Method for calculating extreme wind speed of deep and far sea wind power plant based on artificial intelligence
Patent Information
- Application Number
- CN202511416911.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-30
- Publication Date
- 2026-01-02
AI Technical Summary
Existing technologies lack long-term observation data in deep-sea areas, making it difficult to accurately calculate extreme wind speeds, especially in sea areas unaffected by typhoons, and existing methods lack universality.
Using an artificial intelligence-based approach, combining meteorological reanalysis data and short-term measured data, a hybrid model combining deep neural networks and extreme value theory is constructed to calculate extreme wind speeds in deep-sea wind farms, including data matching, feature extraction, extreme event identification, and the application of probability distribution models.
It can accurately calculate the extreme wind speed of deep-sea wind farms without relying on long-term measured data, providing key wind resistance parameters for wind turbine selection and safe operation, and improving the accuracy and physical interpretation of the calculation.
Smart Images
Figure CN121256261A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of extreme wind speed of deep-sea wind farm, and particularly relates to a method for calculating extreme wind speed of deep-sea wind farm based on artificial intelligence. BACKGROUND
[0002] In recent years, offshore wind power has gradually moved towards deep-sea, but development faces many technical challenges, among which extreme wind speed calculation is an urgent problem to be solved.
[0003] Currently, the extreme wind speed calculation of offshore wind farm mainly adopts extreme probability distribution method (such as Gumbel distribution) and typhoon model numerical simulation method. The former calculates the extreme value by fitting long-term measured historical data, but there are few long-term observation points in deep-sea area, and there are also few island weather stations or offshore buoys nearby. Short-term observation data cannot meet the calculation conditions. The latter substitutes the historical typhoon best path data near the wind farm into the parameterized typhoon model for numerical simulation, so as to estimate the maximum wind speed of typhoon. However, this method is only suitable for high-impact sea areas of typhoon, and is based on the premise that the maximum wind speed in the area is caused by typhoon, which lacks universality.
[0004] Therefore, how to provide a method for calculating the extreme wind speed of deep-sea wind farm without relying on long-term measured data, even in sea areas not affected by typhoon, is a technical problem to be solved at present. SUMMARY
[0005] The purpose of the application is to provide a method for calculating the extreme wind speed of deep-sea wind farm based on artificial intelligence, in order to solve the above problems existing in the prior art.
[0006] In order to achieve the above purpose, the application adopts the following technical scheme: The application provides a method for calculating the extreme wind speed of deep-sea wind farm based on artificial intelligence, comprising: obtaining historical data and short-term measured data of meteorological reanalysis data, wherein the short-term measured data is the meteorological data of a preset observation point within a preset time period; spatially matching the historical data and short-term measured data of meteorological reanalysis data to establish a database, wherein the spatial matching ensures that the historical data and measured data of meteorological reanalysis data point to the same time point and the same geographical location; extracting each data related to wind speed in the database to construct a wind speed time sequence, and calculating wind speed statistical characteristic values using a sliding time window for the wind speed time sequence, wherein the wind speed statistical characteristic values at least include the average value, standard deviation, variance of wind speed in the window, the maximum and minimum values in the window, the maximum and minimum rates of wind speed in the window; An extreme event identification model is constructed to generate dynamic probability features of extreme events for each historical moment. The dynamic probability features and the wind speed statistical feature values are incorporated into the database to form a fusion database. The extreme event identification model is used to identify which extreme weather system is currently in control. Based on the data in the fusion database, a hybrid model combining deep neural networks and extreme value theory is constructed. This hybrid model is used to generate long-term maximum wind speed sequences for the locations of deep-sea wind farms. Using the long-term maximum wind speed sequence, the extreme wind speed at the hub height of the wind turbine in the deep-sea wind farm was calculated by employing a probability distribution model and a wind power exponent equation.
[0007] Optionally, using the long-term maximum wind speed sequence, the extreme wind speed at the turbine hub height of the deep-sea wind farm is calculated using a probability distribution model and a wind power exponential equation. The probability distribution model adopts the Gumbel extreme value type I, and the extreme wind speed at the turbine hub height is calculated using the following formula: middle, For reference height, The height of the wind turbine hub. For height is To measure the wind shear index, For height is The extreme wind speed at that time.
