Multi-source data fusion-based wind turbine generator power generation mode decision-making method and system

Abstract

The embodiment of the invention provides a wind turbine generator power generation mode decision-making method and system based on multi-source data fusion, and the method comprises the steps: carrying out the refined preprocessing and synchronization of multi-source heterogeneous data, such as original radar scanning, SCADA, NWP, CMS vibration, and the like, and extracting the synchronized radar features, high and low frequency SCADA, NWP time sequences, and key fan health indexes. Performing short-term wind regime sudden change prediction by using a deep learning model, and capturing instantaneous wind speed and turbulent flow changes; a GRU model is adopted to carry out medium and long period power generation trend prediction, and confidence evaluation is provided for the two types of prediction. A decision factor fusion evaluation module based on adaptive gating is introduced, the module fuses multivariate information such as a short-term prediction result, medium-and-long-period prediction power, a fan health index, prediction confidence, profit potential, a risk index and power grid constraints, the weight of each decision factor is dynamically adjusted, and optimization decision of a power generation mode and target parameters is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a method and system for making decisions on the power generation mode of wind turbine units based on multi-source data fusion. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, wind power, as an important form of renewable energy, has seen rapid and continuous growth in both installed capacity and electricity generation. However, the inherent intermittency and volatility of wind energy pose significant challenges to the safe and stable operation of wind turbines and grid dispatch. During operation, wind turbines must not only cope with complex and changing wind conditions, such as sudden wind speed changes, strong turbulence, and extreme wind shear, but also consider the health status of their critical components and the grid's capacity to accommodate them. Traditional wind turbine control strategies employ relatively fixed power generation modes, such as maximum power point tracking (MPPT), whose primary objective is to maximize power generation. However, this single-objective control approach often overlooks the mechanical load impacts of extreme wind conditions on the turbine, as well as the risk of failure due to continuous high-load operation when the turbine's own health deteriorates.

[0003] Currently, existing solutions utilize multi-source data for wind farm operation and maintenance and power prediction. For example, some solutions combine data from wind turbine monitoring and data acquisition systems with numerical weather prediction data to predict power generation, or use vibration data from condition monitoring systems to assess the health status of the turbines. However, these existing solutions often have significant drawbacks: First, the application of data sources is relatively isolated, failing to achieve effective integration of multi-dimensional information. For instance, power prediction models rarely consider both ultra-short-term, high-precision wind condition information provided by forward-facing wind radar and the turbine's own health index, making it impossible to accurately predict changes in power generation capacity caused by sudden wind changes or sub-optimal turbine health. Second, the decision-making process is usually step-by-step and static, lacking a unified framework to dynamically weigh the complex relationships between different decision factors (such as power generation revenue, operational risks, and turbine losses). The decision result is often a single predicted value, rather than an optimized power generation mode containing specific target parameters that can directly guide the wind turbine control system. This information fragmentation and lack of a decision-making mechanism prevent wind turbines from proactively and intelligently adjusting their operating strategies according to upcoming specific scenarios, thus missing opportunities to improve overall efficiency.

[0004] Therefore, an optimized decision-making scheme for wind turbine power generation modes is desired. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in the prior art, and provides a method and system for making decisions on the power generation mode of wind turbine units based on multi-source data fusion.

[0006] In a first aspect, embodiments of the present invention provide a wind turbine power generation mode decision-making method based on multi-source data fusion, comprising:

[0007] The original radar scan time series, original SCADA time series, original NWP grid data and original CMS vibration data are preprocessed and synchronized to obtain synchronized radar characteristic time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series and wind turbine health index.

[0008] Short-term wind condition change predictions based on the CNN-LSTM model were performed on the radar characteristic time series and the high-frequency SCADA time series after synchronization to obtain the short-term prediction results and their short-term prediction confidence.

[0009] GRU-based medium- and long-term power generation trend prediction is performed on the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain the medium- and long-term predicted power time series and its medium- and long-term prediction confidence.

[0010] An adaptive gating-based decision factor fusion evaluation was conducted on the short-term forecast results and their short-term forecast confidence, the medium- and long-term forecast power time series and their medium- and long-term forecast confidence, and the wind turbine health index to obtain the recommended power generation mode and target parameters.

[0011] The recommended power generation mode and target parameters are output to the wind turbine control system.

[0012] Secondly, embodiments of the present invention provide a wind turbine power generation mode decision-making system based on multi-source data fusion, comprising:

[0013] The data processing module is used to preprocess and synchronize the original radar scan time series, original SCADA time series, original NWP grid data and original CMS vibration data to obtain synchronized radar characteristic time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series and wind turbine health index.

[0014] The wind condition change prediction module is used to perform short-term wind condition change prediction based on the CNN-LSTM model on the radar characteristic time series and the high-frequency SCADA time series after synchronization to obtain the short-term prediction results and their short-term prediction confidence.

[0015] The power generation trend prediction module is used to perform GRU-based medium- and long-term power generation trend prediction on the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain the medium- and long-term predicted power time series and its medium- and long-term prediction confidence.

[0016] The fusion evaluation module is used to perform adaptive gating-based decision factor fusion evaluation on the short-term forecast results and their short-term forecast confidence, the medium- and long-term forecast power time series and their medium- and long-term forecast confidence, and the wind turbine health index to obtain recommended power generation modes and target parameters.

[0017] The parameter output module is used to output the recommended power generation mode and target parameters to the wind turbine control system.

[0018] Compared with existing technologies, this invention provides a wind turbine power generation mode decision-making method and system based on multi-source data fusion. It refines and synchronizes heterogeneous multi-source data such as raw radar scans, SCADA, NWP, and CMS vibration data, extracting synchronized radar features, high- and low-frequency SCADA and NWP time series, and key wind turbine health indices. Based on this, an advanced deep learning model, namely the CNN-LSTM model, is used to predict short-term wind condition changes, capturing instantaneous wind speed and turbulence variations. Simultaneously, a GRU model is used to predict medium- and long-term power generation trends, grasping the overall power direction and providing confidence assessments for both types of predictions. Finally, an adaptive gating-based decision factor fusion evaluation module is introduced. This module intelligently integrates multi-dimensional information such as short-term prediction results, medium- and long-term predicted power, wind turbine health indices, prediction confidence, revenue potential, risk indicators, and grid constraints. By dynamically adjusting the weights of each decision factor, it achieves intelligent optimization decisions for power generation modes and target parameters. These recommended power generation modes and target parameters will be directly output to the wind turbine control system, enabling the wind turbine to adjust its operating strategy in real time and adaptively, effectively balancing power generation revenue, operational safety and equipment lifespan. This overcomes the shortcomings of isolated data and static decision-making in existing solutions, and realizes closed-loop management from data perception to intelligent decision-making to precise control. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a wind turbine power generation mode decision-making method based on multi-source data fusion according to an embodiment of the present invention;

[0021] Figure 2 This is a data flow diagram illustrating the wind turbine power generation mode decision-making method based on multi-source data fusion according to an embodiment of the present invention.

[0022] Figure 3This is a flowchart illustrating the GRU-based medium-to-long-term power generation trend prediction of the wind turbine generation mode decision method based on multi-source data fusion according to an embodiment of the present invention, to obtain the medium-to-long-term predicted power time series and its medium-to-long-term prediction confidence.

[0023] Figure 4 This is a flowchart illustrating the decoding and prediction of the implicit encoding vector of future power to obtain the medium- and long-term predicted power time series and its medium- and long-term prediction confidence in the wind turbine power generation mode decision method based on multi-source data fusion according to an embodiment of the present invention.

[0024] Figure 5 This is a block diagram of a wind turbine power generation mode decision system based on multi-source data fusion according to an embodiment of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0027] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0028] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0029] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0030] To address the problems raised in the background art, the present invention proposes a wind turbine power generation mode decision-making scheme based on multi-source data fusion. This scheme aims to solve the challenges of power generation efficiency and grid stability caused by the intermittency and volatility of wind energy, as well as the safety risks of the turbine's own health and under extreme wind conditions, which are difficult to address simultaneously with traditional single-objective control. Specifically, this scheme achieves a closed-loop system from data perception to intelligent decision-making and then to precise control through a novel decision-making method. More specifically, firstly, the system performs in-depth preprocessing and synchronization of raw radar, SCADA, NWP, and CMS vibration data, extracting high-precision wind condition characteristics, operating data at different frequencies, and key wind turbine health indices, providing comprehensive and high-quality input for subsequent decision-making. Next, a CNN-LSTM model is used to predict short-term wind condition changes based on synchronized radar features and high-frequency SCADA data, capturing instantaneous wind speed and turbulence changes and assessing their confidence levels to predict potential risks. Simultaneously, a GRU model is employed to predict medium- to long-term power generation trends using low-frequency SCADA and NWP data, obtaining future power sequences and their confidence levels, laying the foundation for revenue assessment. Most importantly, the system introduces an adaptive gating-based decision factor fusion assessment mechanism. This mechanism intelligently integrates multi-dimensional information such as short-term risks, medium- to long-term revenue potential, wind turbine health index, prediction confidence levels, and grid constraints. By dynamically adjusting the weights of each factor, it optimizes the decision-making for power generation modes and target parameters, ensuring that under different operating scenarios, it maximizes power generation revenue while effectively mitigating risks and protecting the turbines. Finally, the generated recommended power generation modes and target parameters are directly output to the wind turbine control system, enabling real-time, adaptive adjustment of the wind turbine operation strategy, thereby effectively improving the overall operational efficiency and reliability of the wind farm.

