Method for predicting monthly scale of fan shutdown risk level, model training method thereof and electronic device

Abstract

The application discloses a training method of a month-scale prediction model of a fan shutdown risk level, comprising the following steps: constructing a training sample set by using a historical meteorological data set of a target region in a first set historical time period and a historical wind speed data set of the target region in the first set historical time period; fitting the historical meteorological data set and the historical wind speed data set in the training sample set to obtain a final target fitting result; generating a maximum wind speed simulation data set in a second set historical time period according to a historical meteorological data set of the target region in the second set historical time period and the final target fitting result; and training an initial model by using the maximum wind speed simulation data set as a training sample to obtain the month-scale prediction model. The application also discloses a prediction method and an electronic device. The fan shutdown risk in a future month can be accurately predicted by using the prediction method.

Description

Technical Field

[0001] This application relates to the field of meteorological data processing technology, specifically to a training method for a monthly-scale prediction model of wind turbine shutdown risk level, a monthly-scale prediction of wind turbine shutdown risk level, and an electronic device. Background Technology

[0002] Wind energy, as a clean energy source, can be used for power generation. However, the natural fluctuations in wind speed make wind power output highly uncertain, especially during periods of extremely high winds (such as wind speeds exceeding rated or cut-off thresholds). This can easily lead to wind turbine shutdowns, equipment fatigue damage, and grid connection stability issues, posing significant challenges to wind farm operation and system dispatch. Summary of the Invention

[0003] This application aims to address one of the technical problems in related technologies to a certain extent. To this end, this application provides a training method for a monthly-scale prediction model of wind turbine downtime risk levels, a monthly-scale prediction method for wind turbine downtime risk levels, and an electronic device.

[0004] As a first aspect of this application, a training method for a monthly-scale prediction model of wind turbine shutdown risk levels is provided, the training method comprising:

[0005] A training sample set is constructed using the historical meteorological dataset of the target area within the first set historical time period and the historical wind speed dataset of the target area within the first set historical time period.

[0006] The historical meteorological dataset and the historical wind speed dataset in the training sample set are fitted to obtain the final target fitting result, which is used to characterize the correspondence between meteorological data and wind speed data.

[0007] The maximum wind speed simulation dataset for the target area within the second set historical time period is generated based on the historical meteorological dataset of the target area within the second set historical time period and the final target fitting result, wherein the time span of the second set historical time period is greater than the time span of the first set historical time period.

[0008] The initial model was trained using the maximum wind speed simulation dataset as training samples to obtain the monthly-scale prediction model.

[0009] Optionally, fitting the historical meteorological dataset in the training sample set with the historical wind speed dataset includes:

[0010] Multiple regression models are used to fit the data in the training sample set to obtain multiple preliminary fitting results, wherein the preliminary fitting results are used to characterize the correspondence between meteorological data and wind speed data;

[0011] Multiple preliminary fitting results are integrated with equal weights to obtain an initial target fitting result, wherein the initial target fitting result is used to characterize the correspondence between meteorological data and wind speed data;

[0012] Using the historical meteorological dataset as input, an initial predicted wind speed dataset is obtained;

[0013] The initial target fitting result is corrected by using the difference between the historical wind speed dataset and the initial predicted wind speed dataset to obtain the final target fitting result.

[0014] Optionally, the step of correcting the initial target fitting result by utilizing the difference between the historical wind speed dataset and the initial predicted wind speed dataset to obtain the final target fitting result includes:

[0015] Determine the empirical distribution function of the initial predicted wind speed dataset and the empirical distribution function of the historical wind speed dataset;

[0016] The empirical distribution function of the initial predicted wind speed dataset is matched with the empirical distribution function of the historical wind speed dataset. At the same cumulative probability position, the predicted wind speed data in the initial predicted wind speed dataset is replaced with the historical wind speed data in the historical wind speed dataset to obtain the final fitting result.

[0017] Optionally, the plurality of said regression models include at least two of the following regression models:

[0018] Random forest regression model, support vector machine regression model, gradient boosting tree regression model, and lightweight gradient boosting machine regression model.

[0019] Optionally, the meteorological data in the historical meteorological dataset includes at least one of the following:

[0020] Ten atmospheric circulation indices, six atmospheric teleconnection indices, six ocean-atmosphere indices, and eleven sea ice and snow cover related indices;

[0021] The wind speed data in the historical wind speed dataset includes wind speed observation data of the wind farm's meteorological tower in the target area at a hub height of 100 meters.

[0022] Optionally, the initial model includes a cascaded temporal convolutional network, a gated recurrent unit, an attention mechanism layer, and a Softmax classifier;

[0023] The input data of the temporal convolutional network is the wind speed observation data of the previous N consecutive months, and the output data of the attention mechanism layer is the wind speed prediction data for the next month.

