Offshore wind power-oriented ERA5 maximum wind speed hybrid machine learning correction method and system
By constructing a hybrid machine learning framework that combines random forest and deep learning models, the error problem of ERA5 wind speed data in offshore wind power under extreme weather conditions is solved, achieving higher accuracy and applicability of wind speed correction, and supporting safety early warning and power generation efficiency optimization of offshore wind farms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-03-31
AI Technical Summary
In offshore wind power, existing technologies show significant errors in ERA5 reanalysis wind speed data under extreme weather conditions. Traditional machine learning models lack adaptability and robustness, failing to meet the needs of safety early warning and power generation efficiency optimization. Furthermore, they fail to effectively integrate marine environmental data with topographic features, ignoring the impact of wind-wave coupling and coastal effects on wind speed.
A hybrid machine learning framework is adopted, combining random forest and deep learning models. By constructing ocean dynamic coupling features and time-series lag features, an oversampling strategy is used to balance typhoon data, and a confidence-driven dynamic selection mechanism is introduced to output the corrected maximum wind speed field.
It improves the accuracy and applicability of maximum wind speed correction, solves the bias problem of traditional models in extreme wind speed prediction, provides more accurate wind speed data support, and ensures safety early warning and power generation efficiency optimization of offshore wind farms.
Smart Images
Figure CN121766485A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a hybrid machine learning correction method and system for ERA5 maximum wind speed for offshore wind power. Background Technology
[0002] Currently, ERA5 reanalysis wind speed data for offshore wind power exhibits significant errors in practical applications, especially under extreme weather conditions such as typhoons. Traditional correction methods struggle to meet accuracy requirements. Existing technologies often employ single machine learning models, such as random forests or deep learning networks, which lack adaptability to different wind speed scenarios. While random forests have advantages in feature selection and interval prediction, they are insufficient in capturing temporal dynamic features. Although deep learning models can model time dependencies, they rely on manual intervention in feature selection and high-dimensional data processing, resulting in limited generalization ability in complex marine meteorological environments. Furthermore, conventional methods have not optimized for the imbalanced typhoon data sample problem, making the models prone to bias in extreme event predictions. They also lack a dynamic model selection mechanism and cannot adaptively output the optimal correction result based on confidence levels, thus limiting their applicability in the refined operation of offshore wind power.
[0003] Existing methods largely rely on a single data source, the ERA5 initial field, for feature construction, failing to effectively integrate marine environmental data with topographic features. They also neglect the impact of key physical processes such as wind-wave coupling and coastal effects on wind speed. Furthermore, during model training, typhoon data, due to its sparse sample size, is often overwhelmed by regular wind speed data, resulting in insufficient learning of extreme wind speed events. Traditional loss functions are sensitive to outliers, easily leading to gradient instability in high-wind-speed typhoon scenarios, affecting convergence. In addition, existing systems typically use fixed model outputs and lack dynamic decision-making mechanisms based on error verification results, failing to ensure the reliability and robustness of correction results under variable meteorological conditions. This makes it difficult to meet the urgent needs of offshore wind farms in safety early warning and power generation efficiency optimization. Therefore, improving the applicability of these methods in the refined operation of offshore wind power and meeting the needs of offshore wind farms in safety early warning and power generation efficiency optimization has become an urgent problem to be solved. Summary of the Invention
[0004] This invention provides a hybrid machine learning correction method and system for ERA5 maximum wind speed in offshore wind power, in order to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides a hybrid machine learning correction method for ERA5 maximum wind speed in offshore wind power, comprising: S1. Based on the input ERA5 initial field, marine environmental data and topographic data, marine dynamic coupling features and temporal lag features are created to obtain the feature dataset; S2, Use the random forest algorithm to pre-train the feature dataset, and filter key features to output wind speed interval prediction in probability form; S3, apply an oversampling strategy to the typhoon data in the feature dataset, perform fine training on the feature dataset and the wind speed interval prediction, and obtain the corrected wind speed value; S4, segment the verification dataset from the feature dataset, perform error verification on the corrected wind speed value based on the verification dataset, and generate error verification results; S5. Based on the error verification results, a confidence-driven dynamic selection mechanism is used to output the corrected maximum wind speed field.
[0006] In a preferred embodiment, the creation of marine dynamic coupling features and temporal lag features based on the input ERA5 initial field, marine environmental data, and topographic data yields a feature dataset, including: Based on the input ERA5 initial field, marine environmental data, and topographic data, feature engineering functions are applied to calculate wind-wave coupling characteristics to obtain wind-wave coupling characteristic values. Based on the input ERA5 initial field, marine environmental data, and topographic data, feature engineering functions are applied to calculate the coastal effect characteristics to obtain the coastal effect characteristic values. Based on the input ERA5 initial field, marine environmental data and topographic data, temporal lag features are extracted using the sliding window method to obtain a temporal lag feature set; Based on the wind-wave coupling feature value, the coastal effect feature value, and the time lag feature set, feature combination is performed to obtain a 15-dimensional feature dataset.
[0007] In a preferred embodiment, the feature combination based on the wind-wave coupling feature value, the coastal effect feature value, and the time-series lag feature set is performed to obtain a 15-dimensional feature dataset, including: The wind and wave coupling feature value, the coastal effect feature value, and the time lag feature set are concatenated according to their dimensions to form a 15-dimensional feature vector, resulting in a 15-dimensional feature dataset. In the 15-dimensional feature vector, each row corresponds to a sample point, and each column corresponds to a feature dimension.