[0008] Optionally, the architecture of the hybrid model combining deep neural networks and extremum theory includes: Input layer: The data in the fusion database; Feature extraction backbone: A 4-layer deep fully connected network with 512 neurons in each layer; Attention mechanism: A multi-head self-attention layer is introduced after the third layer of the fully connected network. The attention mechanism enables the model to automatically weight the key features of extreme events. The primary regression head in the dual-output head outputs the predicted average wind speed under current meteorological conditions. The GPD parameter header in the dual-output header outputs scale and shape parameters. This GPD parameter header defines the probability distribution shape of the extreme wind speed component in the prediction residual.
[0009] Optionally, after constructing a hybrid model combining deep neural networks and extreme value theory based on the data in the fusion database, the hybrid model combining deep neural networks and extreme value theory is further refined using long-term measured data from nearshore wind farms, including: A nearshore dataset is constructed based on long-term measured data of nearshore wind farms and historical data of corresponding meteorological reanalysis data. The model weights are obtained by training the data in the nearshore dataset using a hybrid model that combines the deep neural network with the extreme value theory. The model weights are then used as initial values and loaded back into the hybrid model. Using training data from the deep sea, the hybrid model was fine-tuned with a low learning rate after the initial values were loaded back into the hybrid model, in order to optimize the model that combines deep neural networks with extremum theory.
[0010] Optionally, meteorological feature data related to wind speed can be extracted from historical data in the database, and short-term measured wind speed data can be extracted from measured data. The meteorological feature data and short-term measured wind speed data can be used to construct a wind speed time series.
[0011] Optionally, before spatially matching the historical and measured data of the meteorological reanalysis data to establish a database, the method further includes: The historical and measured data of the meteorological reanalysis data are subjected to quality control and consistency checks to remove abnormal data. The quality control involves removing the historical and measured data that exceed the preset value range.
[0012] Optionally, the historical and measured data of the meteorological reanalysis data are spatially matched to establish a database, including: Time matching is performed on historical data and measured data of meteorological reanalysis data to ensure that the timestamps of historical data and measured data of meteorological reanalysis data are consistent; Bilinear interpolation is used to spatially match historical and measured data of meteorological reanalysis data, so that the historical and measured data of meteorological reanalysis data have the same geographical location. Historical and measured meteorological reanalysis data, after time and spatial matching, are merged to establish a database. The database includes at least timestamp sequences, latitude, longitude, wind speed, relevant meteorological elements, and time.
[0013] Optionally, the method further includes constructing a set of extreme event fingerprint features: They are identified and labeled according to the definition of extreme events, namely typhoon events / cold wave events; Extract multi-dimensional meteorological physical quantities from each labeled cold wave / typhoon event within a preset time period before and after the event to construct a fingerprint feature set of extreme events. The fingerprint feature set includes at least the event type and the meteorological physical quantity features corresponding to the event type.
[0014] Optionally, the method further includes: Each extreme typhoon event was designated as a positive typhoon sample, non-extreme typhoon events as negative typhoon samples, extreme cold wave events as positive cold wave samples, and non-extreme cold wave events as negative cold wave samples. A sample library is created based on each positive and negative sample of a typhoon, a positive and negative sample of a cold wave, and the fingerprint features corresponding to each sample. The model is trained based on the sample library to construct an extreme event recognition model; The extreme event identification model yields the feature patterns of extreme events, which are used to identify and distinguish different types of extreme weather.
[0015] Optionally, the measured data is meteorological data measured by a floating lidar.
[0016] Beneficial effects: This application provides a method for calculating extreme wind speeds that integrates artificial intelligence and reanalysis data. This method can calculate extreme wind speeds in deep-sea wind farms, even in areas unaffected by typhoons, without relying on long-term measured data. It can provide key wind-resistance parameters for wind turbine selection, structural design, and safe operation in deep-sea wind farms. Attached Figure Description
[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application. In the drawings: Figure 1 A flowchart illustrating a method for calculating extreme wind speeds in deep-sea wind farms based on artificial intelligence, provided as an embodiment of this application; Detailed Implementation To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.
[0018] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.
[0019] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.
[0020] Example 1: like Figure 1 The diagram shown is a flowchart illustrating a method for calculating extreme wind speeds in deep-sea wind farms based on artificial intelligence, as proposed in an embodiment of the present invention, including: S1, acquire historical data and measured data of meteorological reanalysis data, wherein the measured data is meteorological data of preset observation points within a preset time period.