[0024] The Softmax classifier is used to output the risk level of wind turbine shutdown based on wind speed forecast data for the next month.

[0025] Optionally, the classifier is used to output the probability distribution of various risk levels, wherein the more days with maximum monthly winds, the higher the risk level.

[0026] Optionally, the loss function of the initial model satisfies the following formula:

[0027] ;

[0028] in, p t Let be the model's predicted probability for the true class, and ;

[0029] For category weight items;

[0030] γ As a focusing factor, it controls the penalty intensity for easily classified samples;

[0031] logp t This is the standard cross-entropy term.

[0032] As a second aspect of this application, a monthly-scale prediction method for wind turbine shutdown risk levels is provided, characterized in that the monthly-scale prediction method includes:

[0033] The wind speed observation data of the previous N consecutive months is input into the monthly scale prediction model to obtain the wind speed prediction data for the next month and output the risk level of wind turbine shutdown. Here, N is a positive integer and N>1. The monthly scale prediction model is a model trained using the training method provided in the first aspect of this application.

[0034] As a third aspect of this application, an electronic device is provided, wherein the electronic device comprises:

[0035] One or more processors;

[0036] A memory having stored one or more computer programs that, when executed by one or more processors, cause the one or more processors to implement the methods provided by the first aspect and / or the second aspect.

[0037] Wind speed data is mostly obtained from observations at wind farms, which are usually built relatively recently. Therefore, using wind speed data alone for model training may result in inaccurate predictions due to insufficient samples.

[0038] In this embodiment, the historical meteorological data in the historical meteorological dataset can come from observation data from meteorological stations in the target region. Each region has meteorological stations, and the history of each regional meteorological station is longer than the history of the local wind farms. In the training method provided in this application, a fitting relationship between meteorological data and wind speed is first established. Then, this fitting relationship is used to construct a large amount of simulated wind speed data as training samples to train the initial model. The resulting monthly-scale prediction model can predict the risk of wind turbine shutdown in future months with relatively high accuracy. Attached Figure Description

[0039] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0040] Figure 1 This is a flowchart of the training method provided in the embodiments of this application;

[0041] Figure 2 This is a flowchart of one embodiment of step S120 in the present application;

[0042] Figure 3 This is a flowchart of one embodiment of step S124 in the present application.

[0043] Figure 4 This is a flowchart of the prediction method provided in the embodiments of this application;

[0044] Figure 5 The cumulative probability density distribution of observed wind speed, simulated wind speed, and PDF-corrected simulated wind speed (test set).

[0045] Figure 6 Time series of observed wind speed, simulated wind speed, and PDF-corrected simulated wind speed (test set);

[0046] Figure 7 A ladder diagram of monthly wind power risk levels and predicted risk levels (test set) is generated.

[0047] Figure 8 This is a schematic diagram of a module of one embodiment of the electronic device provided in this application.

[0048] Figure 9 This is a schematic diagram of the modules of the computer-readable medium provided in the embodiments of this application.

[0049] Explanation of reference numerals in the attached figures

[0050] 101: Processor; 102: Memory

[0051] 103: I / O Interface 104: Bus Detailed Implementation

[0052] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described are intended to explain this application and should not be construed as limiting it.

[0053] The terms "an embodiment," "example," or "example" used in this specification refer to a particular feature, structure, or characteristic described in connection with the embodiment itself that may be included in at least one embodiment disclosed in this application. The phrase "in an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0054] As a first aspect of the present invention, a training method for a monthly-scale prediction model of wind turbine outage risk levels is provided, wherein, as Figure 1 As shown, the training method includes:

[0055] In step S110, a training sample set is constructed using the historical meteorological dataset of the target area within the first set historical time period and the historical wind speed dataset of the target area within the first set historical time period.

[0056] In step S120, the historical meteorological dataset and the historical wind speed dataset in the training sample set are fitted to obtain the final target fitting result, which is used to characterize the correspondence between meteorological data and wind speed data.

[0057] In step S130, a maximum wind speed simulation dataset for the target area within the second set historical time period is generated based on the historical meteorological dataset of the target area within the second set historical time period and the final target fitting result, wherein the time span of the second set historical time period is greater than the time span of the first set historical time period.

[0058] In step S140, the initial model is trained using the maximum wind speed simulation dataset as training samples to obtain the monthly scale prediction model.

[0059] Wind speed data is mostly obtained from observations at wind farms, which are usually built relatively recently. Therefore, using wind speed data alone for model training may result in inaccurate predictions due to insufficient samples.

[0060] In this embodiment, the historical meteorological data in the historical meteorological dataset can come from observation data from meteorological stations in the target region. Each region has meteorological stations, and the history of each regional meteorological station is longer than the history of the local wind farms. In the training method provided in this application, a fitting relationship between meteorological data and wind speed is first established. Then, this fitting relationship is used to construct a large amount of simulated wind speed data as training samples to train the initial model. The resulting monthly-scale prediction model can predict the risk of wind turbine shutdown in future months with relatively high accuracy.