[0008] In a preferred embodiment, the step of pre-training the feature dataset using a random forest algorithm and filtering key features to output probability forms for wind speed interval prediction includes: Based on the aforementioned feature dataset, the top 8 key features are selected using Gini importance ranking to obtain the key feature set; Based on the aforementioned key feature set, a random forest model with 500 decision trees is trained to obtain a pre-trained model; Based on the pre-trained model, the input data is predicted, and the wind speed range prediction is output in probabilistic form.
[0009] In a preferred embodiment, training a random forest model with 500 decision trees based on the key feature set to obtain a pre-trained model includes: Multiple training subsets are generated through autopilot sampling, and a decision tree is trained in each subset. Gini impurity is used as the node splitting criterion. The best features and split points are selected recursively to maximize information gain and build the tree structure until the preset minimum number of leaf nodes or maximum depth is reached. Pre-trained models are obtained by integrating predictions through averaging.
[0010] In a preferred embodiment, the step of applying an oversampling strategy to the typhoon data in the feature dataset to perform fine-tuning training on the feature dataset and the wind speed interval prediction to obtain corrected wind speed values includes: Based on the feature dataset and the wind speed interval prediction, an oversampling strategy is applied to the typhoon data in the feature dataset to obtain an enhanced training dataset. Based on the enhanced training dataset, a 12-hour time series window is processed through an LSTM layer to obtain a time series feature representation; Based on the temporal feature representation, regression training is performed using a Dense layer, and the Huber loss function is used to optimize the wind speed interval prediction to obtain the corrected wind speed value.
[0011] In a preferred embodiment, the mathematical expression of the Huber loss function is as follows: ; In the formula, This represents Huber's loss value. This represents the actual wind speed value. To predict wind speed values, This is the threshold parameter.
[0012] In a preferred embodiment, the step of segmenting a validation dataset from the feature dataset, performing error verification on the corrected wind speed value based on the validation dataset, and generating an error verification result includes: Based on the aforementioned feature dataset, an independent validation dataset is segmented using a random segmentation method to obtain the validation dataset; Based on the validation dataset and the corrected wind speed value, the root mean square error and accuracy index are applied to evaluate the error and obtain the error index. Based on the error metric, an error verification result including model confidence is generated.
[0013] In a preferred embodiment, the step of outputting the corrected maximum wind speed field based on the error verification result using a confidence-driven dynamic selection mechanism includes: Based on the model confidence score in the error verification results, the optimal model is dynamically selected to obtain the selected model; Based on the selected model, the corrected maximum wind speed field is generated.
[0014] To address the aforementioned problems, this invention also provides an ERA5 maximum wind speed hybrid machine learning correction system for offshore wind power, the system comprising: The feature engineering module is used to create marine dynamic coupling features and time-series lag features based on the input ERA5 initial field, marine environmental data and topographic data, to obtain the feature dataset; The random forest pre-training module is used to pre-train the feature dataset using the random forest algorithm, and to filter key features to output wind speed interval predictions in probability form. The DNN fine correction training module is used to apply an oversampling strategy to the typhoon data in the feature dataset, and to perform fine training on the feature dataset and the wind speed interval prediction to obtain the corrected wind speed value. The model validation module is used to segment a validation dataset from the feature dataset, perform error validation on the corrected wind speed value based on the validation dataset, and generate error validation results. The dynamic integrated output module is used to output the corrected maximum wind speed field based on the error verification results and using a confidence-driven dynamic selection mechanism.
[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a hybrid machine learning framework of "random forest + deep learning," effectively integrating the core advantages of both models and significantly improving the accuracy and applicability of ERA5 maximum wind speed correction. The random forest model, leveraging Gini importance ranking, can accurately select the top 8 key features, such as 10-meter wind speed, significant wave height, and coastline distance, quickly completing feature dimensionality reduction and probabilistic wind speed interval prediction, thus solving the problems of feature redundancy and low computational efficiency in traditional models. Meanwhile, the deep learning model can capture the dynamic evolution of wind speed through a 12-hour time series window. Combined with the Huber loss function, it uses squared error to optimize conventional wind speed prediction when the error is less than or equal to the threshold, and uses absolute error to reduce extreme value interference when the error is greater than the threshold, achieving refined regression correction of wind speed. The synergy of the two avoids the shortcomings of random forest in capturing dynamic time series features and makes up for the shortcomings of deep learning in relying on manual feature selection and limited generalization ability, allowing the model to maintain stable prediction performance under both conventional meteorological conditions and complex marine environments.
[0016] 2. To address the pain point of large prediction bias in extreme wind speeds caused by scarce typhoon data samples, this invention balances the dataset through an oversampling strategy, enhancing the model's learning ability for high-wind-speed scenarios such as typhoons. This solves the problem of typhoon data being "overwhelmed" by regular wind speed data and insufficient prediction accuracy for extreme events in traditional methods. Simultaneously, a confidence-driven dynamic selection mechanism is introduced. Based on the root mean square error and accuracy indicators, the model confidence is calculated, and a random forest or deep learning model is dynamically selected as the optimal output model. When the confidence of the deep learning model is higher in extreme scenarios such as typhoons, its correction result is prioritized; in regular scenarios, the random forest output is selected. This ensures that a highly reliable maximum wind speed field can be output under different meteorological conditions. This mechanism avoids the limitations of fixed models in variable scenarios and provides more accurate wind speed data support for safety early warning and power generation efficiency optimization of offshore wind farms. Attached Figure Description
[0017] Figure 1 A flowchart illustrating the ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power provided in an embodiment of the present invention. Figure 2 This is a functional block diagram of an ERA5 maximum wind speed hybrid machine learning correction system for offshore wind power provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0019] This application provides a hybrid machine learning correction method for ERA5 maximum wind speed in offshore wind power. The execution entity of this hybrid machine learning correction method for ERA5 maximum wind speed in offshore wind power includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the hybrid machine learning correction method for ERA5 maximum wind speed in offshore wind power can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.