[0031] Figure 1 This is a flowchart of a wind turbine power generation mode decision-making method based on multi-source data fusion according to an embodiment of the present invention. Figure 2 This is a data flow diagram illustrating a wind turbine power generation mode decision-making method based on multi-source data fusion according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the wind turbine power generation mode decision-making method based on multi-source data fusion according to an embodiment of the present invention includes the following steps: S100, preprocessing and synchronizing the original radar scan time series, original SCADA time series, original NWP grid data and original CMS vibration data to obtain synchronized radar feature time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series and wind turbine health index; S200, performing short-term wind condition surge analysis based on CNN-LSTM model on the synchronized radar feature time series and synchronized high-frequency SCADA time series. S300: Variable prediction is performed to obtain short-term prediction results and their confidence levels; S400: GRU-based medium-to-long-term power generation trend prediction is performed on the synchronized low-frequency SCADA time series and synchronized NWP time series to obtain the medium-to-long-term predicted power time series and its confidence level; S500: Adaptive gating-based decision factor fusion evaluation is performed on the short-term prediction results and their confidence levels, the medium-to-long-term predicted power time series and its confidence level, and the wind turbine health index to obtain recommended power generation modes and target parameters; S500: The recommended power generation modes and target parameters are output to the wind turbine control system.

[0032] Specifically, in step S100, the original radar scan time series, original SCADA time series, original NWP grid data, and original CMS vibration data are preprocessed and synchronized to obtain synchronized radar feature time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series, and wind turbine health index. It should be understood that due to the complexity of the wind turbine operating environment and the differences in acquisition frequency, data format, noise level, and timestamps between multi-source heterogeneous data (such as original radar scans, SCADA, NWP grids, and CMS vibration data), directly using these original data for high-precision prediction and intelligent decision-making is not feasible and may even introduce a large amount of error and uncertainty. Therefore, in the technical solution of this invention, these original data are further subjected to rigorous preprocessing and synchronization to provide high-quality, highly relevant, and time-aligned unified data input for subsequent short-term wind condition change prediction, medium- and long-term power generation trend prediction, and the final adaptive gating decision factor fusion evaluation.

[0033] More specifically, in this embodiment of the invention, the original radar scan time series, original SCADA time series, original NWP grid data, and original CMS vibration data are preprocessed and synchronized to obtain synchronized radar feature time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series, and wind turbine health index. This includes: performing independent data stream cleaning on the original radar scan time series, original SCADA time series, original NWP grid data, and original CMS vibration data to obtain clean radar time series, clean SCADA time series, single-point NWP time series, and clean CMS time series; from the clean radar... The average wind speed, turbulence intensity, and wind shear index are extracted from each clean radar data point in the time series to obtain the radar feature time series. Multi-frequency time series alignment and synchronization are performed on the radar feature time series, clean SCADA time series, single-point NWP time series, and clean CMS time series to obtain synchronized radar feature time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, and synchronized NWP time series. A fast Fourier transform is performed on the clean CMS time series to obtain a key energy amplitude input vector composed of key energy amplitudes. The key energy amplitude input vector is then input into a trained single-class support vector machine model to obtain the wind turbine health index.

[0034] Specifically, the original radar scan time series, original SCADA time series, original NWP grid data, and original CMS vibration data are independently cleaned into clean radar time series, clean SCADA time series, single-point NWP time series, and clean CMS time series. It should be understood that the raw data collected during wind turbine operation, including the original radar scan time series, original SCADA time series, original NWP grid data, and original CMS vibration data, are affected by various factors such as sensor failure, transmission errors, environmental interference, missing data, or outliers, resulting in noise, inconsistency, or incompleteness. Directly using these unprocessed raw data for subsequent feature extraction, model training, and decision evaluation will seriously affect the accuracy and reliability of the system, and may even lead to incorrect predictions and decisions. Therefore, in the technical solution of this invention, these raw data are further independently cleaned to ensure that the data relied upon for subsequent processing is high-quality, reliable, and consistent. This lays a solid data foundation for the effective operation of the entire decision-making system, improving the performance of the prediction model and the accuracy of the final decision.

[0035] More specifically, in a concrete example of the present invention, firstly, a series of data quality control steps are implemented for the original radar scan time series. This includes denoising the radar echo signal, for example, using median filtering or wavelet transform to suppress random noise and clutter interference. Simultaneously, outliers and invalid scan frames caused by sensor malfunctions or severe weather conditions are identified and removed, for example, by setting physical thresholds or using statistical anomaly detection based on historical data distribution. For short-term data gaps, methods such as linear interpolation or spline interpolation can be used to fill the gaps to maintain the continuity of the time series. Secondly, for the original SCADA time series, the cleaning process focuses on identifying and correcting inconsistencies and anomalies in the operational data. This includes physical constraint checks on key parameters such as wind speed, power, rotational speed, and temperature; for example, power output cannot exceed the rated power, and wind speed cannot be displayed as high speed when the system is stopped. Outliers exceeding physical limits are removed by setting reasonable upper and lower thresholds. Simultaneously, long-term data gaps or duplicate records caused by communication interruptions or data acquisition system malfunctions are identified and processed; for short-term gaps, forward filling or interpolation based on historical patterns can be used. In addition, the unit's operating status (such as shutdown, power limitation, and fault) is marked and filtered to ensure that subsequent analysis is based only on data under normal operating conditions. Secondly, for the original NWP grid data, the main goal of cleaning is to transform it into point data applicable to individual wind turbines and ensure its quality. Since NWP data is provided in grid form, accurate spatial interpolation of the wind turbine's location is first required, for example, using bilinear interpolation or inverse distance-weighted interpolation, to convert the grid data into single-point predicted data at the wind turbine tower height. Simultaneously, the interpolated data undergoes a rationality check to ensure that parameters such as wind speed and direction conform to physical laws, and any interpolation errors or boundary effects are addressed. Finally, for the original CMS vibration data, the cleaning process removes high-frequency noise and transient interference and ensures data integrity. This includes applying digital filters (such as bandpass filters) to isolate vibration signals within a specific frequency range, removing background noise and high-frequency interference unrelated to the unit's health status. Simultaneously, spikes and abnormal amplitudes caused by instantaneous sensor impacts or electromagnetic interference are identified and eliminated. For short-term interruptions or data packet loss that occur during data acquisition, zero-padding or simple interpolation methods based on adjacent data can be used to maintain the continuity of the time series.

[0036] Specifically, average wind speed, turbulence intensity, and wind shear index are extracted from each clean radar data point in the clean radar time series to obtain the radar feature time series. It should be understood that while the original clean radar time series provides rich wind field information, its raw data format (e.g., Doppler velocity spectrum, reflectivity, etc.) is high-dimensional and complex, and not a physical quantity directly applicable to wind turbine power generation mode decisions. The operating strategy and load response of wind turbines depend more directly on the average wind speed, its fluctuation level, and vertical variations. Therefore, in the technical solution of this invention, key features such as average wind speed, turbulence intensity, and wind shear index are further extracted from the clean radar data to transform complex radar observation data into engineering parameters that directly guide wind turbine operation. This reduces data dimensionality, improves the training efficiency and accuracy of subsequent prediction models, and provides precise and quantifiable input for predicting short-term wind condition changes, thereby more effectively assessing potential operational risks and power generation potential.

[0037] More specifically, in a concrete example of the present invention, firstly, for the extraction of average wind speed, Doppler velocity measurement data of a specific area in front of the wind turbine (e.g., the hub height and swept area of ​​the turbine) is used. The average wind speed of that area at that moment can be obtained by spatially weighted averaging or simple arithmetic averaging of all valid Doppler velocity measurements within that area. This involves projecting the radar scan data onto the wind turbine coordinate system and selecting a specific height layer or volumetric unit relevant to the turbine's operation for calculation. Secondly, for the extraction of turbulence intensity, wind speed measurement data within the same time window and spatial area is used. Turbulence intensity is defined as the ratio of the standard deviation of wind speed to the average wind speed. Therefore, after calculating the average wind speed, the standard deviation of the wind speed measurements within that time window is further calculated, and then the standard deviation is divided by the average wind speed to obtain the turbulence intensity at that moment. This quantifies the degree of wind speed fluctuation and is an important indicator for evaluating wind turbine fatigue load and instantaneous power fluctuations. Finally, for the extraction of the wind shear index, wind speed measurement data of different height layers within the swept area of ​​the wind turbine are used. The wind shear index describes the variation of wind speed with height, and is fitted using a power-law model, i.e., V (z) =V ref ×(Z / Z ref ) α Among them, V (z) It is the wind speed at altitude Z, V ref Reference height Z ref The wind speed at a given location, α, is the wind shear index. The wind shear index α can be calculated using wind speed data acquired by radar at at least two different altitude levels (e.g., hub height and blade tip height).

