a relationship between or among multiple elements and does not require that the elements be directly connected to one another.

[0019] Although the present application makes various references to certain modules in the system according to embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server. The modules are illustrative only and different aspects of the system and method can use different modules.

[0020] Flow charts are used in the present application to illustrate the operations performed by the system according to embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in the exact order shown. Rather, various steps can be handled in reverse order, or simultaneously, as desired. Other operations can also be added to, or removed from, these processes.

[0021] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and the present application is not limited in these example embodiments described herein.

[0022] To solve the problem of insufficient precision of the power prediction model due to the failure to effectively fuse multi-source heterogeneous data in the prior art, the technical solution of the present application proposes a wind farm environment adaptive power generation capacity dynamic evaluation scheme. Specifically, the technical solution of the present application first starts from the data source, after obtaining the wind turbine SCADA data and the wind farm meteorological data, instead of directly using the obtained data, but pre-processing and feature enrichment of key data, especially through real-time calculation of air density to standardize and correct the active power, so as to eliminate the direct influence of environmental factors on power output, and establish a fair and accurate comparison benchmark for subsequent model. On this basis, a power prediction model with multi-modal guidance mechanism is constructed. The power prediction model no longer regards all input features as equal status, but takes the standardized active power time series as the core mode, and constructs the multi-dimensional data including wind speed, pitch angle and rotating speed into a feature-rich data matrix as the context auxiliary mode. The power prediction model encodes the two kinds of mode data through a deep learning network, then uses the encoding vector of the active power time series mode as a guide signal to perform dynamic context-aware correlation analysis and feature enhancement on the feature-rich data matrix, and finally couples the two deeply fused information streams to generate a feature-rich guided power time series representation vector with high information density and accurate reflection of the current working condition of the unit health status. Using this vector, the power prediction model can output a highly accurate predicted power, and the deviation of the predicted power from the actual standardized power becomes a reliable basis for anomaly detection. Once the deviation is out of limit, the representation vector will be directly used as the input of the fault attribution classifier, so as to not only judge whether the wind turbine has failed, but also accurately diagnose the specific fault type, so as to generate an alarm information and an evaluation report, realize the whole-chain closed loop from data acquisition, accurate prediction to intelligent diagnosis, and solve the technical problems of low fault diagnosis reliability and difficult traceability in the prior art.

[0023] The technical solution of the present application proposes a wind farm environment adaptive power generation capacity dynamic evaluation method. Figure 1 The flowchart of the wind farm environment adaptive power generation capacity dynamic evaluation method according to the embodiment of the present application. Figure 2 The data flow schematic diagram of the wind farm environment adaptive power generation capacity dynamic evaluation method according to the embodiment of the present application. Figure 1 and Figure 2As shown, the wind farm environment adaptive power generation capacity dynamic evaluation method according to the embodiment of the application comprises the following steps: S100, obtaining a real-time SCADA data stream of a wind turbine and a real-time meteorological data stream of a wind farm environment, wherein the real-time SCADA data stream comprises wind speed, active power, pitch angle, rotor speed and nacelle temperature, and the real-time meteorological data stream comprises environmental temperature, air pressure and humidity; S200, performing data preprocessing and feature enrichment on the real-time SCADA data stream and the real-time meteorological data stream to obtain rich feature data; S300, calculating power deviation data based on the rich feature data and a trained power prediction model; S400, starting an attribution classifier to perform fault diagnosis based on the power deviation data to obtain a diagnosis result; and S500, performing quantitative evaluation based on power generation capacity loss based on the diagnosis result and the power deviation data to generate alarm information and an evaluation report.

[0024] Specifically, in step S100, a real-time SCADA data stream of a wind turbine and a real-time meteorological data stream of a wind farm environment are obtained, wherein the real-time SCADA data stream comprises wind speed, active power, pitch angle, rotor speed and nacelle temperature, and the real-time meteorological data stream comprises environmental temperature, air pressure and humidity. It can be understood that the power generation performance of a wind turbine is not only dependent on its own operating state, but also deeply affected by external environmental factors, especially key meteorological parameters such as air density, which directly determine the amount of wind energy. If only the SCADA data of the wind turbine itself is relied on, it will not be possible to accurately distinguish between normal power fluctuations caused by environmental changes and abnormal power losses caused by performance degradation of the wind turbine. Therefore, the embodiment of the application further obtains a real-time SCADA data stream of a wind turbine and a real-time meteorological data stream of a wind farm environment, wherein the real-time SCADA data stream comprises wind speed, active power, pitch angle, rotor speed and nacelle temperature, and the real-time meteorological data stream comprises environmental temperature, air pressure and humidity, so as to construct a multi-dimensional feature space that can comprehensively represent the real-time operating conditions of the wind turbine. In this way, a complete and accurate data basis can be provided for the subsequent training of the power prediction model and fault diagnosis, ensuring the accuracy and reliability of the analysis results.

[0025] More specifically, in one specific example of the present application, the data acquisition process is implemented through a data acquisition and communication system deployed within the wind farm. First, a data acquisition gateway is deployed inside the tower base or nacelle of the wind turbine, which is connected to the main control PLC of the wind turbine through an industrial Ethernet. Using the OPC UA protocol, the data acquisition gateway periodically reads the register addresses of key variables such as wind speed, active power, pitch angle, rotor speed and nacelle temperature in the main control PLC, thereby obtaining a high-frequency, real-time SCADA data stream. At the same time, an integrated meteorological monitoring station is installed outside the wind measurement tower or designated wind turbine nacelle of the wind farm, which contains temperature sensors, barometers and hygrometers. These sensors transmit the collected environmental temperature, barometric pressure and humidity signals to a data logger, which then outputs a structured meteorological data stream. Finally, the SCADA data stream and the meteorological data stream are collected to the central monitoring server or edge computing node within the wind farm, where the two data streams are aligned with a unified timestamp and packaged into a standard data packet format. Through the dedicated fiber ring network or 5G wireless network of the wind farm, the integrated real-time data stream is continuously transmitted to the data server of the remote monitoring center for subsequent processing and analysis.

[0026] Specifically, in step S200, the real-time SCADA data stream and the real-time meteorological data stream are pre-processed and feature-enriched to obtain rich feature data. It should be understood that the original data stream directly obtained from different sources such as PLC and meteorological station is not synchronized in time, and inevitably contains noise, outliers or missing data due to communication interruption. At the same time, core performance indicators such as active power are directly affected by time-varying environmental factors such as air density. If not processed directly for analysis, the internal health status of the wind turbine cannot be accurately evaluated, and normal operating condition fluctuations may be misjudged as faults. Therefore, in the technical solution of the present application, the real-time SCADA data stream and the real-time meteorological data stream are further pre-processed and feature-enriched to obtain rich feature data, thereby constructing a unified data set that is time-synchronized, data-pure and information-dimension-enhanced. In particular, by calculating standardized power and other derived features, the direct interference of environmental changes on power generation performance is actively decoupled and stripped. In this way, a high-information-content input is provided for the subsequent power prediction model, ensuring that the power prediction model learns the true performance baseline of the wind turbine under a specific operating strategy, rather than the apparent performance masked by environmental factors, thereby ensuring accurate calculation of power deviation and reliable fault diagnosis.