[0020] Reference Figure 1The diagram shown is a flowchart illustrating a hybrid machine learning method for ERA5 maximum wind speed correction in offshore wind power, provided by an embodiment of the present invention. In this embodiment, the hybrid machine learning method for ERA5 maximum wind speed correction in offshore wind power includes: S1. Based on the input ERA5 initial field, marine environmental data and topographic data, marine dynamic coupling features and temporal lag features are created to obtain the feature dataset; In this embodiment of the invention, the creation of marine dynamic coupling features and temporal lag features based on the input ERA5 initial field, marine environmental data, and topographic data yields a feature dataset, including: Based on the input ERA5 initial field, marine environmental data, and topographic data, feature engineering functions are applied to calculate wind-wave coupling characteristics to obtain wind-wave coupling characteristic values. Based on the input ERA5 initial field, marine environmental data, and topographic data, feature engineering functions are applied to calculate the coastal effect characteristics to obtain the coastal effect characteristic values. Based on the input ERA5 initial field, marine environmental data and topographic data, temporal lag features are extracted using the sliding window method to obtain a temporal lag feature set; Based on the wind-wave coupling feature value, the coastal effect feature value, and the time lag feature set, feature combination is performed to obtain a 15-dimensional feature dataset.
[0021] It should be noted that the process of calculating the wind-wave coupling characteristics using the feature engineering function is based on the 10-meter wind speed in the ERA5 initial field, the significant wave height in the marine environmental data, and the sea surface temperature in the ERA5 initial field. Specifically, the wind-wave coupling characteristics are obtained by multiplying the 10-meter wind speed by the significant wave height and then dividing the product by the sea surface temperature.
[0022] Furthermore, the wind-wave coupling characteristics are calculated to quantify the conversion efficiency of wind energy to wave energy. Here, the 10-meter wind speed represents the wind dynamics near the sea surface, the effective wave height characterizes the wave state, and the sea surface temperature affects the energy exchange at the air-ocean interface. The influence of temperature variation is eliminated by ratio calculation, highlighting the intensity of wind-wave interaction.
[0023] It should be noted that the wind-wave coupling characteristic value is a scalar value, representing the wind-wave coupling strength at a specific point in time and space. Essentially, it captures the modulation effect of ocean dynamic processes on wind speed.
[0024] It should be noted that the process of calculating the coastal effect features using the feature engineering function is based on the coastline distance in the terrain data and the boundary layer height in the ERA5 initial field. Specifically, the coastal effect features are obtained by multiplying the coastline distance by -0.1 and then performing an exponential operation, and then multiplying the result of the exponential operation by the boundary layer height.
[0025] Furthermore, the coastal effect characteristics are calculated to simulate the damping effect of the coastline on wind speed. The exponential term indicates that the coastal effect decreases exponentially with increasing distance from the coastline, and the boundary layer height represents the depth of the atmospheric mixing layer. The multiplication operation combines the two to quantify the spatial modulation effect of topography on wind speed.
[0026] It should be noted that the coastal effect eigenvalue is a scalar value that represents the degree of influence of coastal topography on wind speed. It is used to enhance the prediction accuracy of the model in nearshore areas and essentially reflects the wind speed gradient change caused by the thermal difference between land and sea.
[0027] It should be noted that the process of extracting time-series lag features using the sliding window method is based on the time series of the input data. A fixed time window of 12 hours is set by default. The feature values of the lag time points are extracted by sliding within the window, generating multiple lag features.
[0028] Furthermore, the sliding window method captures the dynamic evolution patterns of wind speed through time series analysis, such as extracting feature values from the previous hour to capture the inertial effect and periodic changes of wind speed.
[0029] It should be noted that the time-series lag feature set is a collection of multiple feature values representing the meteorological and oceanic conditions at historical moments. It is used to improve the temporal consistency of wind speed corrections and to quantify the memory effect and lag response of wind speed evolution.
[0030] Furthermore, the input data layer includes key variables in the ERA5 initial field: 10m / 100m wind speed (U / V components), sea surface temperature (SST), boundary layer height (BLH), and sea level pressure; Time resolution: 1 hour (encrypted to 30 minutes during typhoons); Marine environmental data: Satellite remote sensing: sea surface roughness (from SAR imagery), significant wave height (Hs, from Jason-3); Buoy measurements: 10-minute wind speed and direction (12 buoy stations in China's coastal waters); Topographic data: seabed slope (derived from ETOPO1 data), coastline distance (used to correct for sea and land breeze effects).
[0031] Furthermore, the feature engineering processing code example is as follows: Python # Constructing marine dynamic coupling characteristics def create_marine_features(era5, ocean): features['wind_wave_coupling'] = era5['ws10']* ocean['Hs'] / era5['SST'] features['coastal_effect']=np.exp(-0.1 *ocean['coast_distance']) *era5['blh'] Return features.