[0038] Specifically, multi-frequency time series alignment and synchronization are performed on radar feature time series, clean SCADA time series, single-point NWP time series, and clean CMS time series to obtain synchronized radar feature time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, and synchronized NWP time series. It should be understood that wind turbine power generation mode decision systems need to comprehensively utilize information from different sensors and data sources, which (such as radar, SCADA, NWP, and CMS) differ in data acquisition frequency, timestamp accuracy, data format, and data update cycle. For example, SCADA systems record high-frequency operating parameters at a second-level frequency, while some SCADA parameters or NWP data are updated only at minute or hourly intervals; radar data is scanned at intervals of several minutes, while CMS data, after processing, provides health indicators at even lower frequencies. This multi-frequency heterogeneity makes it impossible to directly align the various data streams in the time dimension. Without unified time alignment and synchronization, it will be impossible to construct consistent and meaningful feature vectors, thus severely hindering the accuracy of subsequent prediction models, the effectiveness of multi-source data fusion, and the reliability of the final decision. Therefore, in the technical solution of this invention, multi-frequency time series alignment and synchronization are further performed on the radar feature time series, clean SCADA time series, single-point NWP time series, and clean CMS time series to ensure the consistency and comparability of all input data in the time dimension. This provides a unified, high-quality, and time-synchronized input for subsequent short-term wind condition change prediction, medium- and long-term power generation trend prediction, and decision factor fusion evaluation. This effectively eliminates errors introduced by data time inconsistency, improves model training efficiency and prediction accuracy, and ensures that the decision system can make intelligent judgments based on accurate and real-time multi-dimensional information.

[0039] More specifically, in a concrete example of the invention, firstly, a unified time base and target sampling frequency are determined. The system sets a fine, uniform time resolution as the synchronization target for all data streams, for example, one data point per minute. All data will be aligned to this common timestamp sequence. Secondly, timestamp calibration and alignment are performed. For all data streams, their timestamps are first checked and calibrated to eliminate any existing system clock drift or transmission delay, ensuring that all data points accurately correspond to the actual time they occurred or were recorded. Thirdly, data resampling and aggregation are performed. For high-frequency data streams, such as parameters acquired at the second level in the original SCADA, they are aggregated to the target sampling frequency. For example, the SCADA data per second is averaged, summed, or the maximum / minimum value is taken within each minute to generate a synchronized high-frequency SCADA time series; for radar feature time series, they are generated at intervals of several minutes. If their sampling frequency is lower than the target sampling frequency, an interpolation method (such as linear interpolation or nearest neighbor interpolation) is used to upsample them to the target frequency to generate a synchronized radar feature time series. If the sampling frequency is higher than the target frequency, aggregation is performed. For single-point NWP time series, which provide predicted values ​​at hourly or longer intervals, the data is aligned with the target sampling frequency. For example, at each target time point, the most recent NWP predicted value is used for padding, or linear interpolation is performed between two NWP predicted points to generate a synchronized NWP time series. For cleanroom CMS time series, the health index is calculated at hourly or longer intervals. It is aligned with the target sampling frequency, using methods that preserve the previous valid value or interpolation to ensure a corresponding health index at each target time point. For low-frequency SCADA time series, if some SCADA parameters are themselves recorded at a lower frequency (e.g., 10-minute average), they are directly aligned with the target sampling frequency. A synchronized low-frequency SCADA time series is generated by finding or interpolating the most recent low-frequency SCADA value at the target time point. If the low-frequency SCADA is an aggregation result of high-frequency SCADA, further aggregation within a time window (e.g., 10-minute average) is performed on the synchronized high-frequency SCADA, and its timestamp is aligned to the corresponding low-frequency time point.

[0040] Specifically, a Fast Fourier Transform (FFT) is performed on the clean CMS time series to obtain a key energy amplitude input vector composed of key energy amplitudes. It should be understood that the clean CMS time series, i.e., the raw vibration signals of key components of the wind turbine (such as gearboxes, main bearings, generators, etc.), while containing rich information on equipment operating status, often exhibits complex, non-periodic waveforms in the time domain, making it difficult to directly identify potential fault modes through visual inspection or simple statistical methods. Different mechanical faults, such as bearing wear, poor gear meshing, or imbalance, will exhibit specific, identifiable frequency components and changes in energy amplitude in the frequency domain of the vibration signal. Therefore, in the technical solution of this invention, a Fast Fourier Transform is further performed on the clean CMS time series to extract key energy amplitudes, thereby transforming the complex time-domain vibration signal into more diagnostically valuable frequency-domain features. This effectively reveals potential fault information within the equipment, reduces data dimensionality, and provides accurate, sensitive, and physically meaningful input for subsequent wind turbine health index calculations, thereby improving the accuracy and real-time performance of wind turbine health status assessment.

[0041] Specifically, the key energy amplitude input vector is input into a trained single-class support vector machine model to obtain the wind turbine health index. It should be understood that the operating state of wind turbine units is complex and variable, and failures are often low-probability events, resulting in far more data under normal conditions than under fault conditions in actual operation. Traditional supervised learning models require a large number of fault samples for training, which is difficult to obtain in practical applications. Furthermore, the assessment of wind turbine health status requires a mechanism that can accurately identify the boundary between normal and abnormal states, while abnormal states manifest in various unknown or rare patterns. Therefore, in the technical solution of this invention, the key energy amplitude input vector is further input into a trained single-class support vector machine model to obtain the wind turbine health index. This leverages the advantage of the trained single-class support vector machine model, which only needs to learn the features of normal samples, to construct an anomaly detector that can effectively identify deviations from normal operating modes. This overcomes the challenge of scarce fault samples, enabling continuous and sensitive monitoring of wind turbine health status and outputting a quantitative health index. This provides crucial unit status information for subsequent decision factor fusion assessment, thereby effectively supporting predictive maintenance and risk management of wind turbine units.

[0042] More specifically, in a concrete example of this invention, firstly, the training of a single-class support vector machine model is performed. A large amount of clean CMS vibration data collected during normal wind turbine operation is gathered, and the corresponding key energy amplitude input vectors are extracted using the aforementioned method. These vectors constitute the training dataset for OCSVM, and all are labeled as normal samples. The prepared normal sample dataset is input into the OCSVM model for training. The goal of OCSVM is to learn a hyperplane that encloses all normal samples within a minimal hypersphere or hyperplane region. During training, the model determines a decision boundary that maximizes the inclusion of normal samples while excluding abnormal samples. After training, the model can calculate the distance of any input vector to this decision boundary or its score on the decision function. Secondly, the calculation of the wind turbine health index is performed. During the actual operation of the wind turbine, CMS vibration data is collected in real time, and cleaned and FFT transformed using the same method as in the training phase, and the key energy amplitude input vector at the current moment is extracted. The real-time acquired key energy amplitude input vector is input into the trained OCSVM model. The OCSVM model outputs a score or distance value based on the deviation of the input vector from the normal sample distribution. This score or distance value is then mapped to a quantified wind turbine health index. The closer the score is to the decision boundary (i.e., closer to normal samples), the further away the score is from the decision boundary (i.e., more deviating from normal samples), and the lower the health index. Normalization can be used to map the score to a range of 0 to 100. For example, a higher score corresponds to a health index closer to 100, indicating a healthier unit; a lower score corresponds to a health index closer to 0, indicating an abnormal or sub-healthy state of the unit.

[0043] Specifically, in step S200, the radar characteristic time series and the high-frequency SCADA time series after synchronization are used to predict short-term wind condition changes based on a CNN-LSTM model to obtain short-term prediction results and their confidence levels. It should be understood that the operational safety and power generation efficiency of wind turbines are highly susceptible to short-term (minutes to half an hour) wind condition changes, such as sudden strong winds, abrupt changes in wind direction, or a sharp increase in turbulence intensity. These changes cause instantaneous increases in turbine load, drastic fluctuations in power output, and may even lead to shutdowns or component damage. Traditional wind condition prediction methods often struggle to capture such high-frequency, nonlinear instantaneous changes, causing the wind turbine control system to fail to respond in a timely manner, thus missing opportunities to optimize power generation or mitigate risks. Therefore, in the technical solution of this invention, the radar characteristic time series and the high-frequency SCADA time series after synchronization are further used to predict short-term wind condition changes based on a CNN-LSTM model to accurately capture wind speed change trends and potential sudden events in the near future, and to provide prediction confidence levels. This provides high-precision and timely predictive information for the advanced control and risk warning of wind turbine units, thereby effectively improving the unit's operational safety, power generation efficiency, and grid friendliness.

[0044] More specifically, in this embodiment of the invention, short-term wind condition change prediction based on a CNN-LSTM model is performed on the synchronized radar feature time series and the synchronized high-frequency SCADA time series to obtain short-term prediction results and their short-term prediction confidence. This includes: constructing a spatiotemporal input tensor from the synchronized radar feature time series and the synchronized high-frequency SCADA time series to obtain an input tensor; extracting spatial features from the input tensor based on a convolutional neural network model to obtain a spatial feature sequence; performing time dependency modeling on the spatial feature sequence based on a long short-term memory network to obtain an encoded time series vector; and decoding and predicting the encoded time series vector to obtain short-term prediction results and their short-term prediction confidence. The short-term prediction results are short-term wind speed prediction sequences.