[0021] Specifically, historical meteorological reanalysis data can preferentially select ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF). ERA5 data provides hourly estimates of atmospheric, land, and oceanic climate variables, and can be supplemented by CFSR / CFSv2 reanalysis data from the National Center for Environmental Prediction (NCEP) in the United States. Meteorological reanalysis data are not measured data. The measured data in this application are short-term measured data. As one embodiment of this application, the limited measured data are wind speed data, etc., observed by floating lidar at 1-3 year prediction observation points.
[0022] S2, Spatially match the historical data and measured data of the meteorological reanalysis data to establish a database.
[0023] Specifically, "spatiotemporal matching" means ensuring that every set of historical data from meteorological reanalysis and measured data from floating lidar points to the same point in time and the same geographical location.
[0024] To ensure accurate acquisition of extreme wind speeds, in a preferred embodiment of this solution, historical and measured data from the meteorological reanalysis data are spatially matched to establish a database, including: Time matching is performed on historical data and measured data of meteorological reanalysis data to ensure that the timestamps of historical data and measured data of meteorological reanalysis data are consistent; Bilinear interpolation is used to spatially match historical and measured data of meteorological reanalysis data, so that the historical and measured data of meteorological reanalysis data have the same geographical location. Historical and measured meteorological reanalysis data, after time and space matching, are merged to establish a database. The database includes at least timestamp sequences, latitude, longitude, wind speed, and related meteorological elements and times.
[0025] Specifically, I. Spatiotemporal Matching Technology Solution 1. Time matching Objective: To ensure that the timestamps of historical and measured meteorological reanalysis data are completely consistent; Methods: Time resolution matching: ERA5 data uses a fixed time resolution (e.g., once per hour), so the LiDAR data needs to be adjusted to the same time step through interpolation or resampling. UTC time unification: Unify the timestamps of the two data sources to Universal Time (UTC) to avoid time offset caused by time zone differences.
[0026] 2. Spatial matching Objective: To ensure that historical and measured data in meteorological reanalysis are geographically close.
[0027] Method: Bilinear interpolation method: ERA5 data is interpolated to the observation point location using the bilinear interpolation method. The observation point location is the location observed by the floating laser wind radar. The reanalysis data of the observation point obtained by interpolation is extracted and matched with the measured data.
[0028] The reanalysis data of the observation point obtained by interpolation was achieved using the following method: ERA5 uses a 0.25°×0.25° latitude and longitude grid (approximately 28km×28km). Based on the latitude and longitude of the observation point, the ERA5 grid cell in which it is located is found, and data from four neighboring grid points (such as temperature, air pressure, etc.) are extracted. Based on the relative position of the observation point and the four grid points (such as longitude difference, latitude difference), the weights are calculated, and the weighted sum is used to obtain the interpolation result of the observation point.
[0029] 3. Establish a database By fusing historical and measured data from meteorological reanalysis data that have undergone time and spatial matching, a database can be established. Each row in the table below represents a matched spatiotemporal sample, and the columns contain various variables. As an embodiment of this application, before establishing the database, the historical and measured data from the meteorological reanalysis data undergo quality control and consistency checks to remove abnormal data. The quality control involves removing historical and measured data that exceed a preset numerical range. Quality control mainly focuses on verifying the rationality of the measured data, and can be carried out according to the specific requirements of the "Technical Specification for Measurement and Evaluation of Wind Energy Resources in Wind Farms" to ensure the reliability and authenticity of the data. Quality control includes: Range check: removing values that are significantly outside the reasonable range (such as wind speed <0 m / s or >90 m / s). Consistency check: verifying the rationality of the measured data according to the "Technical Specification for Measurement and Evaluation of Wind Energy Resources in Wind Farms".
[0030] Example database structure: S3, extract the wind speed-related data from the database to construct a wind speed time series, and use a sliding time window to calculate the wind speed statistical characteristic value of the wind speed time series.
[0031] Specifically, wind speed and related meteorological characteristics (pressure gradient, temperature gradient, humidity, etc.) are extracted from historical data of meteorological reanalysis data in the database. Short-term measured wind speed data are extracted from the measured data. The meteorological characteristic data and the short-term measured wind speed data are combined to construct a wind speed time series. A sliding time window is used to calculate the wind speed statistical characteristic values (such as mean, variance, standard deviation, etc.) of the wind speed time series. In one embodiment of this application, the historical data in the meteorological reanalysis data is the area near the measured data point (the area of the proposed wind farm), such as extending 50km outward.