[0032] In this embodiment of the invention, the feature combination based on the wind-wave coupling feature value, the coastal effect feature value, and the time-series lag feature set to obtain a 15-dimensional feature dataset includes: The wind and wave coupling feature value, the coastal effect feature value, and the time lag feature set are concatenated according to their dimensions to form a 15-dimensional feature vector, resulting in a 15-dimensional feature dataset. In the 15-dimensional feature vector, each row corresponds to a sample point, and each column corresponds to a feature dimension.
[0033] It should be noted that the 15-dimensional feature dataset is a structured data matrix, with each row corresponding to a sample point and each column corresponding to a feature dimension. It is used for the training and prediction of subsequent random forest and DNN models to construct a feature space that comprehensively represents ocean-atmosphere interactions, in order to support high-precision wind speed correction.
[0034] Furthermore, feature combination integrates multi-source features through vectorization operations, ensuring that each feature dimension represents a different physical process and together describes the key factors affecting wind speed.
[0035] S2, Use the random forest algorithm to pre-train the feature dataset, and filter key features to output wind speed interval prediction in probability form; In this embodiment of the invention, the step of pre-training the feature dataset using the random forest algorithm and filtering key features to output probability forms for wind speed interval prediction includes: Based on the aforementioned feature dataset, the top 8 key features are selected using Gini importance ranking to obtain the key feature set; Based on the aforementioned key feature set, a random forest model with 500 decision trees is trained to obtain a pre-trained model; Based on the pre-trained model, the input data is predicted, and the wind speed range prediction is output in probabilistic form.
[0036] It should be noted that the Gini importance ranking is based on the aforementioned feature dataset. The Gini impurity reduction of each feature at the decision tree node split is calculated, and the feature importance score is obtained by averaging across all decision trees. The top 8 features are then selected in descending order of their scores to obtain the key feature set.
[0037] Furthermore, for a given node in the dataset, its Gini impurity is calculated as follows: ,in, In wind speed range prediction, the number of categories corresponds to ranges such as 3-5 m / s, 5-10 m / s, etc. For the sample to belong to the first The probability of a class, the Gini impurity value ranges from 0 to 1. The smaller the value, the higher the purity of the node, that is, the more consistent the sample classes. In the training of random forest, features and split points are recursively selected to minimize the Gini impurity, thereby splitting the nodes and constructing a decision tree structure to improve the model's classification accuracy for wind speed ranges.
[0038] Furthermore, Gini importance ranking is used to quantify the contribution of each feature to wind speed prediction. By evaluating the information gain of features at split nodes, the most discriminative features are selected, reducing redundancy and noise and improving model efficiency.
[0039] It should be noted that the key feature set is a subset containing the first 8 important features, which is used for subsequent model training. Essentially, it optimizes the feature space, focusing on core physical factors such as sea surface temperature, significant wave height, and coastline distance, in order to improve the accuracy and computational efficiency of wind speed correction.
[0040] It should be noted that the probabilistic form of wind speed interval prediction is the probability distribution output by the pre-trained model, which represents the probability that the input data belongs to each wind speed interval (such as 3-5 m / s, 5-10 m / s, >25 m / s). It is used to quantify the uncertainty of wind speed and support risk assessment and subsequent fine correction of the DNN model.
[0041] In this embodiment of the invention, training a random forest model with 500 decision trees based on the key feature set to obtain a pre-trained model includes: Multiple training subsets are generated through autopilot sampling, and a decision tree is trained in each subset. Gini impurity is used as the node splitting criterion. The best features and split points are selected recursively to maximize information gain and build the tree structure until the preset minimum number of leaf nodes or maximum depth is reached. Pre-trained models are obtained by integrating predictions through averaging.
[0042] It should be noted that training a random forest model with 500 decision trees reduces the risk of overfitting by integrating multiple decision trees. The number of 500 trees balances computational cost and prediction stability, ensuring the model's generalization ability in complex meteorological data and adapting to the variable conditions of the offshore wind power environment.
[0043] Furthermore, the specific training content for training a random forest model with 500 decision trees is based on a key feature set. Multiple training subsets are generated through bootstrap sampling, and each training subset is used to train a decision tree independently. The bootstrap sampling process randomly extracts samples from the key feature set, allowing repeated sampling to form a subset of the same size as the original dataset, ensuring the diversity of training data for each decision tree. During decision tree training, Gini impurity is used as the node splitting criterion. The best features and split points are selected recursively to maximize information gain and build the tree structure until the preset minimum number of samples or maximum depth of leaf nodes is reached. Gini impurity is an indicator used for node splitting in the decision tree algorithm to measure the impurity of sample categories in the dataset.
[0044] Furthermore, the averaging ensemble prediction refers to the random forest model generating the final output by arithmetically averaging the predictions of all decision trees for regression tasks. Specifically, for each input sample, each decision tree independently outputs a wind speed prediction value, and then the prediction values of all trees are averaged to obtain the corrected wind speed value. This ensemble method improves the generalization ability and robustness of the model by reducing the variance and bias of individual trees, adapts to the spatiotemporal variability of wind speed data in offshore wind power environments, and ensures the stability of the pre-trained model under complex meteorological conditions.
[0045] It should be noted that the pre-trained model is a trained random forest model that can output the probability of wind speed range prediction for new input data, which is used to initially correct the normal wind speed and provide a foundation for subsequent fine training.