[0045] Specifically, a spatiotemporal input tensor is constructed from the synchronized radar feature time series and the synchronized high-frequency SCADA time series to obtain the input tensor. It should be understood that short-term wind condition abrupt change prediction requires simultaneously capturing the dynamic changes in the wind field environment and the real-time response of the wind turbine itself. The synchronized radar feature time series (such as average wind speed, turbulence intensity, and wind shear index) provides refined information about the wind field ahead, while the synchronized high-frequency SCADA time series (such as power, speed, pitch angle, and yaw angle) reflects the turbine's operating status and control actions under the current wind conditions. Although these data are time-aligned, they each represent information with different dimensions and physical meanings. Simply treating them as independent input sequences will not effectively uncover the complex correlations between different data sources and between features within the same data source. Therefore, in the technical solution of this invention, a spatiotemporal input tensor is further constructed from the synchronized radar feature time series and the synchronized high-frequency SCADA time series to integrate these heterogeneous but interconnected data into a unified multidimensional structure, providing a comprehensive input for the subsequent CNN-LSTM model that can simultaneously capture spatial correlation and temporal dependence. This allows for the maximum utilization of collaborative information from multiple data sources, enhancing the model's ability to identify complex wind patterns and improve prediction accuracy, thereby enabling more accurate prediction of short-term wind condition changes.

[0046] Specifically, spatial feature extraction based on a convolutional neural network (CNN) model is performed on the input tensor to obtain a spatial feature sequence. It should be understood that although the aforementioned constructed input tensor integrates synchronized radar features and high-frequency SCADA data, there are complex nonlinear correlations between its various feature dimensions. For example, parameters such as wind speed, turbulence intensity, pitch angle, and power are not independent at any given moment, but rather influence each other and jointly characterize the instantaneous operating state of the wind turbine. If the original feature sequence is directly input into a time series model, the model struggles to effectively capture these instantaneous, local collaborative patterns between features, leading to insufficient information utilization and increasing the model's learning burden and overfitting risk. Therefore, in the technical solution of this invention, spatial feature extraction based on a convolutional neural network (CNN) model is further performed on the input tensor to automatically learn and extract local correlation patterns and high-order abstract features between different feature dimensions in the input tensor. In this way, the original, high-dimensional feature vectors can be transformed into more representative and compact spatial feature sequences, thereby effectively reducing the data dimensionality, enhancing the expressive power of the features, and providing higher quality and more discriminative input for subsequent long short-term memory (LSTM) network time-dependent modeling, thus improving the accuracy of short-term wind condition change prediction.

[0047] Specifically, the spatial feature sequence is modeled on time dependence using a Long Short-Term Memory (LSTM) network to obtain an encoded time-series vector. It should be understood that the CNN model has successfully extracted the local correlation patterns between different feature dimensions in the input tensor, generating the spatial feature sequence. However, wind conditions and the operating status of wind turbines are highly dynamic and time-dependent. Future wind speed changes depend not only on current instantaneous characteristics but also on the historical trends of wind field environment and turbine response over a past period. Traditional neural network models or architectures relying solely on CNNs struggle to effectively capture this long-term, non-linear time dependency, leading to insufficient accuracy in predicting sudden wind changes. Therefore, in this invention, the spatial feature sequence is further modeled on time dependence using an LSTM network to deeply learn and understand the dynamic evolution of wind conditions and turbine operating status over time, condensing it into an encoded time-series vector rich in information. This effectively captures the intrinsic dynamics and long-term trends of wind changes, providing a comprehensive and time-memory-based input for subsequent decoding and prediction, improving the accuracy of short-term wind change prediction and adaptability to complex wind patterns.

[0048] Specifically, the encoded time-series vector is decoded and predicted to obtain short-term prediction results and their confidence levels. These short-term prediction results are short-term wind speed prediction sequences. It should be understood that the CNN-LSTM encoder has condensed complex historical spatiotemporal information into a fixed-dimensional encoded time-series vector. While this vector contains all the information needed to predict future wind conditions, it is not a directly usable prediction result. To achieve advanced control and risk assessment of wind turbines, the system needs to transform this highly abstract encoded information into specific, quantifiable future wind speed prediction values, and also needs to assess the reliability of the prediction results, as any prediction involves uncertainty. Therefore, in the technical solution of this invention, the encoded time-series vector is further decoded and predicted to map the abstract temporal features learned by the encoder into short-term wind speed prediction sequences for multiple future time steps, while simultaneously quantifying the uncertainty of the prediction to obtain the short-term prediction confidence level. This provides the wind turbine decision-making system with clear, specific future wind condition information with reliability assessment, thereby supporting more refined and safer adjustments to operating strategies.

[0049] Specifically, in step S300, a GRU-based medium-to-long-term power generation trend prediction is performed on the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain the medium-to-long-term predicted power time series and its medium-to-long-term prediction confidence level. It should be understood that decisions regarding wind turbine operation planning, grid dispatch, and maintenance plans rely on medium-to-long-term (hours to days) power generation trend predictions, not just short-term wind condition changes. While numerical weather prediction (NWP) provides medium-to-long-term macro-meteorological forecasts, its spatial resolution is low and it has systematic biases, limiting its accuracy when directly applied to power prediction for a single wind turbine. Meanwhile, the historical operating status of the wind turbine itself (such as average power and equipment temperature reflected in low-frequency SCADA data) contains valuable information on its response to specific wind conditions. Therefore, in the technical solution of this invention, the low-frequency SCADA time series and the NWP time series after synchronization are further analyzed using a gated cyclic unit (GRU) model to predict the medium- and long-term power generation trends, thereby obtaining the medium- and long-term predicted power time series and its prediction confidence. This effectively integrates macro-weather forecasts and the historical operating characteristics of the generating units themselves, accurately predicting the power generation trend over a longer period and quantifying the reliability of the prediction. This provides accurate and reliable power prediction data for grid dispatching, electricity market trading, and maintenance window planning, improving the overall economic benefits and operation management level of wind farms.

[0050] Figure 3 This is a flowchart illustrating the GRU-based medium-to-long-term power generation trend prediction of the wind turbine power generation mode decision-making method based on multi-source data fusion according to an embodiment of the present invention, to obtain the medium-to-long-term predicted power time series and its medium-to-long-term prediction confidence. Figure 3 As shown, step S300 includes: S310, constructing historical feature sequences from the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain a historical input sequence; S320, inputting the historical input sequence into the encoder of the GRU model to obtain a future power implicit coding vector; S330, decoding and predicting the future power implicit coding vector to obtain a medium- and long-term predicted power time series and its medium- and long-term prediction confidence.

[0051] Specifically, in step S310, historical feature sequences are constructed from the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain a historical input sequence. It should be understood that medium- and long-term power generation trend prediction depends not only on future macro-meteorological conditions but also on the past operating status of wind turbine units and their response patterns to similar meteorological conditions. The synchronized low-frequency SCADA time series (such as 10-minute average power and average speed) reflects the historical performance and actual operating conditions of the units, while the synchronized NWP time series provides key forecast information such as wind speed and direction for the next few hours to days. If these two types of data are processed in isolation, it will be impossible to establish the intrinsic connection between historical operating conditions and future meteorological forecasts, thus limiting the learning ability and accuracy of the prediction model. Therefore, in the technical solution of this invention, historical feature sequences are further constructed from the synchronized low-frequency SCADA time series and the synchronized NWP time series to effectively integrate the SCADA data reflecting historical operating conditions with the NWP data predicting future environments in the time dimension, forming a comprehensive input sequence that includes both past and future information. This provides a complete and relevant input for subsequent GRU models, improving the accuracy of medium- and long-term power generation trend prediction.

[0052] More specifically, in a concrete example of the invention, firstly, historical and future time windows are defined. A historical backtracking time window length (e.g., the past 24 hours) and a future time window length (e.g., the next 48 hours) for prediction are determined. Simultaneously, a uniform time resolution is determined, such as one data point every 15 minutes or hour. Secondly, features are selected and aligned. Key parameters within the defined historical backtracking time window, such as average output power, average rotor speed, and average pitch angle, are extracted from the synchronized low-frequency SCADA time series. These constitute historical operating features. All predicted data covering the period from the current moment to the end of the future prediction window, such as predicted wind speed, predicted wind direction, predicted temperature, and predicted air pressure, are extracted from the synchronized NWP time series. These constitute future environmental features. Thirdly, feature vectors are constructed. At each uniform time step, the historical SCADA features (if the time point is within the historical window) and the future NWP features (if the time point is within the future window) are combined. For a given prediction task (e.g., predicting power over the next 48 hours), the input sequence will contain SCADA features within the historical window and NWP features covering both historical and future windows. Finally, a historical input sequence is formed. The feature vectors constructed at each time step are arranged chronologically to form a two-dimensional matrix, i.e., the historical input sequence. The dimension of this sequence is (total number of time steps T) total , feature dimension F), where T totalIt is the sum of the historical backtracking window length and the future prediction window length, and F is the total number of selected SCADA features and NWP features. This sequence completely describes the environmental and crew status information from the past to the future and will be used as the input to the GRU model encoder.