[0032] 1. The detailed technical solution for calculating the statistical characteristic values of wind speed is as follows: For wind speed time series data in the database, a sliding time window method is used to calculate statistical characteristics. The window time scale is typically defined based on the characteristics of extreme events (e.g., 3 days). For each time point `t_i` in the time series, all data from the preceding 72 hours (including `t_i`) are taken, and the statistics for both the measured data and the ERA5 data are calculated. The window is then slid to the next point, and the calculation is repeated. The calculated characteristic values are shown below: Central trend: (1): The average wind speed within the window, reflecting the average energy level.
[0033] Degree of dispersion: (2): Standard deviation directly reflects the fluctuation of wind speed and the intensity of turbulence, and is a key feature for calculating extreme values.
[0034] (3): Variance, square of standard deviation.
[0035] (4): The maximum value within the window, directly capturing recent extreme events.
[0036] (5): The minimum value within the window.
[0037] (6): (maximum value - minimum value), reflects the total variation.
[0038] (7): The maximum rate of rise and fall of wind speed within the window (e.g., the maximum value of (u_{t+1} - u_t) / Δt). Finally, the calculated wind speed statistical characteristic values are entered into the database above.
[0039] S4. Construct an extreme event identification model, generating dynamic probability features for each historical moment. Incorporate these dynamic probability features and the wind speed statistical feature values into the database to form a fusion database. The extreme event identification model is used to identify the current extreme weather system in control. The fusion database is the final data set after standardization, where all input features (including wind speed statistical feature values and dynamic probability features) are combined into a unified large matrix.
[0040] In order to obtain the characteristic patterns of extreme events, in a preferred embodiment of this solution, the method further includes constructing a set of fingerprint features for extreme events: They are identified and labeled according to the definition of extreme events, namely typhoon events / cold wave events; Extract multi-dimensional meteorological physical quantities from each labeled cold wave / typhoon event within a preset time period before and after the event to construct a fingerprint feature set of extreme events. The fingerprint feature set includes at least the event type and the meteorological physical quantity features corresponding to the event type.
[0041] Specifically, ① Define extreme event labels (typhoon time, cold wave event). 1. Typhoon event identification and labeling: Data source: The CMA (China Meteorological Administration) Tropical Cyclone Best Track Dataset was used. It includes historical typhoon center locations, central pressures, maximum wind speeds, etc.
[0042] Labeling algorithm: Calculate the distance between the location of the wind farm and the path of the nearby typhoon at each time point (e.g., 6-hour interval).
[0043] Define an influence radius R (e.g., 500 km). If the distance d between the typhoon center and the wind farm is less than or equal to R, then the typhoon is considered to be affecting the wind farm at that moment.
[0044] Definition: All periods when the wind farm is affected are labeled as 1 (typhoon extreme event, positive sample), and all other periods are labeled as 0 (typhoon non-extreme event, negative sample). 2. Cold wave event identification and labeling: Cold waves and strong winds are usually associated with strong cold high pressure systems and dramatic pressure gradients. Definition: 24-hour pressure change: ΔP_24h > 10 hPa (a sharp increase in air pressure, indicating the passage of a cold high-pressure front) Sudden temperature drop: ΔT_24h < -10 °C (rapid temperature decrease) Strong wind conditions: Wind speed > 20 m / s (meeting strong wind standard) Labeling algorithm: Calculate the 24-hour air pressure and temperature changes at each point in the reanalysis data. If all three conditions are met, the time is labeled as an extreme cold wave event, and the other times are labeled as non-extreme cold wave events.
[0045] ② Feature extraction—Constructing the “physical fingerprint” of an event For each labeled event (and a time window before and after it, e.g., ±72 hours), multi-dimensional meteorological physical quantities are extracted from each labeled cold wave and / or typhoon event to construct a fingerprint feature set for the extreme event. These quantities collectively constitute the unique "fingerprint" of the event. The fingerprint feature set includes at least the event type and the data features corresponding to the event type. The table below is an example of a fingerprint feature set: To further obtain the characteristic patterns of extreme events, in a preferred embodiment of this solution, the method further includes: Each extreme typhoon event was designated as a positive typhoon sample, non-extreme typhoon events as negative typhoon samples, extreme cold wave events as positive cold wave samples, and non-extreme cold wave events as negative cold wave samples. A sample library is created based on each positive and negative sample of a typhoon, a positive and negative sample of a cold wave, and the fingerprint features corresponding to each sample. The model is trained based on the sample library to construct an extreme event recognition model; The extreme event identification model yields the feature patterns of extreme events, which are used to identify and distinguish different types of extreme weather.