[0046] S3, apply an oversampling strategy to the typhoon data in the feature dataset, perform fine training on the feature dataset and the wind speed interval prediction, and obtain the corrected wind speed value; In this embodiment of the invention, the step of applying an oversampling strategy to the typhoon data in the feature dataset to perform fine training on the feature dataset and the wind speed interval prediction to obtain corrected wind speed values includes: Based on the feature dataset and the wind speed interval prediction, an oversampling strategy is applied to the typhoon data in the feature dataset to obtain an enhanced training dataset. Based on the enhanced training dataset, a 12-hour time series window is processed through an LSTM layer to obtain a time series feature representation; Based on the temporal feature representation, regression training is performed using a Dense layer, and the Huber loss function is used to optimize the wind speed interval prediction to obtain the corrected wind speed value.
[0047] It should be noted that the oversampling strategy is based on typhoon data in the feature dataset. Synthetic samples are generated using the SMOTE algorithm to increase the number of typhoon event samples and obtain an enhanced training dataset. The process of generating synthetic samples is based on minority class samples in the typhoon data. First, for each minority class sample, the algorithm calculates its 5 nearest neighbor samples and uses Euclidean distance to measure similarity in the feature space. Then, a nearest neighbor sample is randomly selected and a synthetic sample is generated. This sample is located on the line connecting the original sample and the nearest neighbor sample.
[0048] Furthermore, the feature values of the synthesized samples are calculated using linear interpolation. For each feature dimension, a random number λ in the range [0,1] is generated, and then the new feature value is calculated. ,in, These are the feature values of the original samples. The nearest neighbor sample feature value is used, and this process is repeated until the number of minority class samples is balanced with that of majority class samples, thereby enhancing the representativeness of typhoon events in the training dataset and solving the sample imbalance problem.
[0049] Furthermore, an oversampling strategy is applied to address the imbalance problem in typhoon data samples. By artificially generating minority class samples, the model's ability to learn about extreme wind speed events is improved, avoiding bias towards the majority class of regular wind speeds and ensuring a more uniform distribution of training data.
[0050] It should be noted that the augmented training dataset is training data that has undergone oversampling processing. It contains original feature data and synthetic typhoon data and is used to train the LSTM-DNN hybrid model. Essentially, it balances the dataset to improve the model's prediction performance under extreme weather conditions and enhance the robustness of corrected wind speeds.
[0051] It should be noted that the 12-hour time series window is processed through the LSTM layer based on the time series sequences in the augmented training dataset, and the time-dependent features are extracted using the Long Short-Term Memory network layer to obtain the time series feature representation.
[0052] Furthermore, the LSTM layer captures long-term dependencies through a gating mechanism and processes 12-hour windows to simulate short-term evolution patterns of wind speed, such as memorizing meteorological conditions from previous moments to capture the inertial effects and periodic changes of wind speed, thereby enhancing temporal consistency.
[0053] It should be noted that the temporal feature representation is the hidden state vector output by the LSTM layer, which represents the compressed temporal information of the input sequence and is used for wind speed correction in subsequent regression layers. Essentially, it encodes the influence of historical meteorological conditions on the current wind speed and quantifies the dynamic process of wind speed evolution.
[0054] It should be noted that the regression training using the Dense layer is based on temporal feature representation. The high-dimensional features are mapped to wind speed values by performing nonlinear transformation through the fully connected layer, and the Huber loss function is used to optimize the model parameters to obtain the corrected wind speed values.
[0055] Furthermore, the process of performing a nonlinear transformation to map high-dimensional features to wind speed values is based on temporal feature representation. The fully connected layer performs a linear transformation on the input features through the weight matrix and bias vector, and then an activation function is applied to introduce nonlinearity for the input temporal feature representation. Where d is the feature dimension; Fully connected layer computation ,in, This is the weight matrix. For bias vectors, Linear output; Through activation function The non-linear activation values are obtained, and multiple Dense layers are stacked to learn the complex interactions between features. Finally, the output layer uses a linear activation function to map the high-dimensional features to a single wind speed value, achieving accurate regression from time-series features to corrected wind speed.
[0056] Furthermore, the initial values of the weights in the weight matrix are usually set to small random numbers, selected from the interval... The weights are extracted from a uniform distribution to avoid gradient vanishing or exploding problems. During training, the weights are iteratively updated using the backpropagation algorithm based on the gradient of the Huber loss function, eventually converging to the optimal value to minimize prediction error. The specific values of the weights depend on the dimension of the input features, the size of the hidden layer, and the distribution of the training data, and have no fixed values. However, regularization constraints prevent overfitting and ensure the model's generalization ability. Furthermore, the weight values are iteratively updated using the backpropagation algorithm based on the gradient of the Huber loss function, as follows: For each weight parameter, the partial derivative of the Huber loss function with respect to that weight is calculated. When the absolute value of the prediction error is less than or equal to the threshold parameter, the gradient is calculated as the prediction error multiplied by the input value of the neuron corresponding to that weight. When the absolute value of the prediction error is greater than the threshold parameter, the gradient is calculated as the threshold parameter multiplied by the sign of the error and then multiplied by the input value of the neuron corresponding to that weight. Finally, the weight values are iteratively adjusted using an optimization algorithm based on the gradient direction. Through multiple iterations, the Huber loss function is gradually minimized, causing the model weights to converge to the optimal value.