[0053] Specifically, in step S320, the historical input sequence is input into the encoder of the GRU model to obtain the future power implicit encoding vector. It should be understood that the aforementioned constructed historical input sequence contains multi-dimensional temporal information spanning the past and future, containing complex dynamic evolution patterns and long-term dependencies. For example, how historical power output responds to specific NWP wind speed patterns, and how the current unit status affects its response to future wind conditions. If this long sequence is directly used for prediction without effective temporal information compression and feature extraction, it will face problems such as high computational complexity and difficulty in capturing key long-term dependencies, thus affecting the accuracy and efficiency of the prediction. Therefore, in the technical solution of this invention, the historical input sequence is further input into the encoder of the Gated Cyclic Unit (GRU) model to obtain the future power implicit encoding vector. This utilizes the GRU model's ability to efficiently process long sequence data, automatically learning and extracting the most critical temporal features and dynamic patterns from the entire historical input sequence. In this way, the lengthy and complex input sequence can be compressed into a fixed-length, highly condensed implicit encoding vector, thereby improving the performance of medium- and long-term power generation trend prediction.

[0054] Specifically, in step S330, the future power implicit coding vector is decoded and predicted to obtain a medium-to-long-term predicted power time series and its medium-to-long-term prediction confidence level. It should be understood that the future power implicit coding vector generated by the GRU encoder is a highly abstract mathematical representation. While it contains a profound understanding of future power generation trends, it is not a directly usable power prediction value. To meet the needs of practical applications such as grid dispatching, power trading, and maintenance planning, this implicit vector must be transformed into a specific, time-step future power prediction sequence. More importantly, any medium-to-long-term prediction inherently involves uncertainty. Providing only a single point prediction value (such as the mean or median) cannot fully reveal potential risks and fluctuation ranges, limiting the accuracy of decision-making. Therefore, in the technical solution of this invention, the future power implicit coding vector is further decoded and predicted to generate a clear medium-to-long-term predicted power time series. Furthermore, the probability distribution of the prediction result is directly estimated through quantile regression, thereby obtaining a quantified medium-to-long-term prediction confidence level. This provides decision-makers with a comprehensive predictive view of future power trajectories and their fluctuation range, thereby supporting more refined and reliable risk management and operational planning.

[0055] Figure 4This is a flowchart illustrating the decoding and prediction of the implicit encoding vector of future power to obtain the medium- and long-term predicted power time series and its medium- and long-term prediction confidence, according to an embodiment of the wind turbine power generation mode decision-making method based on multi-source data fusion. Figure 4 As shown, step S330 includes: S331, inputting the future power hidden encoding vector into the lower quantile head output head to obtain the predicted power lower quantile time series; S332, inputting the future power hidden encoding vector into the median head output head to obtain the predicted power median time series as the medium-to-long-term predicted power time series; S333, inputting the future power hidden encoding vector into the upper quantile head output head to obtain the predicted power upper quantile time series; S334, calculating the medium-to-long-term prediction confidence based on the predicted power lower quantile time series and the predicted power upper quantile time series.

[0056] Specifically, in step S331, the future power implicit encoding vector is input into the lower quantile head output head to obtain the predicted power lower quantile time series. It should be understood that medium- to long-term power generation trend prediction inherently possesses uncertainty. This uncertainty stems from the inherent errors of numerical weather prediction, the randomness of complex flow fields within wind farms, and the minute fluctuations in the state of the wind turbines themselves. When conducting risk assessments and formulating conservative operating strategies, such as determining minimum power generation to fulfill power sales contracts or planning grid reserve capacity, simply knowing the predicted power value (such as the median or mean) is far from sufficient. Decision-makers must understand the lower limit level of power generation achievable in the worst-case scenario. Therefore, in the technical solution of this invention, the future power implicit encoding vector is further input into the lower quantile head output head to obtain the predicted power lower quantile time series, thereby specifically modeling the lower tail of the predicted distribution and directly generating a sequence representing the lower bound of the future power prediction value. This provides the decision-making system with a clear and quantifiable pessimistic scenario prediction, thereby effectively supporting risk-based decision-making and ensuring the safety and economy of wind farm operation under adverse conditions.

[0057] Specifically, in step S332, the future power implicit encoding vector is input into the median point head output head to obtain the predicted power median quantile time series as the medium-to-long-term predicted power time series. It should be understood that in most wind farm daily operation and grid dispatch scenarios, decision-makers are most concerned with the future power generation trajectory. While understanding the predicted fluctuation range (defined by the upper and lower quantiles) is crucial, a clear and representative central trend prediction sequence is the basis for power generation planning, participating in electricity market bidding, and optimizing unit operation strategies. Providing only a probability interval without a central value will leave decision-making without a clear benchmark. Therefore, in the technical solution of this invention, the future power implicit encoding vector is further input into the median point head output head to obtain the predicted power median quantile time series, the central position of the predicted distribution is modeled, and a prediction sequence representing the future power occurrence value is generated, which is then used as the standard medium-to-long-term predicted power time series. This provides a stable, reliable, and statistically significant benchmark prediction for wind farm operation and grid dispatch, thereby simplifying the decision-making process and improving the efficiency and accuracy of operation planning.

[0058] Specifically, in step S333, the future power implicit encoding vector is input into the upper quantile head output head to obtain the predicted power upper quantile time series. It should be understood that in the operation and management of wind farms and participation in electricity market competition, in addition to assessing the worst-case lower limit of power generation, it is also necessary to understand the upper limit of power generation potential under favorable conditions. For example, when making electricity market bids, understanding the maximum achievable power generation helps to formulate a more competitive bidding strategy to capture revenue opportunities during periods of high electricity prices. Simultaneously, when assessing grid acceptance capacity and planning transmission line loads, the upper limit of predicted power prevents grid overload and ensures stable system operation. Therefore, in the technical solution of this invention, the future power implicit encoding vector is further input into the upper quantile head output head to obtain the predicted power upper quantile time series, and the upper tail of the predicted distribution is modeled to generate a sequence representing the upper bound of the predicted future power value. This provides the decision-making system with a clear and quantified optimistic scenario prediction, thereby effectively supporting opportunity-driven decision-making, maximizing the potential economic benefits of wind farms, and ensuring grid security.

[0059] Specifically, in step S334, the medium-to-long-term prediction confidence level is calculated based on the predicted power lower quantile time series and the predicted power upper quantile time series. It should be understood that the preceding steps have already obtained the predicted power lower quantile time series and upper quantile time series representing the lower and upper bounds of the future power prediction, respectively, as well as the median time series representing the trajectory. However, these series themselves do not directly provide a unified, quantitative indicator to describe the reliability or uncertainty of the entire prediction. When evaluating medium-to-long-term power generation plans, it is necessary not only to know the fluctuation range but also to have an intuitive and comparable confidence level measure to determine the extent to which the prediction results are credible, so as to make trade-offs between different prediction scenarios. Therefore, in the technical solution of this invention, the medium-to-long-term prediction confidence level is further calculated based on the predicted power lower quantile time series and the predicted power upper quantile time series, transforming the probability interval of the prediction into a clear, quantitative confidence level indicator. This provides an intuitive assessment of the prediction reliability, making the quantitative assessment of risk and return more accurate, thereby supporting more robust and intelligent operational decisions.

[0060] More specifically, in a concrete example of the invention, firstly, the quantile sequence is obtained. Three key time series are obtained from the preceding steps: the predicted power lower quantile time series (e.g., τ = 0.1, denoted as P). lower(t) Predict the quantile time series of power (τ = 0.5, denoted as P). median(t) ), and the predicted power upper quantile time series (e.g., τ = 0.9, denoted as P). upper(t) These three sequences each have a corresponding power value at each future time step t. Next, the prediction interval width is calculated. At each future time step t, the difference between the upper and lower quantiles is calculated. This difference represents the predicted interval width PIW(t) at that time, and its formula is: PIW(t) = P upper(t) -P lower(t) This width directly reflects the magnitude of the uncertainty in the prediction at that point in time: the wider the width, the higher the uncertainty. Next, normalization is performed to obtain the confidence score. To obtain a more comparable and intuitive confidence score metric, the prediction interval width can be normalized. An effective approach is to use the ratio of the prediction interval width to the center predicted value to represent uncertainty, and then convert it into a confidence score. For example, an uncertainty index U(t) can be defined as: U(t) = PIW(t) / P median(t) To prevent the denominator from being zero, when P... median(t) When the value is close to zero, a small constant can be set.

[0061] Then, the confidence level C(t) for medium- to long-term forecasts can be defined as an inverse mapping of the uncertainty index, for example, by letting C(t) = 1 - U(t), or by using a smoother function to represent C(t), for example, by letting C(t) = exp(-k × U(t)). Here, k is a positive constant used to adjust the sensitivity of the confidence level.