[0027] More specifically, in the embodiments of the present application, the real-time SCADA data stream and the real-time meteorological data stream are data pre-processed and feature enriched to obtain feature-rich data, including: time alignment, cleaning and invalid data filtering processing of the real-time SCADA data stream and the real-time meteorological data stream to obtain cleaned joint data; calculating real-time air density and standardized power based on the cleaned joint data, and integrating them into the cleaned joint data to obtain feature-rich data.

[0028] Specifically, the real-time SCADA data stream and the real-time meteorological data stream are time aligned, cleaned and invalid data filtered to obtain cleaned joint data. It should be understood that due to the inherent differences in data generation frequency and timestamp accuracy of the real-time SCADA data stream and the real-time meteorological data stream collected from different physical devices, the original data cannot be directly aligned in the time dimension. At the same time, the data is susceptible to electromagnetic interference, sensor drift or communication failure during collection and transmission, which can produce physically unreasonable outliers, data missing and invalid data that do not match the actual operation condition of the wind turbine. Therefore, in the technical solution of the present application, the real-time SCADA data stream and the real-time meteorological data stream are further time aligned, cleaned and invalid data filtered to obtain cleaned joint data, so as to construct a data set with unified time reference, accurate data value and true reflection of wind turbine power generation behavior. In this way, the interference of original data quality problems on subsequent analysis can be fundamentally eliminated, thereby ensuring the accuracy of the entire monitoring scheme.

[0029] More specifically, in one specific example of the present application, the data pre-processing process is automatically executed on the data server of the remote monitoring center. First, the time alignment operation is performed, and a fixed time interval, for example 10 minutes, is used as the unified time resolution reference. All SCADA data points and meteorological data points received within this time interval are aggregated, and the arithmetic mean value of continuous variables such as wind speed, active power and temperature is calculated as the representative value of the time point, such as 10 minutes, so as to normalize the two different frequency data streams to the same time grid. Then, the data cleaning process is started, and the data cleaning process automatically identifies and removes abnormal data points that violate the rules, such as wind speed values greater than or equal to zero and active power values between zero and rated power, according to the pre-set physical constraint rules. Subsequently, invalid data filtering is performed, and all data records corresponding to abnormal power generation conditions such as shutdown, failure, maintenance or manual power limiting are filtered out by analyzing the wind turbine status code contained in the SCADA data. After the above steps, the processed SCADA data and meteorological data are merged according to the aligned time stamp to form a structured data table, i.e. the cleaned joint data.

[0030] Specifically, the real-time air density and the normalized power are calculated based on the cleaned joint data, and are integrated into the cleaned joint data to obtain the rich feature data. It should be understood that, due to the active power index in the cleaned joint data, the numerical size thereof not only reflects the health status of the wind turbine itself, but is also closely coupled with the air density determined by the real-time changing environmental temperature, air pressure and humidity. This coupling effect causes that, under different weather conditions, even if the wind speed and the state of the wind turbine are completely the same, the actual power output will also exist differences, thereby covering the real power loss caused by performance degradation or early failure, so that the performance evaluation directly based on the original power loses comparability and accuracy. Therefore, in the technical scheme of the present application, the real-time air density and the normalized power are further calculated based on the cleaned joint data, and are integrated into the cleaned joint data to obtain the rich feature data, so as to actively quantify and decouple the nonlinear influence of environmental factors on power generation performance by constructing a physical model driven derived feature. In this way, a normalized power index that eliminates environmental fluctuation interference and only reflects the inherent conversion efficiency of the wind turbine can be generated, which serves as a core learning target and a comparison benchmark of a subsequent power prediction model, thereby providing data support for accurately identifying slight performance degradation and realizing high-sensitivity fault diagnosis.

[0031] More specifically, in the embodiment of the present application, the real-time air density and the normalized power are calculated based on the cleaned joint data and integrated into the cleaned joint data to obtain the rich feature data, and the calculation of the real-time air density and the normalized power based on the cleaned joint data comprises: based on the cleaned environmental temperature, the cleaned air pressure and the cleaned humidity, the real-time air density is calculated according to the following air density formula: .

[0032] wherein, is the real-time air density, is the cleaned air pressure, is the cleaned environmental temperature, is the water vapor partial pressure, and , is the cleaned humidity, is the saturated water vapor pressure, is the specific gas constant of dry air, is the specific gas constant of water vapor. It should be understood that, since air density is a key physical quantity determining the size of wind energy, it directly affects the power output of the wind turbine, and the air density value is not a constant, but changes dynamically in real time with changes in environmental temperature, air pressure and humidity. If the change is ignored and the standard air density is used for analysis, a physical model error will be introduced, and the actual impact of environmental conditions on power generation performance cannot be accurately quantified. Therefore, in the technical solution of the present application, the real-time air density is calculated based on the cleaned environmental temperature, cleaned air pressure and cleaned humidity using the air density formula, so as to convert the discrete meteorological measurement parameters into a comprehensive physical characteristic that can accurately represent the current wind energy level. In this way, an accurate and dynamic physical benchmark parameter can be provided for subsequent power standardization correction, and a physical foundation is laid for building a high-fidelity power prediction model.

[0033] Based on the cleaned joint data, the real-time air density and the standardized power are calculated, and the method further comprises: based on the real-time air density, the active power in the cleaned joint data is standardized and corrected by using the following correction formula to obtain the standardized power: .

[0034] wherein, is the standardized power, is the active power in the cleaned joint data, is the standard air density, is the real-time air density. It should be understood that, since the actual active power of the wind turbine is the product of its inherent conversion efficiency and the real-time air density, if the effects of the two are not separated, the normal power fluctuation caused by environmental changes will be confused with the abnormal power drop caused by performance degradation of the wind turbine, so that a stable and comparable performance benchmark cannot be established, thereby seriously affecting the sensitivity and accuracy of fault diagnosis. Therefore, in the technical solution of the present application, the active power in the cleaned joint data is standardized and corrected by using the correction formula based on the real-time air density to obtain the standardized power, so as to use the accurately calculated real-time air density to correct the actually measured power value to a unified and standardized reference condition through a correction model based on physical principles. In this way, a standardized power index that has been stripped of the influence of environmental density and only reflects the aerodynamic and mechanical and electrical conversion efficiency of the wind turbine itself can be generated, which provides a stable benchmark for subsequent construction of a high-precision power prediction model and sensitive performance degradation evaluation, thereby improving the reliability of early fault warning.