[0057] Furthermore, the mean value is first calculated to obtain the first moment estimate of the gradient, and the uncentered variance is calculated to obtain the second moment estimate. Then, the first and second moments are biased and corrected. Finally, the weight parameters are updated based on the corrected moment estimates. The bias correction is achieved by dividing by the complement of the decay factor, and the weight update is adjusted by dividing the corrected first moment by the square root of the corrected second moment and then multiplying by the learning rate.
[0058] Furthermore, the decay factor includes a decay factor estimated by the first moment and a decay factor estimated by the second moment, wherein the decay factor estimated by the first moment has a value of 0.9 and the decay factor estimated by the second moment has a value of 0.999. The learning rate has a value of 0.001.
[0059] Furthermore, this update process balances accurate fitting under normal wind speeds with robustness under extreme wind speeds by dynamically adjusting the weight parameters, ensuring the model's generalization ability in the complex environment of offshore wind power.
[0060] Furthermore, the Dense layer learns a complex mapping from features to wind speed through the weight matrix and activation function. The Huber loss function combines the advantages of mean squared error and absolute error, maintaining stable training under both normal and extreme wind speeds, reducing the impact of outliers, and improving the model's generalization ability.
[0061] It should be noted that the corrected wind speed value is a continuous numerical value output by the model, representing the predicted wind speed after fine training. It is used to replace the original ERA5 data to improve the accuracy of wind speed correction, especially in extreme events such as typhoons. Essentially, it provides a more reliable wind speed estimate to support offshore wind power operation decisions.
[0062] Furthermore, the code example for the DNN fine-tuning stage is as follows: keras # LSTM-DNN hybrid structure (specifically for typhoon season) model = Sequential() model.add(LSTM(64, input_shape=(12,15), return_sequences=True)) # Processing 12-hour time series windows model.add(Dropout(0.2)) model.add(Dense(32,activation='relu', kernel_regularizer=l2(0.01))) model.add(Dense(1, activation='linear')) # Outputs the corrected wind speed value. In this embodiment of the invention, the mathematical expression of the Huber loss function is as follows: ; In the formula, This represents Huber's loss value. This represents the actual wind speed value. To predict wind speed values, This is the threshold parameter.
[0063] It should be noted that when the absolute value of the prediction error is less than or equal to the threshold parameter, the squared error is used to calculate the loss. When the absolute value of the prediction error is greater than the threshold parameter, the absolute error is used to calculate the loss and the adjustment term is subtracted. This makes the loss function smooth when the error is small and linearly increase when the error is large, balancing the sensitivity to outliers and optimizing the model's performance under extreme wind speeds.
[0064] Furthermore, the threshold parameter is used to control the critical point at which the loss function transitions from squared error to absolute error, and its value is 1.0.
[0065] S4, segment the verification dataset from the feature dataset, perform error verification on the corrected wind speed value based on the verification dataset, and generate error verification results; In this embodiment of the invention, the step of segmenting a verification dataset from the feature dataset, performing error verification on the corrected wind speed value based on the verification dataset, and generating an error verification result includes: Based on the aforementioned feature dataset, an independent validation dataset is segmented using a random segmentation method to obtain the validation dataset; Based on the validation dataset and the corrected wind speed value, the root mean square error and accuracy index are applied to evaluate the error and obtain the error index. Based on the error metric, an error verification result including model confidence is generated.
[0066] It should be noted that the independent validation dataset is segmented by random partitioning based on the aforementioned feature dataset. Random sampling techniques are used to divide the dataset into training and validation sets. The validation set is used for model evaluation to obtain the model weights and structure of the validation data set, optimize the prediction function, and preliminarily correct conventional wind speeds. Essentially, it provides labeled data to drive model learning and support subsequent validation and deployment.
[0067] Furthermore, the random partitioning method, by ensuring the randomness of data distribution, prevents model overfitting and improves the generalization ability of the evaluation results. It is suitable for the spatiotemporal variability of offshore wind power data. The random partitioning method uses simple random sampling technology to randomly divide the feature dataset into mutually exclusive subsets according to a preset ratio of 70% training set and 30% validation set. Specifically, a pseudo-random number generator is used to assign random numbers to each sample, and the partitioning is sorted according to the random numbers to ensure that each sample has an equal probability of being selected into the training set or validation set. This method, by maintaining the randomness of data distribution, prevents the model from overfitting the training data and improves the generalization ability of the evaluation results, making it suitable for the spatiotemporal variability of offshore wind power data.
[0068] It should be noted that the validation dataset is a subset independently split from the feature dataset, used to test the model's performance on unseen data. Essentially, it provides an unbiased evaluation benchmark to verify the accuracy and reliability of the corrected wind speed values.
[0069] It should be noted that the application of root mean square error and accuracy metrics for error assessment is based on the true wind speed value and corrected wind speed value in the validation dataset. The prediction error is quantified through mathematical calculation to obtain the error metrics.
[0070] Furthermore, the root mean square error is used to calculate the average deviation between the predicted and actual values, while the accuracy metric assesses the proportion of predictions falling within a specific range. Together, they provide a comprehensive error analysis to identify the model's performance under different wind speed conditions.
[0071] It should be noted that the mathematical expression for the root mean square error is as follows: ; In the formula, RMSE is the root mean square error value. For the sample size, For sample index, For the first The true wind speed value of each sample For the first Corrected wind speed values for each sample.
[0072] Furthermore, the calculation process of the root mean square error is as follows: First, calculate the square of the prediction error for each sample, then calculate the average of the squares of the errors for all samples, and finally take the square root of the average. This allows the root mean square error value to reflect the overall magnitude of the prediction error, is sensitive to outliers, and is suitable for evaluating the overall accuracy of wind speed correction.