[0062] Finally, a confidence time series is formed. The confidence C(t) calculated at each time step t is arranged in chronological order to form a medium-to-long-term prediction confidence time series corresponding to the medium-to-long-term prediction power time series. Each value in this series is within a standardized range (e.g., 0 to 1), which intuitively represents the model's confidence in its prediction results at each future time.

[0063] Specifically, in step S400, the short-term forecast results and their short-term forecast confidence levels, the medium- and long-term forecast power time series and their medium- and long-term forecast confidence levels, and the wind turbine health index are evaluated using an adaptive gating-based decision factor fusion assessment to obtain the recommended power generation mode and target parameters. It should be understood that the final operation decision of a wind turbine is a complex multi-objective optimization problem, requiring a simultaneous trade-off between short-term operational safety, medium- and long-term power generation revenue, and the health status of the equipment itself. The aforementioned steps have independently generated three key decision factors: short-term forecasts and their confidence levels reflecting the risk of sudden wind changes in the next few minutes; medium- and long-term power forecasts and their confidence levels indicating the economic benefits of power generation in the next few hours to days; and the wind turbine health index quantifying the current risk of equipment wear. These three factors are often conflicting. For example, pursuing maximum short-term power output will exacerbate unit wear and reduce the health index; while overly conservative derating operation will sacrifice medium- and long-term power generation revenue. Using simple weighted averages or fixed rules for decision-making will fail to adapt to the dynamic changes in wind conditions, electricity prices, and unit status, leading to suboptimal decision-making. Therefore, in the technical solution of this invention, the three decision factors are further evaluated using an adaptive gating-based decision factor fusion mechanism to construct an intelligent arbitration mechanism. This mechanism dynamically adjusts the influence of each factor in the final decision based on the real-time changes in their confidence levels. This achieves an adaptive and dynamic balance between short-term safety, long-term benefits, and equipment health, thereby outputting the optimal recommended power generation mode and target parameters under the current comprehensive conditions.

[0064] More specifically, in this embodiment of the invention, the recommended power generation mode and target parameters are obtained by performing an adaptive gating-based decision factor fusion evaluation on the short-term forecast results and their short-term forecast confidence levels, the medium- and long-term forecast power time series and their medium- and long-term forecast confidence levels, and the wind turbine health index. This includes: extracting the maximum predicted wind speed and maximum turbulence intensity from the short-term forecast results; inputting the maximum predicted wind speed and maximum turbulence intensity into a predefined risk scoring function to obtain risk index values; calculating the revenue potential value based on the medium- and long-term forecast power time series and the real-time electricity price time series; inputting the risk index value, the wind turbine health index, the medium- and long-term forecast confidence levels, and the short-term forecast confidence levels into a neural network-based adaptive gating weight generator to obtain a gating weight vector; and performing scenario decision-making based on weighted factors and a rule base on the risk index value, revenue potential value, wind turbine health index, and grid constraints based on the gating weight vector to obtain the recommended power generation mode and target parameters.

[0065] Specifically, the maximum predicted wind speed and maximum turbulence intensity are extracted from the short-term forecast results. It should be understood that the short-term forecast results are a detailed time series containing wind speed and turbulence intensity at multiple points in the next few minutes to half an hour. Directly using the entire series for risk assessment would be redundant and not intuitive. For the operational safety of wind turbines, the most threatening events are often extreme wind events, namely, peak wind speeds and peak turbulence intensities occurring in the short term. These peaks directly relate to whether the turbine will withstand mechanical and aerodynamic loads exceeding its design limits, and are key indicators for triggering protective shutdowns or implementing emergency derating strategies. Therefore, in the technical solution of this invention, the maximum predicted wind speed and maximum turbulence intensity are further extracted from the short-term forecast results to condense the complex forecast series into two core indicators that best represent short-term extreme risks. This provides a clear and quantifiable short-term risk signal for the subsequent decision fusion module, enabling rapid and accurate responses to potentially destructive wind conditions and effectively ensuring the structural safety of the wind turbine.

[0066] Specifically, the maximum predicted wind speed and maximum turbulence intensity are input into a predefined risk scoring function to obtain a risk index value. It should be understood that while the maximum predicted wind speed and maximum turbulence intensity extracted in the previous step are key risk signals, they are two independent parameters with different physical units and influencing mechanisms. When making comprehensive decisions, the decision-making system needs a unified, standardized index to quantify the overall short-term operational risk, rather than dealing with two independent, interrelated physical quantities. For example, the risk of high wind speed and high turbulence occurring simultaneously is far greater than the risk of only one parameter being high; this nonlinear coupling effect is difficult to assess by directly comparing the original values. Therefore, in the technical solution of this invention, the maximum predicted wind speed and maximum turbulence intensity are further input into a predefined risk scoring function to obtain a risk index value. This integrates these two independent physical quantities through a clear mathematical model, transforming them into a single, dimensionless risk index. In this way, the complex short-term aerodynamic load risk can be quantified into an intuitive and comparable value, thereby simplifying the input of the subsequent decision fusion module and enabling it to more directly and accurately assess and weigh short-term operational risks.

[0067] Specifically, the potential revenue value is calculated based on medium- and long-term predicted power time series and real-time electricity price time series. It should be understood that the medium- and long-term predicted power time series only reflects the potential power generation of wind turbines under standard operating conditions within a certain period; it does not directly reflect economic value. In the modern electricity market environment, electricity prices fluctuate in real time, and the economic returns of the same power generation vary significantly at different points in time. If the decision-making system only plans based on power generation and ignores the dynamic changes in electricity prices, it will be unable to seize profit opportunities during periods of high electricity prices and effectively avoid inefficient power generation during periods of low electricity prices, thus failing to maximize the overall economic benefits of the wind farm. Therefore, in the technical solution of this invention, the potential revenue value is further calculated based on the medium- and long-term predicted power time series and real-time electricity price time series. This deeply couples the physical power generation forecast with market economic signals, quantifying the potential economic returns of future power generation plans. This provides the decision fusion module with a clear, economically oriented decision-making basis, thereby driving the power generation mode towards maximizing economic benefits.

[0068] More specifically, in a concrete example of the invention, firstly, data sequences are acquired and aligned. A medium-to-long-term predicted power time series generated in the preceding steps is acquired, which represents the predicted power value for each time step (e.g., every 15 minutes or hour) within a future period (e.g., 48 hours). Simultaneously, a corresponding real-time electricity price time series is acquired from an electricity market system or external data source, providing the predicted or real-time electricity price for each time step within the same future time period. The two time series are ensured to be perfectly aligned in terms of time axis and resolution. Secondly, instantaneous revenue is calculated step-by-step. At each aligned time step t, the predicted power value P at that moment is calculated. predicted(t) Multiplying this by the electricity price Price(t) at that time step yields the instantaneous revenue potential R at that time step. instant(t) Its calculation formula is: R instant(t) =P predicted(t) ×Price(t)×Δt, where Δt is the length of the time step (e.g., 1 hour or 0.25 hours), used to convert the product of power (energy / time) and electricity price (currency / energy) into the total revenue (currency) for that time period.

[0069] Next, the total return potential is calculated by summing up the instantaneous return potentials at all time steps within the entire medium-to-long-term forecast window. This summates to obtain a scalar value representing the total economic return over the entire forecast period, i.e., the return potential value RP. The formula for RP is: RP = ∑R instant(t) The summation symbol ∑ represents the accumulation of all time steps t within the prediction time window.

[0070] Finally, the revenue potential value is output. The calculated revenue potential value RP is used as the final output. This single value highly summarizes the maximum potential economic value that the wind turbine can create under the current predicted power generation and electricity price trends. This value will serve as the core economic benefit indicator and will be input into the subsequent adaptive gating decision fusion module.

[0071] Specifically, risk index values, turbine health index, medium-to-long-term forecast confidence levels, and short-term forecast confidence levels are input into a neural network-based adaptive gating weight generator to obtain a gating weight vector. It should be understood that the final operational decision for wind turbine units requires a trade-off between several conflicting objectives: short-term operational safety, long-term equipment health, and medium-to-long-term economic benefits. The preceding steps have quantified these objectives into risk index values, turbine health index, and revenue potential values. However, dynamically determining the relative importance of these core decision factors at a specific moment is a highly complex nonlinear problem. For example, when the turbine health index is extremely low, even if the revenue potential is high, equipment safety should be prioritized; when the forecast confidence level is very low, the reliability of the revenue potential value calculated based on that forecast should also decrease accordingly. Using fixed weights or simple linear rules cannot capture this complex and dynamic dependency, leading to rigid decision-making and suboptimal results. Therefore, in the technical solution of this invention, the risk index value, wind turbine health index, and the confidence levels of the two predictions are further input into an adaptive gating weight generator based on a neural network to obtain a gating weight vector. This utilizes the powerful nonlinear mapping capability of the neural network to learn and generate a set of dynamic weights that reflect the relative importance of each decision factor under the current comprehensive operating conditions. This enables the final decision fusion process to be adaptive, ensuring that the most reasonable trade-off is made between the three core concerns of safety, health, and profitability at any given time.