[0046] The feature patterns of extreme events are used to identify and differentiate different types of extreme weather. A "feature pattern" is not a single indicator, but a multi-dimensional feature vector. Each dimension of this vector represents a meteorological parameter extracted from historical data in meteorological reanalysis that characterizes the physical nature of a specific extreme weather system (such as a typhoon or cold wave). Through this set of "feature patterns," AI models can identify which weather system is dominating the current meteorological field. Feature patterns are physics-oriented, used to identify and differentiate different types of extreme weather systems, while wind speed statistical feature values are data-oriented, used to quantify the local state and behavior of the wind speed sequence itself.
[0047] Specifically, 1. Each of the aforementioned extreme typhoon events is designated as a positive typhoon sample, the non-extreme typhoon events are designated as negative typhoon samples, the extreme cold wave events are designated as positive cold wave samples, and the non-extreme cold wave events are designated as negative cold wave samples. A sample library is created based on each typhoon positive and negative sample, cold wave positive and negative sample, the fingerprint features corresponding to each sample, and the time points marked as extreme events; 2. Feature standardization: The feature matrix generated from the positive and negative samples in the previous step is standardized to eliminate the influence of dimensions.
[0048] 3. Model training and pattern recognition: The model is trained based on the aforementioned sample database to construct an extreme event recognition model. Supervised learning algorithms are used to learn the complex mapping relationship between "fingerprint features" and "event labels." The trained model itself is an "extreme event recognizer." When new reanalysis data is input, it can not only determine whether it is an extreme event, but also provide a probability value. This probability value itself is an extremely powerful feature that can be input into the final model.
[0049] Based on the extreme event recognition model, the characteristic patterns of extreme events are obtained. A well-trained extreme event recognition model can generate extreme event probability features for each historical moment. These dynamic probability features will be incorporated into the statistical features and used as input to the DNN model, enabling it to "know" which weather system is currently in control, thereby making more accurate extreme wind speed predictions.
[0050] Explanation of why it is necessary to include characteristic patterns for identifying extreme events (such as typhoons and cold waves).
[0051] I. Even without constructing extreme event characteristic patterns, the core process of this solution can still run, and extreme wind speeds can be calculated in the following ways: 1. Data-driven: The model (DNN) can still learn the statistical relationship between common meteorological features (such as basic wind speed, air pressure, and temperature) in ERA5 reanalysis data and short-term measured wind speeds in the field. It remains a powerful nonlinear regression model.
[0052] 2. Application of extreme value theory: The model can still use GPD to model the tail of the prediction residual.
[0053] 3. Output results: The model can still output a long-term wind speed sequence, and the return period wind speed value can be obtained through extreme value analysis.
[0054] From a purely technical perspective, this process is complete and can produce a numerical result.
[0055] II. However, the lack of extreme event characteristic patterns will lead to a series of fundamental problems: 1. Lacking physical mechanisms, it has become a "black box" model. The model cannot understand the physical causes behind the strong winds. This makes its decision-making process lack a physical basis and reduces its interpretability.
[0056] 2. Insufficient extrapolation capability, leading to increased risk in predicting unknowns. Without characteristic patterns to describe the physical structure of extreme events, when a model encounters a typhoon of unprecedented intensity that has never appeared in short-term measured data, it cannot physically understand the system, and its predictions may be severely distorted or completely fail. Capturing such unprecedented extreme events is precisely the core purpose of extreme value analysis.
[0057] 3. Inability to effectively utilize transfer learning If the model does not build these feature patterns, then the knowledge learned from nearshore areas (how to identify typhoons) is difficult to transfer effectively to the deep sea. The advantage of using nearshore data to assist training is almost negated, and the model has to rely more on limited short-term data from the deep sea, making it more prone to overfitting.
[0058] S5. Based on the data in the fusion database, a hybrid model combining deep neural networks and extreme value theory is constructed. The hybrid model combining deep neural networks and extreme value theory is used to generate the long-term maximum wind speed sequence of the location of deep-sea wind farms.
[0059] Specifically, based on the fusion database constructed in stages S1-S4, a hybrid modeling method combining deep neural networks and extreme value theory is used to achieve accurate prediction of the long-term maximum wind speed sequence of observation points. The specific process is as follows: I. First stage: Data preparation.
[0060] 1. Integrate databases.
[0061] 2. Dataset partitioning: Divide the data into a training set (70%) and a validation set (30%).