[0061] In this embodiment, no special limitation is made on how to obtain the fitting relationship between meteorological data and wind speed data. As an optional implementation, such as... Figure 2 As shown, the fitting of the historical meteorological dataset in the training sample set with the historical wind speed dataset includes:

[0062] In step S121, multiple regression models are used to fit the data in the training sample set to obtain multiple preliminary fitting results, wherein the preliminary fitting results are used to characterize the correspondence between meteorological data and wind speed data;

[0063] In step S122, multiple preliminary fitting results are integrated with equal weights to obtain an initial target fitting result, wherein the initial target fitting result is used to characterize the correspondence between meteorological data and wind speed data;

[0064] In step S123, the historical meteorological dataset is used as input to obtain the initial predicted wind speed dataset;

[0065] In step S124, the initial target fitting result is corrected by using the difference between the historical wind speed dataset and the initial predicted wind speed dataset to obtain the final target fitting result.

[0066] In the embodiments of this application, the fitting results of multiple regression models are fused using a weighted integration method, which can improve the stability and generalization ability of the final target fitting result when making predictions.

[0067] In this embodiment of the application, it is also proposed to use the difference between the historical wind speed dataset and the initial predicted wind speed dataset to correct the initial target fitting result, which can improve the accuracy of the simulation results.

[0068] Extreme tailwind speed events are a significant factor causing wind turbine damage. In this embodiment, to reduce systematic biases in tailwind speed prediction and enhance the simulation accuracy of extreme events, optionally, as follows: Figure 3As shown, the step of correcting the initial target fitting result by utilizing the difference between the historical wind speed dataset and the initial predicted wind speed dataset to obtain the final target fitting result includes:

[0069] In step S124a, the empirical distribution function of the initial predicted wind speed dataset and the empirical distribution function of the historical wind speed dataset are determined.

[0070] In step S124b, the empirical distribution function of the initial predicted wind speed dataset is matched with the empirical distribution function of the historical wind speed dataset. At the same cumulative probability position, the predicted wind speed data in the initial predicted wind speed dataset is replaced with the historical wind speed data in the historical wind speed dataset to obtain the final fitting result.

[0071] In step S124b, considering that extreme wind speed practices are easily diluted by mean dominance during the modeling process, a probability density matching method is introduced to significantly reduce the systematic bias in tail wind speed prediction and enhance the simulation accuracy of extreme events.

[0072] As an optional implementation, the plurality of regression models include at least two of the following regression models:

[0073] Random forest regression model, support vector machine regression model, gradient boosting tree regression model, and lightweight gradient boosting machine regression model.

[0074] Furthermore, in this embodiment of the application, in step S121, the above five regression models are used to fit the data in the training sample set, and in step S122, the fitting results of the above five regression models are weighted and fused.

[0075] As an optional implementation, the meteorological data in the historical meteorological dataset includes at least one of the following:

[0076] Ten atmospheric circulation indices, six atmospheric teleconnection indices, six ocean-atmosphere indices, and eleven sea ice and snow cover related indices;

[0077] The wind speed data in the historical wind speed dataset includes wind speed observation data of the wind farm's meteorological tower in the target area at a hub height of 100 meters.

[0078] For example, using the final fitting results, the daily maximum wind speed at a hub height of 100 meters for the target wind farm can be calculated using a sample of meteorological station data from January 1981 to May 2023.

[0079] As an optional implementation, the daily meteorological data in the historical meteorological dataset includes 33 climate data points: 10 atmospheric circulation indices, 6 atmospheric teleconnection indices, 6 ocean-atmosphere indices, and 11 sea ice and snow cover related indices. Optionally, the above data can be collected from the databases of the National Climate Center and the National Oceanic and Atmospheric Administration (NOAA).

[0080] It should be noted that, in the embodiments of this application, 33 climate driving factors from multiple physical spheres such as the atmosphere, ocean, sea ice and snow cover are integrated, which significantly enhances the model's ability to characterize and predict extreme wind events under different seasons and climate backgrounds.

[0081] The following example, using a 50MW wind farm in Province A, briefly explains how to obtain the final target fitting result:

[0082] (1) Hourly wind speed observation data from the wind measurement tower at a height of 100 meters in the target wind farm, covering the period from September 12, 2021 to May 31, 2023; the daily maximum wind speed was extracted as the prediction target of the model; (2) Daily meteorological observation data from three reference meteorological stations near the target wind farm during the same period, selected according to the "Technical Specification for Wind Energy Resource Measurement and Assessment of Wind Farm Projects" (NB / T 31147—2018), the reference stations must meet the standards of "similar terrain, consistent underlying surface, and good correlation with wind speed of the wind farm". Each meteorological station provides 11 basic meteorological elements, including: daily average air pressure, daily maximum air pressure, daily minimum air pressure, 10-minute maximum wind speed, 10-minute average wind speed, daily average temperature, daily maximum temperature, daily minimum temperature, daily relative humidity, daily precipitation and sunshine hours. A total of 33 basic meteorological prediction factors were formed by the three stations as input features of the model.