[0072] More specifically, in a concrete example of the present invention, firstly, an input vector is constructed. The four key scalar values ​​calculated in the preceding steps are combined to form a fixed-dimensional input vector. This vector includes: a risk indicator value representing the risk of short-term extreme wind conditions, a wind turbine health index quantifying the current equipment status, a medium-to-long-term prediction confidence level reflecting the reliability of medium-to-long-term power prediction, and a short-term prediction confidence level characterizing the reliability of short-term prediction. Secondly, a weight generator network structure is designed. This adaptive gated weight generator is a feedforward neural network, whose structure consists of one or more hidden layers and an output layer. The number of nodes in the input layer is the same as the dimension of the input vector, i.e., 4. The hidden layers use nonlinear activation functions such as ReLU to learn the complex nonlinear relationships between input factors. The number of nodes in the output layer corresponds to the number of decision factors that need to be weighted. In this invention, corresponding to the three aspects of risk, health, and return, the output layer has 3 nodes. Thirdly, a Softmax activation function is applied. To ensure that the generated weights have clear physical meaning, i.e., the sum of all weights is 1 and each weight is between 0 and 1, a Softmax activation function is applied to the output layer of the neural network. The Softmax function transforms the raw output values ​​of the previous layer into a probability distribution. Thus, the three output values ​​can be directly interpreted as the relative importance weights of risk, health, and return at the current moment. Finally, a gating weight vector is generated. The constructed input vector is fed into this trained adaptive gating weight generator. After forward propagation computation, a three-dimensional vector, the gating weight vector, is obtained at the output layer. The three elements of this vector correspond to the dynamic weights of the three objectives: risk aversion, health maintenance, and return pursuit. For example, the vector [0.7, 0.2, 0.1] indicates that risk should be considered with 70% weight, health with 20% weight, and return with only 10% weight.

[0073] Specifically, based on the gating weight vector, scenario-based decision-making using weighted factors and a rule base is performed on risk index values, return potential values, wind turbine health index, and grid constraints to obtain recommended power generation modes and target parameters. It's worth noting that the grid constraint here refers to the active power ceiling. It should be understood that while the gating weight vector generated in the previous step provides dynamic and quantitative guidance for decision-making, it is not the final control command. The decision-making system still needs a clear mechanism to apply these weights to specific decision factors (risk, return, health) and combine them with insurmountable external rigid conditions (such as grid constraints) to ultimately transform them into recommended power generation modes and target parameters that the wind turbine can execute. Without subsequent decision-making logic, the entire system cannot close the loop. Furthermore, certain extreme situations (such as emergency grid commands or severe turbine failures) require mandatory intervention beyond conventional weighted balancing, necessitating a rule base for handling. Therefore, in the technical solution of this invention, based on a gating weight vector, scenario-based decision-making is performed on risk index values, revenue potential values, wind turbine health index, and grid constraints using weighted factors and a rule base to obtain recommended power generation modes and target parameters. This constructs the final decision-making layer, transforming abstract weights and quantified indicators into specific, executable operating strategies. This ensures that the final decision is not only intelligent and adaptive, but also comprehensive and safe, ultimately enabling precise control of the wind turbine units.

[0074] More specifically, in a concrete example of the present invention, firstly, a comprehensive decision objective function is constructed. An objective function is designed to maximize a comprehensive score, which is a weighted combination of returns, risks, and health. The gated weight vector [w] generated in the preceding steps is then... risk ,w health ,w revenue ] are used as weights and multiplied by the corresponding decision factors. Where, w risk To mitigate risk, w health For health maintenance weight, w revenue Weighting based on returns, Revenue potential value potential Health Index of Fans Index It refers to the benefit item (the higher the value, the better), while the risk indicator value is Risk. Indicator Since this is a cost item (the higher the value, the worse the situation), the objective function can be designed as follows:

[0075] Score = w revenue ×f(Revenue potential )-w risk ×g(Risk Indicator )+w health ×h(Health Index )

[0076] Here, f, g, and h are functions used to normalize different decision factors to a comparable scale. The core idea of ​​this objective function is to find the power generation mode that maximizes the overall score under the current weight allocation. Secondly, the power generation modes and target parameter sets are defined. A discrete set of power generation modes is predefined, for example: {maximum power output mode, standard power generation mode, flexible derating mode, conservative derating mode, emergency shutdown mode}. Each mode corresponds to a specific set of target parameters, such as target active power, power change rate limit, and pitch angle adjustment range. For example, the flexible derating mode corresponds to a target power of 80% of the rated power. Thirdly, priority decisions based on a rule base are executed. Before weighted evaluation, high-priority rules in the rule base are checked first. This rule base contains a series of IF-THEN conditional statements to handle scenarios that must be enforced. For example:

[0077] IF (Grid Command = Emergency Power Reduction to X MW) THEN Recommended Generation Mode = Flexible Degradation Mode, Target Parameter = {Active Power = X}.

[0078] IF (Wind turbine health index < preset severe fault threshold) THEN Recommended power generation mode = emergency shutdown mode.

[0079] If (risk index value > preset extreme danger threshold) THEN, the recommended power generation mode equals the emergency shutdown mode. If any rule in the rule base is triggered, the recommended power generation mode and target parameters specified by that rule are directly output, skipping subsequent weighted evaluation. Finally, a weighted mode selection is performed. If no high-priority rule is triggered, the system iterates through all feasible power generation modes. For each mode, the system evaluates its expected results after execution and substitutes them into the aforementioned comprehensive decision objective function to calculate its score. For example, the maximum power output mode has the greatest potential for profit, but it may lead to an increase in risk indicators or have an impact on health. The system will select the power generation mode that results in the highest comprehensive score and output it and its corresponding target parameters as the final recommended power generation mode and target parameters.

[0080] Specifically, in step S500, the recommended power generation mode and target parameters are output to the wind turbine control system. It should be understood that all the aforementioned data fusion, prediction, evaluation, and decision-making steps ultimately result in a set of recommended power generation modes and target parameters representing the current optimal operating strategy. However, if this information merely remains at the decision-making system level without being received and executed by the physical execution mechanisms, the entire complex technical solution will lose its practical meaning and will not have any substantial impact on the operation of the wind turbine. The value of decision-making lies in execution; the goal of the entire system is to optimize the actual operating state of the wind turbine. Therefore, in the technical solution of this invention, the recommended power generation mode and target parameters are further output to the wind turbine control system to complete the closed loop from intelligent decision-making to physical control, transforming the results of upper-level optimization into specific actions of the lower-level execution mechanisms. This allows for direct intervention and optimization of the wind turbine's operating behavior, thereby achieving the ultimate technical goals of improving power generation efficiency, ensuring operational safety, and extending equipment lifespan.

[0081] In summary, the wind turbine power generation mode decision-making method based on multi-source data fusion according to embodiments of the present invention is elucidated. It refines and synchronizes multi-source heterogeneous data such as raw radar scans, SCADA, NWP, and CMS vibration data to extract synchronized radar features, high- and low-frequency SCADA and NWP time series, and key wind turbine health indices. Based on this, an advanced deep learning model, namely the CNN-LSTM model, is used to predict short-term wind condition changes, capturing instantaneous wind speed and turbulence variations. Simultaneously, a GRU model is used to predict medium- and long-term power generation trends, grasping the overall power direction and providing confidence assessments for both types of predictions. Finally, an adaptive gating-based decision factor fusion evaluation module is introduced. This module intelligently integrates multi-dimensional information such as short-term prediction results, medium- and long-term predicted power, wind turbine health indices, prediction confidence, revenue potential, risk indicators, and grid constraints. By dynamically adjusting the weights of each decision factor, it achieves intelligent optimization decisions on power generation modes and target parameters. These recommended power generation modes and target parameters will be directly output to the wind turbine control system, enabling the wind turbine to adjust its operating strategy in real time and adaptively, effectively balancing power generation revenue, operational safety and equipment lifespan. This overcomes the shortcomings of isolated data and static decision-making in existing solutions, and realizes closed-loop management from data perception to intelligent decision-making to precise control.

[0082] This invention also provides a wind turbine power generation mode decision system based on multi-source data fusion. Figure 5 This is a block diagram of a wind turbine power generation mode decision system based on multi-source data fusion according to an embodiment of the present invention. Figure 5As shown, the wind turbine power generation mode decision system 500 based on multi-source data fusion according to an embodiment of the present invention includes: a data processing module 510, used to preprocess and synchronize the original radar scan time series, original SCADA time series, original NWP grid data and original CMS vibration data to obtain synchronized radar feature time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series and wind turbine health index; and a wind condition change prediction module 520, used to perform short-term wind condition change prediction based on a CNN-LSTM model on the synchronized radar feature time series and synchronized high-frequency SCADA time series. The system obtains short-term forecast results and their short-term forecast confidence levels; the power generation trend forecast module 530 is used to perform GRU-based medium- and long-term power generation trend forecasts on the synchronized low-frequency SCADA time series and synchronized NWP time series to obtain the medium- and long-term predicted power time series and their medium- and long-term prediction confidence levels; the fusion evaluation module 540 is used to perform adaptive gating-based decision factor fusion evaluation on the short-term forecast results and their short-term forecast confidence levels, the medium- and long-term predicted power time series and their medium- and long-term prediction confidence levels, and the wind turbine health index to obtain recommended power generation modes and target parameters; the parameter output module 550 is used to output the recommended power generation modes and target parameters to the wind turbine control system.