[0062] Phase Two: Construction and training of a hybrid model combining deep neural networks (DNNs) and extremum theory (such as GPD) (core process) Model architecture: Input layer: Receives data (feature vectors) from the fused database after standardization.
[0063] Feature extraction backbone: A 4-layer deep fully connected network (DNN) is used, with 512 neurons in each layer.
[0064] Attention Mechanism: A multi-head self-attention layer is introduced after the third layer of the DNN. This mechanism enables the model to automatically weight key features of extreme events (such as high-pressure gradients and typhoon labels), significantly enhancing its ability to capture extreme wind processes.
[0065] Dual output heads: Main regression head: Outputs a scalar, namely the predicted mean wind speed under the current meteorological conditions (ŷ).
[0066] GPD Parameter Header: Outputs two parameters: scale parameter σ (ensured to be positive using Softplus activation) and shape parameter ξ. This header defines the probability distribution shape of the extreme value portion in the prediction residual (y - ŷ).
[0067] Among them, the scale parameter σ reflects the dispersion of the intensity of extreme events. The larger the σ is, the wider the range of extreme wind speed fluctuations. The shape parameter ξ determines the decay rate of the distribution tail. Softplus activation ensures that σ>0, which conforms to physical constraints.
[0068] Define the residual (y - ŷ) to characterize the model prediction error. When the residual > u (high-order threshold), the residual is considered an extreme event signal. By modeling the tail distribution of the residual using GPD, it is possible to separate the statistical characteristics of normal wind speed from those of extreme wind speed, while preserving the spatiotemporal correlation of the original wind speed.
[0069] The loss function plays a central role in deep learning and machine learning, and its significance is mainly reflected in: 1. The core driving force of model optimization The loss function quantifies the difference between the model's predicted values and the true values, providing a clear direction for optimizing model parameter updates. Huber Loss combines the advantages of MSE and MAE, maintaining gradient stability when the error is small and reducing the impact of outliers when the error is large.
[0070] 2. Regression tasks require matching a specific loss function.
[0071] Loss function: The total loss function consists of two weighted parts: Total Loss = HuberLoss(ŷ, y_true) + α * GPD_NLLLoss(ε, σ, ξ) In the formula: HuberLoss: Used for the main regression task, it is less sensitive to outliers than MSE.
[0072] GPD_NLLLoss: Negative log-likelihood loss for the generalized Pareto distribution. It is calculated only for sample points where the actual wind speed exceeds a preset high-order threshold (u, such as the 95th quantile). This allows the model to focus on learning the statistical behavior of the extreme tails.
[0073] α: A hyperparameter that weighs the two loss terms.
[0074] Third stage of model training (based on training set data): Optimizer: AdamW parameter optimization improves generalization ability Initial learning rate: 1e-4 (using a learning rate decay strategy) Batch size: 128 Training cycle: Training will be stopped early when the loss on the validation set no longer decreases (Early Stopping), usually 100-200 cycles.
[0075] Core objective: To enable the model to learn to use meteorological field information provided by ERA5 to accurately calculate wind speeds in deep-sea locations, especially to model the statistical distribution of extreme wind speeds.
[0076] Once the model is trained, it can generate long-term maximum wind speed sequences for the locations of deep-sea wind farms. The long-term maximum wind speed sequence was obtained using the following method: 1. Long-term wind speed reconstruction: 2. Input 30 years of ERA5 feature data into the trained final model. The data used here are for illustrative purposes only.
[0077] 3. The model will output a 30-year hourly conditional mean wind speed sequence and the corresponding GPD distribution parameter sequence.
[0078] 4. Monte Carlo simulation generates annual extreme values: For each year, a synthetic wind speed sequence for the entire year is randomly generated based on hourly conditional distributions through 10,000 Monte Carlo simulations. The maximum value is extracted from the annual sequence of each simulation to obtain a sample of the annual maximum wind speed, and the average of these 10,000 simulations is taken as the maximum wind speed for that year.
[0079] Repeat this process to generate a 30-year annual maximum wind speed sequence.