[0083] In the data preprocessing stage, the wind speed data and meteorological factor data were first time-aligned and divided into a training set (80%) and a test set (20%). Subsequently, the input features were standardized to have zero mean and unit standard deviation, aiming to improve the numerical stability and generalization performance of the model training.

[0084] To fully explore the nonlinear modeling capabilities of various models, this invention designs an integrated machine learning framework that integrates the following four regression models: (1) Random Forest Regression Model, (2) Support Vector Machine Regression Model, (3) Gradient Boosting Tree Regression Model, and (4) Lightweight Gradient Boosting Machine Regression Model. The above four models are fitted to the training set and used to predict test samples. Finally, the average value is calculated using an equal-weighted ensemble strategy to form preliminary simulation results. _pred_m.

[0085] To further improve the model's performance in the maximum wind speed range at the tail end, this invention introduces a probability density matching correction method (PDF Matching) based on the empirical distribution function. Its core idea is to adjust the model output... The empirical distribution function of _pred_m and the actual wind speed values ​​in the training set The empirical distribution function of the training data is matched, and at positions with the same cumulative probability, the predicted value is replaced with the corresponding quantile in the training value to obtain the corrected prediction sequence. _pred_m_1( Figure 5 ).exist Figure 5 As shown in the figure, the red solid line represents the observation data, the green dashed line represents the simulation results, and the blue dashed line represents the correction results. After correction, the correction results almost overlap with the observation data. It can be seen that this method can effectively eliminate systematic biases, spikes and tail offsets in the simulation results and improve the simulation capability of extreme events.

[0086] Based on the evaluation results, the Mean Absolute Percentage Error (MAPE) was used to measure the performance of the simulation results: uncorrected ensemble results. The MAPE of _pred_m is 14.93%, while the result after PDF matching correction is... The MAPE of _pred_m_1 (see Equation 1 below) decreased to 13.64%, demonstrating the significant advantage of the correction method in improving simulation accuracy (see Equation 1 below). Figure 6 ).

[0087] (1)

[0088] in, n The number of samples; For the first i One observation (true value) For the first i One predicted value.

[0089] After completing the training and verification of the integrated 100-meter hub height wind speed estimation model (i.e., the final target fitting result), this embodiment further utilizes the model to perform long-term simulation of the historical wind speed of the target wind farm. Specifically, based on the daily meteorological observation data (a total of 33 meteorological factors) from three reference meteorological stations around the target wind farm for 509 consecutive months (in this embodiment, from January 1981 to May 2023), these data are used as model inputs, and the maximum wind speed value at the 100-meter hub height of the target wind farm is generated daily through an inference process.

[0090] Ultimately, a simulated dataset of daily maximum wind speeds at a hub height of 100 meters was constructed, covering the period from January 1981 to May 2023. This dataset effectively compensates for the problems of short coverage time and unstable frequency of the original wind farm observation data, and provides continuous, complete and high-quality data support for long-term statistical analysis and risk modeling of extreme wind weather events.

[0091] The 509 months of simulated data serve as the core input for the monthly maximum wind weather probability prediction model, significantly improving the sample coverage and statistical significance of medium- and long-term risk identification, and laying a solid data foundation for subsequent probability model training, grade mapping, and prediction system deployment.

[0092] Optionally, the meteorological prediction factor matrix constructed by this method (i.e., the historical meteorological dataset) contains 33 monthly-scale variables, covering the period from January 1981 to May 2023. These variables include: 10 atmospheric circulation indices from the 88 circulation indices published by the National Climate Center, including: Northern Hemisphere subtropical high area index, Western Pacific subtropical high intensity index, Western Pacific subtropical high ridge position index, Western Pacific subtropical high westward extension ridge point index, Eurasian zonal circulation index, Eurasian meridional circulation index, East Asian trough position index, East Asian trough intensity index, Tibetan Plateau-1 index, and Tibetan Plateau-2 index; and 6 atmospheric teleconnection indices published by the NOAA Climate Prediction Center, including: North Atlantic Oscillation Index, East Atlantic Type Index, Western Pacific Type Index, Pacific / North American Type Index, East Atlantic / West Russian Type Index, and Arctic Oscillation Index. The NOAA Physical Sciences Laboratory publishes six ocean and atmospheric indices, including: Nino1+2, Nino3.4, Nino3, Nino4, the Pacific Decadal Oscillation Index, and the Southern Oscillation Index; the NOAA National Center for Environmental Information publishes 11 sea ice, snow cover, and sea surface temperature indices, including: Barents Sea Ice Index, Beaufort Sea Ice Index, Bering Sea Ice Index, Chukchi Sea Ice Index, East Siberian Sea Ice Index, Greenland Sea Ice Index, Kara Sea Ice Index, Sea of ​​Okhotsk Ice Index, Eurasian Snow Cover Index, Northern Hemisphere Snow Cover Index, and the Atlantic Decadal Oscillation Index.