[0083] Furthermore, the data processing module 510 is specifically used to: perform independent data stream cleaning on the original radar scanning time series, original SCADA time series, original NWP grid data, and original CMS vibration data to obtain a clean radar time series, a clean SCADA time series, a single-point NWP time series, and a clean CMS time series; extract average wind speed, turbulence intensity, and wind shear index from each clean radar data in the clean radar time series to obtain a radar feature time series; and perform multi-frequency time series alignment and synchronization on the radar feature time series, clean SCADA time series, single-point NWP time series, and clean CMS time series to obtain a synchronized radar feature time series, a synchronized high-frequency SCADA time series, a synchronized low-frequency SCADA time series, and a synchronized NWP time series.

[0084] As described above, the wind turbine power generation mode decision-making system 500 based on multi-source data fusion according to embodiments of the present invention can be implemented in various wireless terminals, such as servers with wind turbine power generation mode decision-making algorithms based on multi-source data fusion. In one possible implementation, the wind turbine power generation mode decision-making system 500 based on multi-source data fusion according to embodiments of the present invention can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the wind turbine power generation mode decision-making system 500 based on multi-source data fusion can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the wind turbine power generation mode decision-making system 500 based on multi-source data fusion can also be one of many hardware modules of the wireless terminal.

[0085] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for determining the power generation mode of a wind turbine based on multi-source data fusion, characterized in that, include: The original radar scan time series, original SCADA time series, original NWP grid data and original CMS vibration data are preprocessed and synchronized to obtain synchronized radar characteristic time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series and wind turbine health index. Short-term wind condition change predictions based on the CNN-LSTM model were performed on the radar characteristic time series and the high-frequency SCADA time series after synchronization to obtain the short-term prediction results and their short-term prediction confidence. GRU-based medium- and long-term power generation trend prediction is performed on the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain the medium- and long-term predicted power time series and its medium- and long-term prediction confidence. An adaptive gating-based decision factor fusion evaluation was conducted on the short-term forecast results and their short-term forecast confidence, the medium- and long-term forecast power time series and their medium- and long-term forecast confidence, and the wind turbine health index to obtain the recommended power generation mode and target parameters. The recommended power generation mode and target parameters are output to the wind turbine control system.

2. The wind turbine power generation mode decision-making method based on multi-source data fusion according to claim 1, characterized in that, The original radar scan time series, original SCADA time series, original NWP grid data, and original CMS vibration data are preprocessed and synchronized to obtain synchronized radar characteristic time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series, and wind turbine health index, including: Independent data stream cleaning was performed on the original radar scan time series, original SCADA time series, original NWP grid data and original CMS vibration data to obtain clean radar time series, clean SCADA time series, single-point NWP time series and clean CMS time series. The average wind speed, turbulence intensity, and wind shear index are extracted from each clean radar data point in the clean radar time series to obtain the radar characteristic time series. Multi-frequency time series alignment and synchronization were performed on the radar characteristic time series, clean SCADA time series, single-point NWP time series and clean CMS time series to obtain synchronized radar characteristic time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series and synchronized NWP time series.

3. The wind turbine power generation mode decision-making method based on multi-source data fusion according to claim 2, characterized in that, Preprocessing and synchronization of the original radar scan time series, original SCADA time series, original NWP grid data, and original CMS vibration data yield synchronized radar characteristic time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series, and wind turbine health index, including: A fast Fourier transform is performed on the clean CMS time series to obtain a key energy amplitude input vector composed of key energy amplitudes; The key energy amplitude input vector is input into the trained single-class support vector machine model to obtain the wind turbine health index.

4. The wind turbine power generation mode decision-making method based on multi-source data fusion according to claim 1, characterized in that, Short-term wind condition abrupt changes are predicted using a CNN-LSTM model based on the radar characteristic time series and the high-frequency SCADA time series after synchronization to obtain short-term prediction results and their confidence levels, including: Spatiotemporal input tensors are constructed from the synchronized radar feature time series and the synchronized high-frequency SCADA time series to obtain the input tensor; Spatial feature sequence is obtained by performing spatial feature extraction on the input tensor based on a convolutional neural network model. Temporal dependency modeling of spatial feature sequences based on long short-term memory networks is performed to obtain encoded temporal vectors; The encoded time-series vector is decoded and predicted to obtain the short-term prediction result and its short-term prediction confidence. The short-term prediction result is a short-term wind speed prediction sequence.

5. The wind turbine power generation mode decision-making method based on multi-source data fusion according to claim 1, characterized in that, GRU-based medium- and long-term power generation trend predictions were performed on the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain the medium- and long-term predicted power time series and their medium- and long-term prediction confidence levels, including: Historical feature sequences are constructed from the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain the historical input sequence; The historical input sequence is input into the encoder of the GRU model to obtain the future power hidden encoding vector; Decode and predict the hidden coding vector of future power to obtain the medium- and long-term predicted power time series and its medium- and long-term prediction confidence.

6. The wind turbine power generation mode decision-making method based on multi-source data fusion according to claim 5, characterized in that, Decoding and predicting the future power latent coding vector yields the medium- and long-term predicted power time series and its medium- and long-term prediction confidence, including: Input the future power hidden encoding vector into the lower quantile head output head to obtain the predicted power lower quantile time series; The future power implicit coding vector is input into the midpoint head and output head to obtain the predicted power quantile time series as the medium-to-long-term predicted power time series. Input the future power hidden encoding vector into the upper quantile head output head to obtain the predicted power upper quantile time series; The confidence level of medium- and long-term predictions is calculated based on the lower quantile time series and the upper quantile time series of predicted power.

7. The wind turbine power generation mode decision-making method based on multi-source data fusion according to claim 1, characterized in that, An adaptive gating-based decision factor fusion evaluation is performed on the short-term forecast results and their short-term forecast confidence levels, the medium- and long-term forecast power time series and their medium- and long-term forecast confidence levels, and the wind turbine health index to obtain recommended power generation modes and target parameters, including: The maximum predicted wind speed and maximum turbulence intensity are extracted from the short-term prediction results. Input the maximum predicted wind speed and the maximum turbulence intensity into a predefined risk scoring function to obtain risk index values; Based on medium- and long-term forecast power time series and real-time electricity price time series, calculate the revenue potential value; The risk indicator values, wind turbine health index, medium- and long-term prediction confidence and short-term prediction confidence are input into a neural network-based adaptive gating weight generator to obtain the gating weight vector. Based on the gating weight vector, scenario-based decision-making based on weighted factors and rule base is performed on risk index values, revenue potential values, wind turbine health index and grid constraints to obtain recommended power generation modes and target parameters.

8. The wind turbine power generation mode decision-making method based on multi-source data fusion according to claim 7, characterized in that, The power grid constraint is the upper limit of active power.

9. A wind turbine power generation mode decision-making system based on multi-source data fusion, characterized in that, include: The data processing module is used to preprocess and synchronize the original radar scan time series, original SCADA time series, original NWP grid data and original CMS vibration data to obtain synchronized radar characteristic time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series, synchronized NWP time series and wind turbine health index. The wind condition change prediction module is used to perform short-term wind condition change prediction based on the CNN-LSTM model on the radar characteristic time series and the high-frequency SCADA time series after synchronization to obtain the short-term prediction results and their short-term prediction confidence. The power generation trend prediction module is used to perform GRU-based medium- and long-term power generation trend prediction on the synchronized low-frequency SCADA time series and the synchronized NWP time series to obtain the medium- and long-term predicted power time series and its medium- and long-term prediction confidence. The fusion evaluation module is used to perform adaptive gating-based decision factor fusion evaluation on the short-term forecast results and their short-term forecast confidence, the medium- and long-term forecast power time series and their medium- and long-term forecast confidence, and the wind turbine health index to obtain recommended power generation modes and target parameters. The parameter output module is used to output the recommended power generation mode and target parameters to the wind turbine control system.

10. The wind turbine power generation mode decision system based on multi-source data fusion according to claim 9, characterized in that, The data processing module is further used for: Independent data stream cleaning was performed on the original radar scan time series, original SCADA time series, original NWP grid data and original CMS vibration data to obtain clean radar time series, clean SCADA time series, single-point NWP time series and clean CMS time series. The average wind speed, turbulence intensity, and wind shear index are extracted from each clean radar data point in the clean radar time series to obtain the radar characteristic time series. Multi-frequency time series alignment and synchronization were performed on the radar characteristic time series, clean SCADA time series, single-point NWP time series and clean CMS time series to obtain synchronized radar characteristic time series, synchronized high-frequency SCADA time series, synchronized low-frequency SCADA time series and synchronized NWP time series.

Images