[0080] To verify and optimize the hybrid model combining deep neural networks and extreme value theory, in a preferred embodiment of this scheme, after constructing the hybrid model combining deep neural networks and extreme value theory based on the data in the fusion database, long-term measured data from near-shore wind farms are used to assist in optimizing the hybrid module, including: A nearshore dataset is constructed based on long-term measured data of nearshore wind farms and historical data of corresponding meteorological reanalysis data. The model weights are obtained by training the data in the nearshore dataset using a hybrid model that combines the deep neural network with the extreme value theory. The model weights are then used as initial values and loaded back into the hybrid model. Using training data from the deep sea, the hybrid model was fine-tuned with a low learning rate after the initial values were loaded back into the hybrid model, in order to optimize the model that combines deep neural networks with extremum theory.
[0081] Specifically, this step is an optional optimization strategy to further improve the robustness and accuracy of the model, especially effective when the quality of short-term experimental data is poor or the period is extremely short. Model weights are core parameters of neural networks in deep learning, essentially numerical coefficients connecting different neurons. Taking the linear transformation y = Wx + b as an example, W is the weight matrix, which determines the linear combination relationship of the features in the input x.
[0082] In hybrid models that combine deep learning and extremum theory, model weights are obtained after training. These weights are the core components of the model parameters and are used to adjust the relationship between the input data and the output prediction. After training the initial model with near-shore datasets, the model learns a set of initial weights; subsequently, when fine-tuning with deep-sea data at a low learning rate, these weights are further optimized to adapt to the new data distribution.
[0083] Specific operation method: Obtain long-term (e.g., 10-year) SCADA measured data and corresponding ERA5 data for another near-shore wind farm.
[0084] The nearshore dataset was constructed using the exact same method as in the first phase.
[0085] The constructed DNN-GPD model was pre-trained on a nearshore dataset. At this stage, the model learns a more general mapping relationship between wind speed and meteorological features.
[0086] The pre-trained model weights are used as initial values and loaded back into the model.
[0087] Using training data from the deep sea, the initial values are loaded back into the model at a low learning rate (e.g., 1e-5) for fine-tuning of the entire model or the last few layers. This allows the model to draw prior knowledge from abundant near-shore data and quickly adapt to the local characteristics of the deep sea, effectively mitigating the overfitting problem caused by insufficient deep-sea data.
[0088] S6. Using the long-term maximum wind speed sequence, the extreme wind speed at the hub height of the wind turbine in the deep-sea wind farm is calculated using a probability distribution model and a wind power exponent equation.
[0089] Specifically, using the long-term maximum wind speed sequence, the extreme wind speed at the hub height of the wind turbine in the deep-sea wind farm is calculated using a probability distribution model (optional extreme value type I probability distribution function).
[0090] Using the long-term maximum wind speed sequence, the extreme wind speed at the turbine hub height of the deep-sea wind farm was calculated using a probability distribution model and a wind power exponential equation. The probability distribution model adopted was of the Günbel extreme value type I, and the extreme wind speed at the turbine hub height was calculated using the following formula: middle, For reference height, The height of the wind turbine hub. For height is To measure the wind shear index, For height is The extreme wind speed at that time.
[0091] For example, using the generated 30-year annual maximum wind speed sequence, the maximum wind speed value once every 50 years at a height of 100m is calculated using a probability distribution model (with an optional extreme value type I probability distribution function). After obtaining the maximum wind speed value, it is substituted into the equation to obtain the extreme wind speed at the hub height of the wind turbine in the deep-sea wind farm.
[0092] This embodiment provides a non-transitory computer-readable storage medium that stores computer instructions, which cause a computer to execute the methods provided in the above-described method embodiments. Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium, and when executed, it performs the steps of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0093] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0094] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.
[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for calculating extreme wind speeds in deep-sea wind farms based on artificial intelligence, characterized in that, include: Acquire meteorological reanalysis data and short-term measured data, wherein the short-term measured data are meteorological data of preset observation points within a preset time period; Spatial matching is performed on the meteorological reanalysis data and the measured data to ensure that the historical data of the meteorological reanalysis data and the measured data both point to the same time point and the same geographical location, and a database containing the reanalysis data and the measured data is established. Extract wind speed-related data from the database to construct a wind speed time series. Use a sliding time window to calculate wind speed statistical feature values for the wind speed time series. The wind speed statistical feature values include at least the average wind speed, standard deviation, variance, maximum and minimum data values within the window, and the maximum rate of wind speed increase and decrease within the window. An extreme event identification model is constructed to generate dynamic probability features of extreme events for each historical moment. The dynamic probability features and the wind speed statistical feature values are incorporated into the database to form a fusion database. The extreme event identification model is used to identify which extreme weather system is currently in control. Based on the data in the fusion database, a hybrid model combining deep neural networks and extreme value theory is constructed. This hybrid model is used to generate long-term maximum wind speed sequences for the locations of deep-sea wind farms. Using the long-term maximum wind speed sequence, the extreme wind speed at the hub height of the wind turbine in the deep-sea wind farm was calculated using a probability distribution model and a wind power exponent equation.