[0073] It should be noted that the accuracy metric is the ratio of the number of correctly predicted samples to the total number of samples, i.e. ,in, For accuracy, A true positive result indicates that the predicted interval matches the actual interval. The result is a true negative, indicating that the non-target interval has been correctly excluded. It was a false positive. It is a false negative.
[0074] Furthermore, for each sample, the model outputs a wind speed interval prediction in probabilistic form, selects the interval with the highest probability as the prediction result, compares it with the actual interval to count the number of correct predictions, and finally calculates the accuracy value.
[0075] Furthermore, the accuracy metric is used to evaluate the model's performance in wind speed range classification, such as determining whether the prediction falls within the true range, providing a measure of classification accuracy, and supporting risk assessment.
[0076] It should be noted that the error metric is a combination of root mean square error and accuracy, representing the quantitative result of the model's prediction error. It is used to compare the performance of different models and essentially provides an objective evaluation standard to guide model optimization.
[0077] It should be noted that the generation of error validation results including model confidence is based on the error index. The model confidence is calculated using statistical methods to obtain the error validation results. The specific calculation process of the confidence is as follows: First, the ratio of the root mean square error to the maximum allowable root mean square error is calculated. Then, this ratio is multiplied by the accuracy. Finally, the confidence value is obtained by subtracting the product from 1.
[0078] Furthermore, model confidence reflects the reliability of model predictions. High confidence indicates small errors and stable performance, which can be used for subsequent dynamic selection mechanisms.
[0079] It should be noted that the error validation results are structured outputs that include error indicators and model confidence, used for decision support. Essentially, they summarize evaluation information to determine the optimal model output and ensure the accuracy of correcting the wind speed field.
[0080] S5. Based on the error verification results, a confidence-driven dynamic selection mechanism is used to output the corrected maximum wind speed field.
[0081] In this embodiment of the invention, the step of outputting the corrected maximum wind speed field based on the error verification result using a confidence-driven dynamic selection mechanism includes: Based on the model confidence score in the error verification results, the optimal model is dynamically selected to obtain the selected model; Based on the selected model, the corrected maximum wind speed field is generated.
[0082] It should be noted that the dynamic selection of the optimal model is based on the model confidence in the error verification results. By comparing the confidence values of the random forest model and the DNN model, the model with the higher confidence is selected as the optimal model, thus obtaining the selected model.
[0083] Furthermore, the dynamic selection of the optimal model is used to prioritize high-confidence models based on the model's performance reliability on validation data. For example, when the confidence of the DNN model is higher than that of the random forest model, the DNN model is selected first to ensure the accuracy and stability of the output corrected wind speed and adapt to different wind speed conditions.
[0084] It should be noted that the selected model is the optimal model chosen by a dynamic selection mechanism, such as a random forest model or a DNN model, used to generate corrected wind speed values. Essentially, it optimizes the model output to minimize errors and improve the reliability of wind speed corrections.
[0085] It should be noted that the generation of the corrected maximum wind speed field is based on the selected model, performing a prediction operation on the input data, and outputting the spatially gridded maximum wind speed value to obtain the corrected maximum wind speed field.
[0086] Furthermore, the corrected maximum wind speed field is generated and the wind speed value of each grid point is calculated through model inference to form a wind speed field with a resolution of 1km. This field is used for wind speed risk assessment and operational decision-making in offshore wind farms. Essentially, it provides high-precision and reliable wind speed estimates to support energy management.
[0087] like Figure 2 The diagram shown is a functional block diagram of an ERA5 maximum wind speed hybrid machine learning correction system for offshore wind power provided in an embodiment of the present invention.
[0088] The ERA5 maximum wind speed hybrid machine learning correction system 100 for offshore wind power described in this invention can be installed in an electronic device. Depending on the functions implemented, the ERA5 maximum wind speed hybrid machine learning correction system 100 for offshore wind power may include a feature engineering module 101, a random forest pre-training module 102, a DNN fine-tuning correction training module 103, a model validation module 104, and a dynamic integration output module 105. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.
[0089] In this embodiment, the functions of each module / unit are as follows: The feature engineering module is used to create marine dynamic coupling features and time-series lag features based on the input ERA5 initial field, marine environmental data and topographic data, to obtain a feature dataset; The random forest pre-training module is used to pre-train the feature dataset using the random forest algorithm, and to filter key features to output wind speed interval predictions in probability form. The DNN fine correction training module is used to apply an oversampling strategy to the typhoon data in the feature dataset, and to perform fine training on the feature dataset and the wind speed interval prediction to obtain the corrected wind speed value. The model verification module is used to segment a verification dataset from the feature dataset, perform error verification on the corrected wind speed value based on the verification dataset, and generate an error verification result. The dynamic integrated output module is used to output the corrected maximum wind speed field based on the error verification results and using a confidence-driven dynamic selection mechanism.
[0090] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.
[0091] The modules described as separate components may or may not be physically separate. The components shown as modules 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.
[0092] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.
[0093] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.
[0094] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.
[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.