[0035] Specifically, in step S300, power deviation data is calculated based on the rich feature data and the trained power prediction model. It should be understood that since the health status of a wind turbine ultimately reflects its power generation capacity under certain working conditions, merely observing the absolute value of the normalized power cannot determine whether it is in an optimal state, because the theoretical optimal power of a healthy wind turbine is dynamically changing under different combinations of wind speed, pitch angle, etc. Without an accurate theoretical power benchmark, it is impossible to quantify the gap between actual performance and ideal performance, thus it is impossible to effectively identify early and minor performance degradation. Therefore, in the technical solution of the present application, power deviation data is further calculated based on the rich feature data and the trained power prediction model, so as to use the power prediction model that has learned the operation rules of a healthy wind turbine to generate a high-fidelity theoretical optimal power prediction value for the current real-time working condition, and compare the theoretical optimal power prediction value with the actual normalized power value, thereby converting the performance degradation problem into a quantifiable power deviation data. In this way, the complex performance evaluation problem can be simplified into monitoring of a single deviation index, and the power deviation data not only directly quantifies the loss of power generation capacity, but also provides a trigger basis and quantitative input for subsequent fault attribution diagnosis due to its high signal-to-noise ratio, thereby improving the sensitivity and accuracy of the entire evaluation method.

[0036] Figure 3 A flowchart of the step of calculating power deviation data based on the rich feature data and the trained power prediction model in the wind farm environment adaptive power generation capacity dynamic evaluation method according to the embodiment of the present application is shown in FIG. 3. As shown in FIG. 3, step S300 includes: S310, passing the rich feature data through the trained power prediction model to obtain the predicted power of the wind turbine under the current working condition; and S320, calculating the difference between the predicted power and the normalized power in the rich feature data as the power deviation data. Figure 3

[0037] ​Specifically, in step S310, the rich feature data is passed through the trained power prediction model to obtain the predicted power of the wind turbine under the current working condition. It should be understood that, since the operation of the wind turbine is a complex nonlinear process affected by multiple variables, the theoretical optimal power at any moment cannot be obtained by simple lookup table or static formula, and must rely on a dynamic model that can accurately depict its healthy operation mode. If such a high-fidelity model is lacking to generate a theoretical power benchmark, it is impossible to determine whether the current actual power output is reasonable, and thus effective performance evaluation cannot be carried out. Therefore, in the technical solution of the present application, the rich feature data is further passed through the trained power prediction model to obtain the predicted power of the wind turbine under the current working condition, so as to utilize the complex mapping relationship learned from the massive historical health data by the power prediction model to reason the current real-time multi-dimensional working condition, thereby generating an optimal power value that the wind turbine should theoretically be able to achieve at this moment. In this way, an accurate and dynamic benchmark reference can be provided for subsequent calculation of power deviation, and a complex, multi-variable performance evaluation problem is converted into direct quantification of the difference between the actual power and the theoretically predicted power, thereby accurately identifying performance degradation and starting the subsequent fault diagnosis process.

[0038] Figure 4 The flowchart of passing the rich feature data through the trained power prediction model to obtain the predicted power of the wind turbine under the current working condition according to the wind farm environment adaptive power generation capacity dynamic evaluation method of the embodiment of the present application. As shown in Figure 4 S310, step S310 includes: S311, extracting the time sequence of active power from the rich feature data; S312, constructing other data in the rich feature data except the time sequence of active power into a rich feature data matrix according to the time dimension and the data sample dimension; S313, passing the time sequence of active power and the rich feature data matrix through the trained power prediction model to obtain a power time sequence representation vector under the guidance of rich features; S314, decoding the power time sequence representation vector under the guidance of rich features to obtain the predicted power.

[0039] Specifically, in step S311, a time series of active power is extracted from the rich feature data. It should be understood that, since the rich feature data is a structured data set containing multiple dimensional variables, and the power prediction model proposed in the present application adopts a multi-modal guiding architecture which requires the core indicators directly reflecting the performance output of the wind turbine and the auxiliary features describing the external working conditions as inputs of two different modalities to be processed separately, if all features are treated equally, the deep fusion mechanism of guiding and analyzing other features with the power time series mode as the core cannot be realized. Therefore, in the technical solution of the present application, the time series of active power is further extracted from the rich feature data, so as to separate the most critical performance characterization signal from the multi-dimensional feature matrix and make it an independent one-dimensional input stream with clear time series dynamic information. In this way, data meeting the structural requirements of the specific input port of the power prediction model can be provided, which is the premise and basis for realizing subsequent multi-modal guiding time series correlation analysis and generating a deep coupling characterization vector.

[0040] More specifically, in one specific example of the present application, the extraction process of the time series of active power can be automatically completed by a program on a data server. First, the generated rich feature data table is located, which is indexed by time stamp, and each column corresponds to different features, including wind speed, rotational speed, and standardized and corrected active power, etc. Then, all data values of a specific column labeled as standardized power are selected and read. Since the rich feature data table itself is sorted by time, this reading operation naturally preserves the inherent time sequence between power data points. Subsequently, in order to adapt to the input requirements of the power prediction model, the sliding window technique is used to slice the time series of active power. For example, a time step of 144 length (corresponding to 24 hours of 10 minute data) is set, a window starts from the starting point of the time series of active power, and the first 144 power data points are intercepted to form the first time series sample; then, the window slides one time step backward, and the second to 145th data points are intercepted to form the second sample, and so on, until the entire power data long sequence, i.e. the time series of active power, is traversed. Finally, a large number of fixed-length, overlapping active power time series samples are generated, which constitute the core modal input required by the power prediction model.

[0041] Specifically, in step S312, the data other than the time series of the active power in the rich feature data is constructed into a rich feature data matrix according to the time dimension and the data sample dimension. It should be understood that, since the multi-modal guided power prediction model adopted in the present application has a specific double-path input structure, in which one path is specially used to receive the time series of the active power as the core guiding signal, and the other path needs to receive multi-dimensional feature information capable of comprehensively describing the operating context of the wind turbine, if all the features are mixed together or input in an unstructured manner, the power prediction model will not be able to distinguish between the guiding modal and the context modal, and the core guiding and correlation analysis mechanism will not be able to start. Therefore, in the technical solution of the present application, the data other than the time series of the active power in the rich feature data is further constructed into a rich feature data matrix according to the time dimension and the data sample dimension, so as to create a structured and high-dimensional information carrier for the context input end of the power prediction model. In this way, it can be ensured that the complete information containing all other operating variable conditions, which is strictly aligned in time with the time series of the active power, is input into the power prediction model as a whole, and a data organization form is provided for the power prediction model to perform in-depth analysis of the complex dynamic relationship between multiple variables guided by the power time sequence.