[0093] The target variable is the number of wind turbine cut-out risk days per month from January 1981 to May 2023. All data are time-aligned and merged into a unified time-series data structure. The input features are climate factors of 3 consecutive months × 33 dimensions, and the output is the number of cut-out risk days for the following month.

[0094] In this embodiment of the application, the specific structure of the initial model is not specifically limited. For example, the initial model includes a cascaded temporal convolutional network, a gated recurrent unit, an attention mechanism layer, and a Softmax classifier.

[0095] The input data of the temporal convolutional network is the wind speed observation data of the previous N consecutive months, and the output data of the attention mechanism layer is the wind speed prediction data for the next month.

[0096] The Softmax classifier is used to output the risk level of wind turbine shutdown based on wind speed forecast data for the next month.

[0097] A temporal convolutional network structure is introduced to establish short-term and medium-term climate dependence, a gated recurrent unit recursive structure is introduced to capture long-term climate trends, and an attention mechanism layer is used to extract key monthly information in a weighted manner, thereby comprehensively improving the model's expressive power and interpretability.

[0098] Specifically, the Temporal Convolutional Network (TCN) module consists of multiple layers of one-dimensional causal convolutions, used to efficiently capture short- to medium-term dependency features in time series. The causal structure ensures that the output at each time step depends only on the input at the current and previous time steps, effectively avoiding the leakage of future information and ensuring the temporal consistency of the prediction task. By setting different dilation rates, this module can expand its receptive field to 1 to 6 months.

[0099] Gated Recurrent Unit (GRU): The feature sequences extracted by the convolutional module are fed into the GRU to further establish the dependencies between time series. In this embodiment, return_sequences=True is set to retain the hidden state at each time step, which is then used as input to the subsequent attention mechanism layer. The GRU effectively establishes the cumulative influence and lag effect of climate factors, enhancing the model's ability to express complex temporal features.

[0100] To enhance the model's ability to focus on key time segments, this application introduces an attention mechanism layer on top of the GRU output. This mechanism not only "remembers the complete history" but also "focuses on key segments," significantly improving the ability to identify typical anomalous climate events (such as the persistent Western Pacific Subtropical High and the strong Arctic Oscillation) and enhancing the model's interpretability.

[0101] Softmax Classification Layer: The context vector aggregated by the attention mechanism layer is used as the overall semantic feature and input to a fully connected network layer. The output is a Softmax probability vector of length 4, corresponding to four wind power outage risk levels (0-3).

[0102] Optionally, the network structure of the initial model is shown in Table 1 below:

[0103] Table 1

[0104]

[0105] In this embodiment, the format of the classifier output data is not specifically limited. Optionally, the classifier is used to output the probability distribution of various risk levels, wherein the more days with maximum monthly winds, the higher the risk level.

[0106] As an optional implementation method, the risk level is 0 when there are 0 days with maximum winds in a month; the risk level is 1 when there are 1 to 2 days with maximum winds in a month; the risk level is 2 when there are 3 to 4 days with maximum winds in a month; and the risk level is 3 when there are 5 or more days with maximum winds in a month.

[0107] The aforementioned risk level mapping mechanism converts continuous variables output by the model into risk level labels that are easy to interpret in practical business operations through conditional judgment and discrete classification. The mapping rules have good scalability and flexibility, and can be adjusted and customized according to the operating standards, maintenance strategies, or insurance underwriting requirements of different wind farms. For example, more suitable risk classification thresholds can be set for specific geographical areas or turbine types.

[0108] Optionally, to improve the interpretability of the model and the intuitiveness of the results, this application proposes a visualization mechanism. Typical samples are selected from the test set, and the risk level sequences predicted by the model and the actual observed risk level sequences are plotted and displayed in the form of a "ladder diagram" (e.g., Figure 7 This intuitively presents the model's predictive trends, level changes, and error distribution over continuous time series, enhancing business understanding and scenario adaptability. Figure 7 In the diagram, the solid blue line represents the actual risk level, while the dashed orange line represents the predicted risk level.

[0109] To comprehensively evaluate the model's performance in classification tasks, particularly its ability to identify medium- to high-level extreme wind events, the embodiments of this application employ the following three core evaluation metrics:

[0110] Overall classification accuracy: measures the proportion of all samples in which the predicted level is completely consistent with the true level. It is a basic indicator for evaluating the overall recognition ability of the model.