Claims
1. A hybrid machine learning method for ERA5 maximum wind speed correction in offshore wind power, characterized in that, The method includes: S1. Based on the input ERA5 initial field, marine environmental data and topographic data, marine dynamic coupling features and temporal lag features are created to obtain the feature dataset; S2, Use the random forest algorithm to pre-train the feature dataset, and filter key features to output wind speed interval prediction in probability form; S3, apply an oversampling strategy to the typhoon data in the feature dataset, perform fine training on the feature dataset and the wind speed interval prediction, and obtain the corrected wind speed value; S4, segment the verification dataset from the feature dataset, perform error verification on the corrected wind speed value based on the verification dataset, and generate error verification results; S5. Based on the error verification results, a confidence-driven dynamic selection mechanism is used to output the corrected maximum wind speed field.
2. The ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power as described in claim 1, characterized in that, The ocean dynamic coupling features and time-series lag features are created based on the input ERA5 initial field, marine environmental data, and topographic data, resulting in a feature dataset including: Based on the input ERA5 initial field, marine environmental data, and topographic data, feature engineering functions are applied to calculate wind-wave coupling characteristics to obtain wind-wave coupling characteristic values. Based on the input ERA5 initial field, marine environmental data, and topographic data, feature engineering functions are applied to calculate the coastal effect characteristics to obtain the coastal effect characteristic values. Based on the input ERA5 initial field, marine environmental data and topographic data, temporal lag features are extracted using the sliding window method to obtain a temporal lag feature set; Based on the wind-wave coupling feature value, the coastal effect feature value, and the time lag feature set, feature combination is performed to obtain a 15-dimensional feature dataset.
3. The ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power as described in claim 2, characterized in that, The feature combination based on the wind-wave coupling feature value, the coastal effect feature value, and the time-series lag feature set is used to obtain a 15-dimensional feature dataset, including: The wind and wave coupling feature value, the coastal effect feature value, and the time lag feature set are concatenated according to their dimensions to form a 15-dimensional feature vector, resulting in a 15-dimensional feature dataset. In the 15-dimensional feature vector, each row corresponds to a sample point, and each column corresponds to a feature dimension.
4. The ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power as described in claim 1, characterized in that, The step of pre-training the feature dataset using the random forest algorithm and filtering key features to output probability-based wind speed interval predictions includes: Based on the aforementioned feature dataset, the top 8 key features are selected using Gini importance ranking to obtain the key feature set; Based on the aforementioned key feature set, a random forest model with 500 decision trees is trained to obtain a pre-trained model; Based on the pre-trained model, the input data is predicted, and the wind speed range prediction is output in probabilistic form.
5. The ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power as described in claim 4, characterized in that, The process of training a random forest model with 500 decision trees based on the key feature set to obtain a pre-trained model includes: Multiple training subsets are generated through autopilot sampling, and a decision tree is trained in each subset. Gini impurity is used as the node splitting criterion. The best features and split points are selected recursively to maximize information gain and build the tree structure until the preset minimum number of leaf nodes or maximum depth is reached. Pre-trained models are obtained by integrating predictions through averaging.
6. The ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power as described in claim 1, characterized in that, The step of applying an oversampling strategy to the typhoon data in the feature dataset, and performing fine-tuning on the feature dataset and the wind speed interval prediction to obtain corrected wind speed values includes: Based on the feature dataset and the wind speed interval prediction, an oversampling strategy is applied to the typhoon data in the feature dataset to obtain an enhanced training dataset. Based on the enhanced training dataset, a 12-hour time series window is processed through an LSTM layer to obtain a time series feature representation; Based on the temporal feature representation, regression training is performed using a Dense layer, and the Huber loss function is used to optimize the wind speed interval prediction to obtain the corrected wind speed value.
7. The ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power as described in claim 6, characterized in that, The mathematical expression for the Huber loss function is as follows: ; In the formula, This represents Huber's loss value. This represents the actual wind speed value. To predict wind speed values, This is the threshold parameter.
8. The ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power as described in claim 1, characterized in that, The step of segmenting a validation dataset from the feature dataset, performing error verification on the corrected wind speed value based on the validation dataset, and generating an error verification result includes: Based on the aforementioned feature dataset, an independent validation dataset is segmented using a random segmentation method to obtain the validation dataset; Based on the validation dataset and the corrected wind speed value, the root mean square error and accuracy index are applied to evaluate the error and obtain the error index. Based on the error metric, an error verification result including model confidence is generated.
9. The ERA5 maximum wind speed hybrid machine learning correction method for offshore wind power as described in claim 1, characterized in that, Based on the error verification results, a confidence-driven dynamic selection mechanism is used to output the corrected maximum wind speed field, including: Based on the model confidence score in the error verification results, the optimal model is dynamically selected to obtain the selected model; Based on the selected model, the corrected maximum wind speed field is generated.
10. A hybrid machine learning correction system for maximum wind speed in ERA5 for offshore wind power, used to implement the hybrid machine learning correction method for maximum wind speed in ERA5 for offshore wind power as described in claims 1-9, characterized in that, The system includes: The feature engineering module is used to create marine dynamic coupling features and time-series lag features based on the input ERA5 initial field, marine environmental data and topographic data, to obtain the feature dataset; The random forest pre-training module is used to pre-train the feature dataset using the random forest algorithm, and to filter key features to output wind speed interval predictions in probability form. The DNN fine correction training module is used to apply an oversampling strategy to the typhoon data in the feature dataset, and to perform fine training on the feature dataset and the wind speed interval prediction to obtain the corrected wind speed value. The model validation module is used to segment a validation dataset from the feature dataset, perform error validation on the corrected wind speed value based on the validation dataset, and generate error validation results. The dynamic integrated output module is used to output the corrected maximum wind speed field based on the error verification results and using a confidence-driven dynamic selection mechanism.