[0042] More specifically, in one specific example of the present application, the construction process of the rich feature data matrix is synchronized with the extraction process of the aforementioned active power time series. First, the generated rich feature data table is locked on the data server, and all other feature columns except the standardized power are identified, such as wind speed, pitch angle, rotor speed, cabin temperature, ambient temperature, air pressure, humidity, and calculated real-time air density. Then, the sliding window mechanism defined in the aforementioned step is reused, for example, a window with a length of 144 time steps. When the sliding window intercepts a segment of the active power time series sample, corresponding data segments are simultaneously extracted from all other feature columns at the same time start and end points. Then, these extracted data segments from different feature columns are spliced in the feature dimension to form a two-dimensional array, i.e. a rich feature data matrix. The number of rows of the rich feature data matrix is equal to the time step length of the sliding window (for example, 144), and the number of columns of the rich feature data matrix is equal to the total number of features except the active power. This process is repeated with each movement of the sliding window, and finally a series of rich feature data matrices are generated, each of which has a one-to-one correspondence with an active power time series sample, thereby providing a perfectly aligned data pair for the double-path input of the power prediction model.

[0043] Specifically, in step S313, the time sequence of the active power and the rich feature data matrix are passed through the trained power prediction model to obtain a rich feature guided power time sequence representation vector. It should be understood that since the power output dynamics of a wind turbine not only depend on its own time evolution law, but also are complexly and nonlinearly modulated by many external working condition variables such as wind speed and pitch angle, if the power time sequence and the rich feature data matrix describing the working conditions are processed as two independent or simply spliced information sources, the power prediction model will not be able to capture the deep coupling relationship between the specific time sequence pattern of the power and the multi-dimensional working condition features, thereby leading to a one-sided understanding of the running state of the wind turbine and failing to generate an accurate state representation that can comprehensively reflect the internal running mechanism. Therefore, in the technical solution of the present application, the time sequence of the active power and the rich feature data matrix are further passed through the trained power prediction model to obtain a rich feature guided power time sequence representation vector, so as to perform a multi-modal guided deep fusion process, that is, using the comprehensive working condition information contained in the rich feature data matrix to dynamically and selectively focus on and analyze the key patterns and dynamic changes in the active power time sequence, thereby deeply fusing the two heterogeneous data sources and compressing them into a low-dimensional but highly information-dense state vector. In this way, a rich feature guided power time sequence representation vector can be generated, which not only contains the time dynamics of the power itself, but also embeds the modulating effect of multi-dimensional working conditions, thereby achieving comprehensive and refined analysis of the current health state of the wind turbine. As a direct input for subsequent power prediction and fault attribution, the rich feature guided power time sequence representation vector improves the sensitivity of the power prediction model to minor performance deviations and the prediction accuracy of the power behavior under complex working conditions.

[0044] More specifically, in the embodiment of the present application, passing the time sequence of the active power and the rich feature data matrix through the trained power prediction model to obtain a rich feature guided power time sequence representation vector comprises: vectorizing and encoding the time sequence of the active power to extract active power time sequence features to obtain an active power time sequence pattern encoding vector; performing implicit time sequence correlation coding on the rich feature data matrix to obtain a rich feature implicit correlation feature map; and performing multi-modal guided time sequence correlation analysis on the active power time sequence pattern encoding vector and the rich feature implicit correlation feature map to obtain the rich feature guided power time sequence representation vector.

[0045] Specifically, the time sequence of the active power is vectorized and encoded to extract the active power time sequence feature to obtain an active power time sequence mode encoding vector. It should be understood that since the original active power time sequence is a high-dimensional, time-ordered original data stream, the dynamic patterns such as trends, periodicity and mutations contained therein are implicit and cannot be directly used as structured features to participate in subsequent guidance and fusion calculations. If not refined by encoding, the power prediction model will be difficult to capture and quantify the evolution law of the power sequence itself, and thus an effective guidance signal cannot be formed. Therefore, in the technical solution of the present application, the time sequence of the active power is further vectorized and encoded to extract the active power time sequence feature to obtain an active power time sequence mode encoding vector. In this way, the neural network structure specially processing sequence data is used to compress and map the input long sequence data into a low-dimensional, fixed vector space, so as to convert the time sequence dynamic information into a static, highly condensed feature vector. In this way, an active power time sequence mode encoding vector can be generated, which represents the dynamic characteristics of the power as a power behavior pattern in the time period, and provides a structured guidance source for the subsequent multi-modal fusion process of guiding and analyzing other multi-dimensional working condition features based on the mode.

[0046] More specifically, in one specific example of the present application, the vectorization and encoding process is a core sub-module inside the power prediction model, for example, a Long Short-Term Memory (LSTM) encoder. First, an active power time sequence sample, for example, a sequence containing 144 time steps, is input to the input layer of the LSTM encoder. The LSTM encoder processes each power value in the sequence one time step at a time. At each time step, the LSTM encoder combines the current power input value with the hidden state and cell state passed from the previous time step, and updates its internal memory through its internal forget gate, input gate and output gate structure to determine which historical information to retain and which new information to store. This process propagates forward along the time axis until the last data point in the active power time sequence sample is processed. After processing the entire sequence, the hidden state vector output by the LSTM encoder at the last time step is defined as the active power time sequence mode encoding vector. The dimension of the active power time sequence mode encoding vector is predetermined, and it has compressed and encoded the power dynamic evolution information of the entire 144 time steps therein.

[0047] Specifically, the rich feature data matrix is implicitly time-correlation coded to obtain a rich feature implicit correlation feature map. It should be understood that, since the rich feature data matrix contains comprehensive information describing the operating conditions of the wind turbine, but the information is in the form of multiple parallel time series, the mutual influence and lag correlation between variables such as wind speed, pitch angle, rotor speed, etc. in the time dimension is implicit and not explicitly expressed. If the rich feature data matrix is directly used for subsequent guided fusion, the power prediction model will be difficult to effectively capture the complex dynamic cooperation between variables, thereby limiting its depth understanding of the operating condition context. Therefore, in the technical solution of the present application, the rich feature data matrix is further implicitly time-correlation coded to obtain a rich feature implicit correlation feature map, so as to utilize the deep learning structure capable of automatically extracting local and global correlations to autonomously learn and code the mutual dependence of different operating condition variables in the time evolution process without pre-defining physical rules. In this way, a rich feature implicit correlation feature map can be generated, which is no longer a simple stacking of original data, but a high-order information representation that explicitly expresses the time correlation pattern between multiple variables, providing an information-rich and structurally optimized context information carrier for the subsequent guided fusion step.