[0111] High-risk identification capability (Recall@2+): The proportion of samples with a true risk level of 2 or 3 that also predict a risk level of 2 or 3, used to measure the effectiveness of the model in early warning of medium- and high-risk events;

[0112] Quadratic Weighted Kappa (QWK): Considers the order information between categories and measures the consistency between the predicted results and the true ranking in the overall sequence. It is often used in multi-level classification tasks to evaluate the ranking reliability of the model.

[0113] The results of the three core evaluation indicators are shown in Table 2 below: QWK reached 0.754, indicating that the model's predicted ranking is very close to reality and is suitable for graded risk tasks; although there is still room for improvement in high-risk recall (Recall@2+), it has been increased to 50%, achieving a good balance between accuracy and risk identification ability.

[0114] Table 2

[0115]

[0116] As an alternative implementation, a wind speed threshold can be set, and winds exceeding the wind speed threshold can be defined as extreme winds.

[0117] Considering the extreme imbalance of the samples (with level 0 and level 1 dominating), this embodiment introduces a category-weighted Focal Loss function to improve the identification ability of level 2 and level 3 (medium-high risk) events. Specifically, the loss function of the initial model satisfies the following formula:

[0118] ;

[0119] in, t Identifier for the target category;

[0120] This represents the model's predicted probability of the true class (i.e., the value of the true label position in the Softmax output).

[0121] This is a category weight term used to adjust the relative importance of samples from different categories;

[0122] As a focusing factor, it controls the penalty intensity for easily classified samples;

[0123] This is the standard cross-entropy term.

[0124] In this invention, to enhance the model's ability to identify high-risk levels, the category weight vector is explicitly set as follows: .

[0125] By introducing the aforementioned Focal Loss during the training of the deep neural network, this invention significantly improves the model's recall capability at high risk levels (2 and 3). This mechanism is one of the key innovations of this invention in extreme event classification tasks.

[0126] Step S140 is briefly described below. The model training in step S140 employs a mini-batch gradient descent approach, with training samples used to update network parameters batch by batch via backpropagation. The optimizer is Adam, and the learning rate is adaptively adjusted to improve convergence speed and numerical stability. To prevent overfitting, an early stopping mechanism is introduced: training is terminated early if the validation set loss shows no significant improvement after 16 consecutive training rounds. Throughout the training process, the system dynamically monitors training and validation errors and records the model's performance on key evaluation metrics such as Accuracy, Recall@2+, and QWK after each training round, ensuring the model has business deployment value in terms of both overall accuracy and high-risk identification.

[0127] In step S140, the training samples are divided into a training set (70%), a validation set (20%), and a test set (10%) according to a random layer strategy. The training process is set as follows:

[0128] Hyperparameter settings: epochs = 30, batch_size = 16, optimizer is Adam;

[0129] Online evaluation metrics include Accuracy, Recall@2+, F1-score, and QWK (all implemented using custom callback functions).

[0130] Early stopping mechanism: If the validation set metrics do not improve for 8 consecutive epochs, training is terminated to avoid overfitting.

[0131] In this embodiment of the application, no special limitation is made on how to use the output values ​​of the target simulation results to form the initial predicted wind speed dataset. Optionally, the steps to form the initial predicted wind speed dataset may include:

[0132] Based on the data time span, a continuous and complete Gregorian calendar sequence is generated, and wind speed records are aligned day by day. If there are missing days, null values ​​can be used to place them to ensure the continuity of the time series and provide a basis for subsequent sliding window construction and statistical analysis.

[0133] Clean the initial simulation data and unify the date field in the initial data to a standard time format (e.g., YYYY-MM-DD).

[0134] The fields of wind speed observation data in the initial simulation data are standardized to floating point type to remove non-numerical, missing or physically unreasonable abnormal data;

[0135] Set daily cut-out thresholds for the cleaned data. Specifically, refer to the technical parameter manual of the deployed wind turbine and combine it with the actual measured wind conditions in the area to set the critical cut-out wind speed value for the turbine. For example, the cut-out wind speed for some mainstream models is set to 20 m / s.

[0136] The daily wind speed is compared with a set threshold. If the wind speed is greater than the threshold, the day is marked as "extreme wind day" (True); otherwise, it is marked as "non-extreme wind day" (False). This process forms a binary identifier sequence, which provides a basis for monthly risk statistics (this step is the daily threshold determination step).

[0137] Group the data by year and month to generate a time hierarchy structure indexed by "year-month";

[0138] Valid recorded days in the current month: The number of days in the month for which valid wind speed observation data was actually available;

[0139] Maximum wind days: The number of days in the month when wind speeds exceeded the cut-out threshold;

[0140] Maximum wind frequency: The number of days with maximum winds divided by the number of valid observation days in the month reflects the relative density of the downtime risk in that month.