2. The method according to claim 1, characterized in that, Using the long-term maximum wind speed sequence, the extreme wind speed at the hub height of the wind turbine in the deep-sea wind farm was calculated using a probability distribution model and a wind power exponential equation. The probability distribution model adopted was of the Gumbel extreme value type I. The extreme wind speed at the hub height was calculated using the following formula: middle, For reference height, The height of the wind turbine hub. For height is High wind speed, To measure the wind shear index, For height is The extreme wind speed at that time.
3. The method according to claim 1, characterized in that, The architecture of the hybrid model combining deep neural networks and extremum theory includes: Input layer: The data in the fusion database; Feature extraction backbone: A 4-layer deep fully connected network with 512 neurons in each layer; Attention mechanism: A multi-head self-attention layer is introduced after the third layer of the fully connected network. This attention mechanism enables the model to automatically weight the key features of extreme events. The primary regression head in the dual-output head outputs the predicted average wind speed under current meteorological conditions. The GPD parameter header in the dual-output header outputs scale and shape parameters. This GPD parameter header defines the probability distribution shape of the extreme wind speed component in the prediction residual.
4. The method according to claim 1, characterized in that, After constructing a hybrid model combining deep neural networks and extreme value theory based on the data in the fused database, the hybrid model is then optimized using long-term measured data from nearshore wind farms, including: A nearshore dataset is constructed based on long-term measured data of nearshore wind farms and historical data of corresponding meteorological reanalysis data. The model weights are obtained by training the data in the nearshore dataset using a hybrid model that combines the deep neural network with the extreme value theory. The model weights are then used as initial values and loaded back into the hybrid model. Using training data from the deep sea, the hybrid model was fine-tuned with a low learning rate after the initial values were loaded back into the hybrid model, in order to optimize the model that combines deep neural networks with extremum theory.
5. The method according to claim 1, characterized in that, Meteorological feature data related to wind speed is extracted from historical data in the database, and short-term measured wind speed data is extracted from measured data. The meteorological feature data and short-term measured wind speed data are then used to construct a wind speed time series.
6. The method according to claim 1, characterized in that, Before spatially matching the historical and measured data of the meteorological reanalysis data to establish a database, the following steps are also included: The historical and measured data of the meteorological reanalysis data are subjected to quality control and consistency checks to remove abnormal data. The quality control involves removing the historical and measured data that exceed the preset value range.
7. The method according to claim 6, characterized in that, The historical and measured data of the meteorological reanalysis data are spatially matched to establish a database, including: Time matching is performed on historical data and measured data of meteorological reanalysis data to ensure that the timestamps of historical data and measured data of meteorological reanalysis data are consistent; Bilinear interpolation is used to spatially match historical and measured data of meteorological reanalysis data, so that the historical and measured data of meteorological reanalysis data have the same geographical location. Historical and measured meteorological reanalysis data, after time and spatial matching, are merged to establish a database. The database includes at least timestamp sequences, latitude, longitude, wind speed, relevant meteorological elements, and time.
8. The method according to claim 1, characterized in that, The method also includes constructing a set of fingerprint features for extreme events: They are identified and labeled according to the definition of extreme events, namely typhoon events / cold wave events; Extract multi-dimensional meteorological physical quantities from each labeled cold wave / typhoon event within a preset time period before and after the event to construct a fingerprint feature set of extreme events. The fingerprint feature set includes at least the event type and the meteorological physical quantity features corresponding to the event type.
9. The method according to claim 8, characterized in that, The method further includes: Each extreme typhoon event was set as a positive typhoon sample, non-extreme typhoon events were set as negative typhoon samples, extreme cold wave events were set as positive cold wave samples, and non-extreme cold wave events were set as negative cold wave samples. A sample library is created based on each positive and negative sample of a typhoon, a positive and negative sample of a cold wave, and the fingerprint features corresponding to each sample. The model is trained based on the sample library to construct an extreme event recognition model; The extreme event identification model yields the feature patterns of extreme events, which are used to identify and distinguish different types of extreme weather.
10. The method according to claim 1, characterized in that, The measured data are meteorological data measured by a floating lidar.