[0048] More specifically, in one specific example of the present application, the process of implicitly time-correlation coding of the rich feature data matrix is realized by a Convolutional Neural Network (CNN) module inside the power prediction model. First, a rich feature data matrix, such as a matrix with dimensions 144x8 (144 time steps, 8 features), is taken as the input of the CNN module. The CNN module contains multiple one-dimensional convolution layers. The first convolution layer uses a set of small-sized convolution kernels, such as a 3x8 convolution kernel, to perform sliding convolution operations along the time dimension (i.e. the direction of 144 time steps) of the input matrix. At each sliding position, the convolution kernel performs weighted summation on all 8 feature data within a small window of 3 time steps in the matrix, thereby capturing the local correlation pattern of multiple variables in a very short time. After all the convolution kernels complete a complete sliding, the first layer feature map is generated. Then, the feature map is processed by a nonlinear activation function (such as ReLU) and sent to the next convolution layer. The subsequent convolution layers operate on the feature map output by the previous convolution layer. Since their receptive field is larger, they can capture longer time-span and more abstract variable correlation patterns. After multiple convolution stacking, the two-dimensional tensor output by the last convolution layer is the rich feature implicit correlation feature map. Each element of the rich feature implicit correlation feature map encodes the complex time correlation information of the original multi-variable data at different time scales.

[0049] Specifically, the active power time sequence mode coding vector and the rich feature implicit association feature map are subjected to multi-modal guided time sequence association analysis to obtain the power time sequence representation vector guided by the rich feature. It should be understood that, since the active power time sequence mode coding vector and the rich feature implicit association feature map are two heterogeneous information representations, the former condenses the dynamic mode of the wind turbine performance output, and the latter depicts the internal association of the multi-dimensional operating condition. If only simple feature-level splicing or parallel processing is performed on the two, it is impossible to realize which association in the complex condition feature map is the key to causing the specific dynamic mode of the power, thereby losing the deep causal and guided relationship between the two. Therefore, in the technical solution of the present application, the active power time sequence mode coding vector and the rich feature implicit association feature map are further subjected to cross-modal guided time sequence association analysis, so as to perform a paradigm transition from data interaction to structure interaction, that is, the active power time sequence mode coding vector is used as a high-level semantic prior to dynamically modulate and reconstruct the topological association inside the rich feature implicit association feature map, generate a context relationship graph highly relevant to the current power mode, and perform deep information aggregation on the dynamic graph. In this way, a power time sequence representation vector guided by the rich feature can be generated, which is not a simple linear superposition of the two modal information, but a highly condensed fusion representation obtained after deep context reasoning, which has encoded the multivariate time sequence association structure most relevant to the current unit state guided by the power dynamic, thereby providing the context adaptability and reasoning depth input for subsequent power prediction.

[0050] More specifically, in the embodiment of the present application, the multi-modal guided time sequence association analysis of the active power time sequence mode coding vector and the rich feature implicit association feature map to obtain the power time sequence representation vector guided by the rich feature includes: nodeizing the rich feature implicit association feature map along the channel dimension to obtain a set of rich feature implicit association coding vectors; taking the active power time sequence mode coding vector as a modulator, calculating a time-varying context association matrix of the set of rich feature implicit association coding vectors to obtain a rich feature implicit association time-varying context association matrix; inputting the set of rich feature implicit association coding vectors and the rich feature implicit association time-varying context association matrix into a graph convolutional neural network model to obtain a rich feature implicit association context enhanced representation matrix; and performing multi-modal dynamic interaction on the rich feature implicit association context enhanced representation matrix and the active power time sequence mode coding vector to obtain the power time sequence representation vector guided by the rich feature.

[0051] Specifically, the feature-rich latent association feature map is node-based along the channel dimension to obtain a set of feature-rich latent association encoding vectors. This is expressed by the following formula: .

[0052] in, For feature-rich latent association feature maps, For feature reshaping, It is a set of feature-rich latent association encoding vectors. These are the 1st, 2nd, and 3rd vectors in the set of feature-rich latent association encoding vectors. A feature-rich latent association encoding vector.

[0053] It should be understood that, since the rich latent association feature map is essentially a continuous, image-like grid structure data, its inherent spatial adjacency relationships cannot directly support subsequent more flexible and higher-order graph structure relationship reasoning. Without a paradigm shift, it is impossible to introduce models such as graph neural networks to transcend local temporal adjacency and capture and construct non-local, semantic-level operational feature associations guided by power modes. Therefore, in the technical solution of this invention, the rich latent association feature map is further vectorized and reconstructed along the channel dimension to obtain a set of rich latent association encoding vectors. This performs the key step of deconstructing and nodeifying the continuous feature map, abstracting the feature information representing different temporal positions or local regions in the rich latent association feature map into a series of discrete, independent encoding vectors, thereby completing the key transformation from the grid paradigm to the graph structure paradigm. This lays a structural foundation for using the active power time-series pattern encoding vector as a semantic prior and dynamically constructing the contextual adjacency relationships between these encoding vectors (i.e., graph nodes), enabling the power prediction model to reason about the correlation between operating condition features from a higher dimension, rather than being limited to their fixed adjacency relationships in the original time series.

[0054] Specifically, using the active power time-series pattern coding vector as a modulator, the time-varying context correlation matrix of the set of feature-rich latent correlation coding vectors is calculated to obtain the feature-rich latent correlation time-varying context correlation matrix. This is expressed by the following formula: ; .

[0055] in, for The transpose of , and These are the first in the set of feature-rich latent association encoding vectors. The and the first A feature-rich latent association encoding vector, It is the active power timing pattern encoding vector. For matrix multiplication, For learnable projection matrices, To balance the hyperparameters of content relevance and structural similarity, For Gaussian similarity kernel, for function, For feature-rich latent associations in time-varying contextual association matrices Location feature value For feature-rich, latent association time-varying contextual association matrices, Let the first norm of the vector be 1. Let e ​​be the value of the logarithmic function with the natural constant e as the base. These are learnable weights.

[0056] It should be understood that although the set of feature-rich latent correlation encoding vectors contains operating condition feature information at various times, the correlation strength and pattern between them are not static, but dynamically change with the actual power output state of the wind turbine. If a fixed, predefined correlation structure is used, it cannot reflect the fact that the interaction between certain operating condition features becomes particularly critical when power exhibits a specific pattern (such as a sudden drop or oscillation), thus causing the power prediction model to have a static and sluggish understanding of the context. Therefore, in the technical solution of this invention, the active power time-series pattern encoding vector is further used as the context modulation signal to construct a time-varying context correlation matrix of the set of feature-rich latent correlation encoding vectors to obtain a feature-rich latent correlation time-varying context correlation matrix. This completely overturns the idea of ​​constructing a static graph, and uses the encoding vector representing power dynamics as a high-level semantic prior to dynamically generate an adjacency matrix describing the correlation strength between each operating condition feature node for each specific power pattern instance. In this way, a feature-rich, latently correlated, time-varying contextual correlation matrix can be generated. The topological structure of this matrix itself becomes part of the contextual information. It directly injects the global semantic information of the power mode into the structural definition of the operating condition features. For example, when the power mode is cut-out, the feature-rich, latently correlated, time-varying contextual correlation matrix will adaptively enhance the connection weight between the node representing ultra-high wind speed and the node representing the extreme pitch angle, thereby providing a highly task-related, non-local, power mode-prior-calibrated inference path for subsequent graph convolution operations.