[0141] In this embodiment, to fully capture the time-dependent characteristics of the number of wind turbine downtime risk days, a fixed-length sliding window mechanism is proposed to transform historical climate factor sequences into an input format recognizable by a deep learning model. Specifically, each training sample consists of meteorological prediction factors for several consecutive months, and the output is the number of maximum wind downtime days in the month following the end of the window.

[0142] In this implementation, the sliding window length is set to 3 months, and the prediction step size is 1 month, ensuring that each set of inputs can cover short- to medium-term climate impact characteristics. The sliding window process proceeds step by step on the complete data sequence in chronological order, constructing multiple sets of input-output pairs to form a well-structured and sufficiently large set of supervised learning samples. All samples maintain consistent feature dimensions, facilitating subsequent batch training and parallel optimization of the neural network model.

[0143] After constructing the sliding window samples, the system randomly shuffles the sample data and divides it into training, validation, and test sets in a ratio of 70%:20%:10%, which are used for model parameter fitting, hyperparameter tuning, and final performance evaluation, respectively, to ensure that the model has good generalization ability in the development, validation, and deployment stages.

[0144] Considering the significant differences in the numerical scale and distribution of various meteorological factors, directly inputting them into the model might affect the stability and accuracy of the training process. Therefore, the mean and standard deviation of all climate factors are calculated on the training set, and the input features in the training, validation, and test sets are standardized based on these statistical parameters, ensuring that the mean of each feature is 0 and the standard deviation is 1. This preprocessing not only improves the model's convergence efficiency but also prevents a factor from obtaining an excessively high gradient during backpropagation due to a large numerical value, thus avoiding an imbalance in model weight updates.

[0145] In this embodiment of the application, the classification criteria for wind turbine shutdown risk are shown in Table 3.

[0146] Table 3

[0147]

[0148] By dividing the data into the above categories, the monthly maximum wind days are converted into a risk level sequence, and a standardized and discrete output result is constructed, which facilitates rapid interpretation and response in scenarios such as wind farm operation and maintenance, scheduling, and insurance.

[0149] In this application, the output of the prediction model may include the following key fields:

[0150] year: year;

[0151] month: month;

[0152] strong_wind_days: Number of days with maximum wind speed in a month;

[0153] risk_level: The corresponding risk level.

[0154] In this embodiment, the output results can be exported as a CSV file, supporting seamless integration with subsequent predictive modeling processes, risk warning systems, and visualization platforms.

[0155] As a second aspect of the present invention, a monthly-scale prediction method for wind turbine shutdown risk level is provided, wherein the monthly-scale prediction method includes:

[0156] In step S210, wind speed observation data from the previous N consecutive months are input into the monthly-scale prediction model, wherein the monthly-scale prediction model is a model trained using the above training method, and the wind speed prediction data for the next month is used to output the risk level of wind turbine shutdown.

[0157] This application proposes a multi-module deep neural network integrating TCN, GRU, attention mechanism layer, and Softmax classifier to address the extremely imbalanced sample distribution and missed reporting of high-risk levels (2 and 3) in wind farm turbine shutdown risk level prediction under extreme wind conditions. This method is innovative in its network structure, loss function, and evaluation index system, significantly improving high-risk identification capabilities and providing more reliable early warning support for wind power operation and maintenance scheduling.

[0158] Considering that high-risk samples account for a very small proportion, this invention introduces class-weighted Focal Loss as the optimization objective. This function, by adjusting the focusing factor (γ) and class weight (α), assigns higher loss weights to hard-to-classify samples (i.e., samples with large differences between the true label and the predicted probability) during training, thereby effectively suppressing the main class dominance effect and improving the model's ability to identify high-risk samples at the tail end.

[0159] In summary, the deep neural network structure proposed in this invention integrates the time-awareness capability of TCN, the temporal dependency modeling capability of GRU, the key feature aggregation capability of the attention mechanism, and the sample imbalance adaptation capability of Focal Loss to construct a prediction model for the maximum wind outage risk level with strong generalization, interpretability, and business sensitivity. It can meet the forward-looking early warning needs of high-risk meteorological events in multiple scenarios such as wind farm operation and maintenance, power dispatching, and risk insurance.

[0160] As a third aspect of the embodiments of this application, an electronic device is provided, wherein, as Figure 8 As shown, the electronic device includes:

[0161] One or more processors 101;

[0162] The memory 102 stores one or more computer programs that, when executed by the one or more processors 101, cause the one or more processors 101 to implement the training method provided in the first aspect of the embodiments of this application and / or the prediction method provided in the second aspect.

[0163] The electronic device may also include one or more I / O interfaces 103 connected between the processor 101 and the memory 102, configured to enable information interaction between the processor 101 and the memory 102.