[0057] Specifically, the set of rich-feature latent association encoding vectors and the rich-feature latent association time-varying context association matrix are input into a graph convolutional neural network model to obtain a rich-feature latent association context enhancement representation matrix. This is expressed by the following formula: .

[0058] wherein, is a graph convolutional neural network model, is a rich feature implicit correlation context enhanced representation matrix.

[0059] It can be understood that, since the rich feature implicit correlation time-varying context correlation matrix only defines a blueprint or path of how each operating condition feature node should be associated in a specific power mode, the information of each operating condition feature node (i.e. the rich feature implicit correlation coding vector) itself has not yet been propagated and fused along these newly constructed paths. Without performing information aggregation, this structured advantage guided by the power mode cannot be translated into the enhancement of the feature representation itself, and the context information is still fragmented. Therefore, in the technical solution of the present application, the set of rich feature implicit correlation coding vectors and the rich feature implicit correlation time-varying context correlation matrix are further input into a graph convolutional neural network model to obtain a rich feature implicit correlation context enhanced representation matrix, so as to utilize the powerful information propagation and node feature updating capability of the graph convolutional neural network to perform iterative message passing on the graph topology just calibrated by the power time sequence mode coding vector. In this way, information can be integrated on an optimal path guided by semantics, so that the representation of each operating condition feature node can absorb the information of the most important associated node in the current power mode, thereby generating a rich feature implicit correlation context enhanced representation matrix, which is a structured and overall operating condition representation obtained after deep context reasoning, and its internal has encoded the most relevant topological relationship with the current power mode.

[0060] Specifically, the rich feature implicit correlation context enhanced representation matrix and the active power time sequence mode coding vector are subjected to multi-modal dynamic interaction to obtain the rich feature guided power time sequence feature vector. This can be expressed by the following formula: .

[0061] wherein, is a rich feature guided power time sequence feature vector.

[0062] It should be understood that, although the rich-feature implicit correlation context-enhanced representation matrix is a structured overall working condition characterization after power mode guidance, it is still a distributed and high-dimensional information set, and has not been finally aligned and fused with the power mode itself as the guidance source. If the last step of interaction and contraction is not performed, the power prediction model will lack a single and highly information-concentrated final characterization to drive subsequent prediction or classification tasks. Therefore, in the technical solution of the present application, the rich-feature implicit correlation context-enhanced representation matrix and the active power time series mode encoding vector are further coupled in a multi-modal time series to perform a secondary focusing and final alignment operation, i.e., the active power time series mode encoding vector is again used as the final query vector to perform a precise and global weighted reading of the working condition characterization matrix that has been sufficiently contextually inferred. In this way, the most critical information can be selectively extracted and fused from the structured reorganized working condition information according to the final pointing of the power mode, and finally contracted into a single and highly information-concentrated rich-feature guided power time series characterization vector, which is the inevitable result of the deep and structured interaction of the two modalities, ensuring that the final output features not only fuse the information of the two modalities, but also provide the most refined and relevant decision basis for subsequent tasks.

[0063] Specifically, in step S314, the rich-feature guided power time series characterization vector is decoded to obtain power deviation data. It should be understood that, since the rich-feature guided power time series characterization vector is a highly abstract and condensed expression of the wind turbine operating state in a low-dimensional latent space, it is not a prediction value that can be directly compared with an actual physical quantity. If it is not mapped back to the physical space, the theoretical optimal performance of the wind turbine under the current complex working condition cannot be quantified, and a precise performance evaluation benchmark cannot be established. Therefore, in the technical solution of the present application, the rich-feature guided power time series characterization vector is further decoded to translate and reconstruct the deeply fused state vector into a theoretical power prediction value that is most likely to occur in the current working condition context. In this way, a more accurate power prediction result can be obtained than traditional methods, and the power prediction result already contains a deep understanding of the complex coupling relationship of multiple variables. By comparing the high-precision predicted power with the actual standardized power, the power deviation data obtained can effectively eliminate the influence of normal working condition fluctuations on power, thereby becoming a high signal-to-noise ratio and directly pointing to the quantitative indicator of potential performance degradation or failure of the wind turbine.

[0064] Specifically, in step S400, based on the power deviation data, a cause classification classifier is started to perform fault diagnosis to obtain a diagnosis result. It should be understood that, since the power deviation data is a sensitive index capable of accurately quantifying the performance deviation of the wind turbine from the theoretical optimal state, but it only reflects the existence and degree of the abnormality, and does not reveal the specific physical root cause or fault type leading to the deviation, if only the deviation value is provided without attribution, the operation and maintenance personnel still need to perform complex troubleshooting to locate the problem, and the final goal of intelligent diagnosis cannot be achieved. Therefore, in the technical solution of the present application, based on the power deviation data, a cause classification classifier is started to perform fault diagnosis to obtain a diagnosis result, so as to improve the problem from abnormality detection to fault attribution, that is, using a specially trained classification model as a cause classification classifier, the current observed high signal-to-noise ratio power deviation signal and its mode are matched and identified with the characteristics corresponding to each type of known fault (such as pitch system abnormality, inaccurate yaw-to-wind, sensor failure, etc.) learned in advance. In this way, a clear diagnosis conclusion with physical meaning can be directly obtained, for example, pitch angle control system response delay or generator cooling system efficiency decrease, thereby replacing the traditional time-consuming and laborious fault troubleshooting process relying on manual experience, achieving automatic and accurate positioning of early and small faults of the wind turbine, and improving the efficiency and accuracy of operation and maintenance decisions.