[0164] The processor 101 is a device with data processing capabilities, including but not limited to a central processing unit (CPU); the memory 102 is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read-write interface) is connected between the processor and the memory, enabling information exchange between the processor and the memory, including but not limited to a data bus (Bus).

[0165] In some embodiments, the processor 101, memory 102, and I / O interface 103 are interconnected via bus 104, and thus connected to other components of the computing device.

[0166] like Figure 9 As shown, embodiments of this application also provide a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the training method provided in the first aspect and / or the prediction method provided in the second aspect of embodiments of this application.

[0167] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. Accordingly, the computer program can be stored in a non-volatile computer-readable storage medium, and when executed, the computer program can implement the methods of any of the above embodiments. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0168] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Those skilled in the art should understand that this application includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments above. Any modifications that do not depart from the functional and structural principles of this application will be included within the scope of the claims.

Claims

1. A training method for a monthly-scale prediction model of wind turbine shutdown risk levels, characterized in that, The training method includes: A training sample set is constructed using the historical meteorological dataset of the target area within the first set historical time period and the historical wind speed dataset of the target area within the first set historical time period. The historical meteorological dataset and the historical wind speed dataset in the training sample set are fitted to obtain the final target fitting result, which is used to characterize the correspondence between meteorological data and wind speed data. The maximum wind speed simulation dataset for the target area within the second set historical time period is generated based on the historical meteorological dataset of the target area within the second set historical time period and the final target fitting result, wherein the time span of the second set historical time period is greater than the time span of the first set historical time period. The initial model was trained using the maximum wind speed simulation dataset as training samples to obtain the monthly-scale prediction model. The fitting of the historical meteorological dataset in the training sample set with the historical wind speed dataset includes: Multiple regression models are used to fit the data in the training sample set to obtain multiple preliminary fitting results, wherein the preliminary fitting results are used to characterize the correspondence between meteorological data and wind speed data; Multiple preliminary fitting results are integrated with equal weights to obtain an initial target fitting result, wherein the initial target fitting result is used to characterize the correspondence between meteorological data and wind speed data; Using the historical meteorological dataset as input, an initial predicted wind speed dataset is obtained; The initial target fitting result is corrected by using the difference between the historical wind speed dataset and the initial predicted wind speed dataset to obtain the final target fitting result; The step of correcting the initial target fitting result by utilizing the difference between the historical wind speed dataset and the initial predicted wind speed dataset to obtain the final target fitting result includes: Determine the empirical distribution function of the initial predicted wind speed dataset and the empirical distribution function of the historical wind speed dataset; The empirical distribution function of the initial predicted wind speed dataset is matched with the empirical distribution function of the historical wind speed dataset. At the same cumulative probability position, the predicted wind speed data in the initial predicted wind speed dataset is replaced with the historical wind speed data in the historical wind speed dataset to obtain the final fitting result.

2. The training method according to claim 1, characterized in that, The plurality of regression models include at least two of the following regression models: Random forest regression model, support vector machine regression model, gradient boosting tree regression model, and lightweight gradient boosting machine regression model.

3. The training method according to claim 1 or 2, characterized in that, The meteorological data in the historical meteorological dataset includes at least one of the following: Ten atmospheric circulation indices, six atmospheric teleconnection indices, six ocean-atmosphere indices, and eleven sea ice and snow cover related indices; The wind speed data in the historical wind speed dataset includes wind speed observation data of the wind farm's meteorological tower in the target area at a hub height of 100 meters.

4. The training method according to claim 1 or 2, characterized in that, The initial model includes a cascaded temporal convolutional network, a gated recurrent unit layer, an attention mechanism layer, and a Softmax classifier; The input data of the temporal convolutional network is the wind speed observation data of the previous N consecutive months, and the output data of the attention mechanism layer is the wind speed prediction data for the next month. The Softmax classifier is used to output the risk level of wind turbine shutdown based on wind speed forecast data for the next month.

5. The training method according to claim 4, characterized in that, The Softmax classifier is used to output the probability distribution of various risk levels, where the more days with maximum monthly winds, the higher the risk level.

6. The training method according to claim 4, characterized in that, The loss function of the initial model satisfies the following formula: ； in, p t Let be the model's predicted probability for the true class, and ; For category weight items; As a focusing factor; logp t This is the standard cross-entropy term.

7. A monthly-scale prediction method for wind turbine downtime risk level, characterized in that, The monthly-scale prediction method includes: The wind speed observation data of the previous N consecutive months are input into the monthly scale prediction model to obtain the wind speed prediction data for the next month and output the risk level of wind turbine shutdown, where N is a positive integer and N>1. The monthly scale prediction model is a model trained using the training method described in any one of claims 1 to 6.

8. An electronic device, characterized in that, The electronic device includes: One or more processors; A memory having stored thereon one or more computer programs that, when executed by the one or more processors, cause the one or more processors to implement the method according to any one of claims 1 to 7.

Images