[0065] More specifically, in the embodiments of the present application, based on the power deviation data, the attribution classifier is started to perform fault diagnosis to obtain a diagnosis result, including: when the absolute value of the power deviation data is greater than a preset abnormality judgment threshold, performing attribution diagnosis to obtain a diagnosis result. Wherein, the attribution diagnosis includes: adding the power deviation data to the tail end of the rich feature guided power time sequence representation vector to obtain an attribution feature vector; passing the attribution feature vector through the attribution classifier to obtain the diagnosis result, the diagnosis result including an attribution diagnosis type label and its confidence. It can be understood that since the power deviation data is only a single scalar representing the degree of abnormality, and the rich feature guided power time sequence representation vector contains complex working condition context information leading to the deviation, if the two are processed separately, the attribution classifier will not be able to establish a mapping relationship between a specific working condition and a specific deviation, making it difficult to distinguish the subtle differences between different faults under similar deviations, resulting in ambiguity and uncertainty in diagnosis. Therefore, in the technical solution of the present application, the power deviation data is further added to the tail end of the rich feature guided power time sequence representation vector to obtain an attribution feature vector, and the attribution feature vector is passed through the attribution classifier to obtain the diagnosis result, so as to construct an information complete attribution feature vector, which not only explicitly contains the quantitative result of performance abnormality (power deviation), but also retains the full-dimensional deep working condition snapshot (representation vector) when the corresponding abnormality occurs, and takes it as the only input of the attribution classifier. In this way, the attribution classifier can learn and reason in the feature space, thereby improving the accuracy and robustness of fault recognition, and finally obtaining a clear diagnosis result containing the specific fault type and its occurrence probability, realizing a decisive leap from detecting abnormality to diagnosing cause.

[0066] Specifically, in step S500, a quantitative assessment based on the loss of power generation capacity is performed based on the diagnostic result and the power deviation data to generate an alarm information and an assessment report. It should be understood that, since the diagnostic result only indicates the type of fault, and the power deviation data only quantifies the instantaneous performance loss, neither of them directly translates into a business indicator that is meaningful for operational decision-making. Without a link between technical diagnosis and economic impact, the operational manager will have difficulty in determining the urgency and priority of the problem from numerous alarms, and cannot compare the economic losses caused by different faults horizontally, thereby affecting the optimal allocation of maintenance resources. Therefore, in the technical solution of the present application, a quantitative assessment based on the loss of power generation capacity is further performed based on the diagnostic result and the power deviation data to generate an alarm information and an assessment report, so as to convert the abstract fault label and the instantaneous power deviation value into an intuitive and quantifiable power generation loss and economic loss indicator through accumulation in the time dimension and combination with economic parameters (such as electricity price). In this way, a clear and value-based decision-making basis can be provided for the operation team, so that the alarm information is no longer a mere technical notification, but an action guide containing explicit business impact, thereby realizing a complete closed loop from fault diagnosis to asset management efficiency assessment, and ensuring that the operation activities can prioritize the problems that have the greatest impact on power generation efficiency.

[0067] More specifically, in one specific example of the present application, the quantitative assessment and report generation process is activated after the attribution classifier outputs a diagnostic result with a confidence exceeding a preset threshold (e.g. 90%). First, the power deviation data stream output by the power prediction model is continuously recorded, and the on-grid electricity price information applicable to the wind farm is obtained from the configuration database. Then, the power deviation data within a preset assessment period (e.g. every hour or every day since the fault was diagnosed) is subjected to a time integration operation. Specifically, the power deviation value (in kilowatts) of each data sampling point is multiplied by the time interval of the data sampling (e.g. 10 minutes, i.e. 1 / 6 hours), and then the calculation results of all sampling points within the assessment period are accumulated to obtain the total power generation loss (in kilowatt-hours) within the assessment period. Subsequently, the calculated total power generation loss is multiplied by the preset on-grid electricity price to obtain the direct economic loss amount caused by the fault within the assessment period. Finally, these quantitative results are structured and integrated with the diagnostic result to generate an assessment report. The assessment report includes wind turbine number, fault diagnosis type, fault confidence, assessment time period, power deviation time series graph, cumulative power generation loss, and cumulative economic loss, etc. key fields. At the same time, if the instantaneous power deviation or the cumulative economic loss exceeds the preset alarm threshold, a high-priority alarm information will be automatically generated, including the fault type and the current hour loss amount, etc. core summary, and pushed to the central monitoring system and the mobile terminal of the operation personnel.

[0068] In summary, the wind farm environment-adaptive dynamic power generation capacity assessment method according to embodiments of the present invention is elucidated. It breaks away from the limitation of traditional methods that treat all monitoring data as equal inputs, abandoning simple feature concatenation and instead using active power time-series data as the core mode. By constructing a new power prediction model, it utilizes the inherent patterns of the power sequence as contextual prior information to dynamically guide and analyze the hidden temporal correlations in other feature-rich data (such as wind speed, pitch angle, meteorological parameters, etc.). This multimodal guidance and coupling mechanism generates a unified state representation vector that deeply integrates the operating context. This vector not only improves the accuracy of power prediction and makes anomaly detection based on power deviation more reliable, but it also becomes a highly condensed data package, providing high-quality input for subsequent attribution classifiers, thereby achieving closed-loop monitoring from accurate state assessment to precise fault tracing.

[0069] This invention also provides a dynamic evaluation system for wind farm power generation capacity that is environmentally adaptive.

[0070] Figure 5 This is a block diagram of a wind farm environment-adaptive dynamic power generation capacity assessment system according to an embodiment of the present invention. Figure 5 As shown, the wind farm environment-adaptive dynamic power generation capacity assessment system 500 according to an embodiment of the present invention includes: a real-time data acquisition module 510, used to acquire real-time SCADA data streams of wind turbine generators and real-time meteorological data streams of the wind farm environment, wherein the real-time SCADA data streams include wind speed, active power, pitch angle, rotor speed, and nacelle temperature, and the real-time meteorological data streams include ambient temperature, air pressure, and humidity; a feature-rich data generation module 520, used to perform data preprocessing and feature enrichment on the real-time SCADA data streams and the real-time meteorological data streams to obtain feature-rich data; a power deviation calculation module 530, used to calculate power deviation data based on the feature-rich data and a trained power prediction model; a fault diagnosis module 540, used to activate an attribution classifier to perform fault diagnosis based on the power deviation data to obtain a diagnosis result; and a quantitative assessment module 550, used to perform a quantitative assessment based on power generation capacity loss based on the diagnosis result and the power deviation data to generate alarm information and an assessment report.

[0071] The specific implementation method of the wind farm environment-adaptive dynamic power generation capacity assessment system provided in this embodiment of the invention can be found in the description of the wind farm environment-adaptive dynamic power generation capacity assessment method provided in this embodiment of the invention, and will not be repeated here.

[0072] The wind farm environment adaptive power generation capacity dynamic evaluation system 500 according to the embodiments of the present application can be implemented in various wireless terminals, such as a server with a wind farm environment adaptive power generation capacity dynamic evaluation control algorithm, etc. In a possible implementation, the wind farm environment adaptive power generation capacity dynamic evaluation system 500 according to the embodiments of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the wind farm environment adaptive power generation capacity dynamic evaluation system 500 can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the wind farm environment adaptive power generation capacity dynamic evaluation system 500 can also be one of the many hardware modules of the wireless terminal.

[0073] Embodiments of the present disclosure have been described above, the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles, practical applications, or improvements to the technology in the market of the embodiments, or to enable other ordinary skilled in the art to understand the embodiments disclosed herein.

Claims

1. A wind farm environment adaptive power generation capacity dynamic evaluation method, characterized in that, The method comprises: obtaining a real-time SCADA data stream of a wind turbine and a real-time meteorological data stream of a wind farm environment, the real-time SCADA data stream comprising wind speed, active power, pitch angle, rotor speed and nacelle temperature, and the real-time meteorological data stream comprising ambient temperature, air pressure and humidity; data preprocessing and feature enrichment are performed on the real-time SCADA data stream and the real-time meteorological data stream to obtain enriched feature data; based on the enriched feature data and a trained power prediction model, power deviation data is calculated; based on the power deviation data, a fault diagnosis is performed by starting an attribution classifier to obtain a diagnosis result; based on the diagnosis result and the power deviation data, a quantitative assessment based on power generation capacity loss is performed to generate alarm information and an assessment report.

2. The wind farm environment adaptive power generation capacity dynamic assessment method according to claim 1, characterized in that, The data preprocessing and feature enrichment performed on the real-time SCADA data stream and the real-time meteorological data stream to obtain enriched feature data comprise: time alignment, cleaning and invalid data filtering are performed on the real-time SCADA data stream and the real-time meteorological data stream to obtain cleaned joint data; based on the cleaned joint data, real-time air density and normalized power are calculated and integrated into the cleaned joint data to obtain the enriched feature data.

3. The wind farm environment adaptive power production capability dynamic assessment method of claim 2, wherein, The calculation of real-time air density and normalized power based on the cleaned joint data comprises: based on the cleaned ambient temperature, the cleaned air pressure and the cleaned humidity, the real-time air density is calculated according to the following air density formula: ； wherein, is the real-time air density, is the air pressure after cleaning, is the ambient temperature after cleaning, is the water vapor partial pressure and , is the humidity after cleaning, is the saturated water vapor pressure, is the specific gas constant of dry air, is the specific gas constant of water vapor.

4. The wind farm environment adaptive power production capability dynamic assessment method of claim 3, wherein, The calculation of real-time air density and normalized power based on the cleaned joint data further comprises: based on the real-time air density, the active power in the cleaned joint data is normalized and corrected according to the following correction formula to obtain the normalized power: ； wherein, is the standardised power, is the active power in the cleaned combined data, is the standard air density, is the real-time air density.

5. The wind farm environment adaptive power production capability dynamic assessment method of claim 4, wherein, based on the enriched feature data and the trained power prediction model, power deviation data is calculated, which comprises: the enriched feature data is input into the trained power prediction model to obtain the predicted power of the wind turbine under the current working condition; the difference between the predicted power and the normalized power in the enriched feature data is calculated as the power deviation data.

6. The wind farm environment adaptive power generation capability dynamic assessment method according to claim 5, characterized in that, The input of the enriched feature data into the trained power prediction model to obtain the predicted power of the wind turbine under the current working condition comprises: a time sequence of active power is extracted from the enriched feature data; other data in the enriched feature data except the time sequence of active power are constructed into an enriched feature data matrix according to the time dimension and the data sample dimension; the time sequence of active power and the enriched feature data matrix are input into the trained power prediction model to obtain a power time sequence representation vector under the guidance of enriched features; the power time sequence representation vector under the guidance of enriched features is decoded to obtain the predicted power.

7. The wind farm environment adaptive power production capability dynamic assessment method of claim 6, wherein, The input of the time sequence of active power and the enriched feature data matrix into the trained power prediction model to obtain a power time sequence representation vector under the guidance of enriched features comprises: the time sequence of active power is vectorized and encoded to extract active power time sequence features to obtain an active power time sequence pattern encoding vector; performing implicit temporal correlation coding on the feature-rich data matrix to obtain a feature-rich implicit correlation feature map; performing multi-modal guided temporal correlation analysis on the active power temporal pattern coding vector and the feature-rich implicit correlation feature map to obtain the feature-rich guided power temporal representation vector.

8. The wind farm environment adaptive power generation capacity dynamic assessment method according to claim 7, characterized in that, performing multi-modal guided temporal correlation analysis on the active power temporal pattern coding vector and the feature-rich implicit correlation feature map to obtain the feature-rich guided power temporal representation vector, comprising: vectorizing the feature-rich implicit correlation feature map along the channel dimension to obtain a set of feature-rich implicit correlation coding vectors; constructing a time-varying context correlation matrix of the set of feature-rich implicit correlation coding vectors to obtain a feature-rich implicit correlation time-varying context correlation matrix, with the active power temporal pattern coding vector as a context modulation signal; inputting the set of feature-rich implicit correlation coding vectors and the feature-rich implicit correlation time-varying context correlation matrix into a graph convolutional neural network model to obtain a feature-rich implicit correlation context-enhanced representation matrix; performing multi-modal temporal coupling on the feature-rich implicit correlation context-enhanced representation matrix and the active power temporal pattern coding vector to obtain the feature-rich guided power temporal representation vector.

9. The wind farm environment adaptive power generation capability dynamic assessment method according to claim 7, characterized in that, based on the power deviation data, starting an attribution classifier to perform fault diagnosis to obtain a diagnosis result, comprising: performing attribution diagnosis when the absolute value of the power deviation data is greater than a preset abnormality judgment threshold to obtain a diagnosis result; wherein the attribution diagnosis comprises: adding the power deviation data to the tail end of the feature-rich guided power temporal representation vector to obtain an attribution feature vector; inputting the attribution feature vector through an attribution classifier to obtain the diagnosis result, the diagnosis result comprising an attribution diagnosis type label and a confidence thereof.

10. A wind farm environment adaptive power generation capacity dynamic assessment system, characterized in that, comprising: a real-time data acquisition module configured to acquire a real-time SCADA data stream of a wind turbine and a real-time meteorological data stream of a wind farm environment, the real-time SCADA data stream comprising wind speed, active power, pitch angle, rotor speed and nacelle temperature, and the real-time meteorological data stream comprising ambient temperature, air pressure and humidity; a feature-rich data generation module configured to perform data preprocessing and feature enrichment on the real-time SCADA data stream and the real-time meteorological data stream to obtain feature-rich data; a power deviation calculation module configured to calculate power deviation data based on the feature-rich data and a trained power prediction model; a fault diagnosis module configured to start an attribution classifier based on the power deviation data to perform fault diagnosis and obtain a diagnosis result; a quantitative evaluation module configured to perform quantitative evaluation based on power generation capacity loss based on the diagnosis result and the power deviation data to generate alarm information and an evaluation report.

Images