Method and system for determining wind power in different weathers, storage medium and equipment

Abstract

The embodiment of the invention discloses a method for determining wind power in different weathers. The method comprises the following steps: collecting historical weather data and carrying out clustering classification on the historical weather data to obtain wind profile index data, and carrying out data enhancement on the wind profile index data to obtain a first data set; determining a plurality of wind profile indexes according to the first data set, and determining a corresponding wind profile index under an extreme weather condition according to a nonlinear relation model of the plurality of wind profile indexes and meteorological parameters; determining a dynamic prediction model of the wind profile index according to the corresponding wind profile index under the extreme weather condition by using a bidirectional long-short-term memory network; collecting ground wind speed data in real time, and determining the wind speed of the hub height of the wind turbine generator according to the dynamic prediction model of the wind profile index; determining power output curves at different wind speeds according to the corresponding relation between the historical ground wind speed data and the historical wind turbine generator power output data; and determining the wind power corresponding to the wind speed of the hub height of the wind turbine generator according to the power output curve.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, system, storage medium and device for determining wind power in different weather conditions. Background Art

[0002] As the world's dependence on renewable energy increases, wind power, as an important component of clean energy, the accuracy of its power forecast is crucial to the stable operation and dispatch optimization of the power system. However, under extreme weather conditions, the frequent occurrence of natural phenomena such as strong storms, heat waves, and cold fronts has brought unprecedented challenges to the accurate prediction of wind power; the wind profile index, as a key indicator to measure the changing characteristics of wind speed at different altitudes, shows great instability in extreme climate events. These extreme weather events not only lead to drastic changes in wind speed, but are also accompanied by a significant increase in turbulence intensity and significant changes in atmospheric stability. These factors work together on the wind profile index, making it difficult for traditional prediction models based on historical data to accurately capture its dynamic change characteristics.

[0003] Most existing wind power prediction models rely on historical data under normal weather conditions and predict future wind power through statistical analysis or physical model simulation. However, when faced with extreme weather, the scarcity and brevity of such weather events result in extremely limited sample data, making it difficult to construct a data set that fully reflects the characteristics of extreme weather. In addition, the complex physical mechanisms of extreme weather far exceed the descriptive capabilities of traditional physical models, greatly reducing the model's prediction accuracy under extreme conditions.

[0004] In order to improve the prediction accuracy of wind power under extreme weather conditions, there is an urgent need for a new method that can fully consider the special changes in wind profile characteristics under extreme meteorological conditions, integrate conventional meteorological data and extreme meteorological data of extreme weather, extract key meteorological features through advanced machine learning technology, and accurately predict ramp events and extreme power based on these features. Summary of the invention

[0005] Based on this, it is necessary to propose a method for determining wind power in different weather conditions to address the above issues.

[0006] A method for determining wind power in different weather conditions, the method comprising the following steps:

[0007] Collect historical weather data and perform cluster classification on the data to obtain wind profile index data, and perform data enhancement on the wind profile index data to obtain a first data set;

[0008] Determine a plurality of wind profile indexes according to the first data set, and determine a corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indexes and meteorological parameters;

[0009] Using a bidirectional long short-term memory network, a dynamic prediction model of a wind profile index is determined according to the corresponding wind profile index under the extreme weather conditions;

[0010] Collecting ground wind speed data in real time, and determining the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index;

[0011] According to the correspondence between historical ground wind speed data and historical wind turbine power output data, the power output curve under different wind speeds is determined;

[0012] The wind power corresponding to the wind speed at the hub height of the wind turbine generator set is determined according to the power output curves under different wind speeds.

[0013] In the above scheme, the collecting of historical weather data and clustering and classifying the data to obtain wind profile index data, and performing data enhancement on the wind profile index data to obtain a first data set specifically includes:

[0014] The historical weather data is cleaned and missing values ​​and outliers are removed, and the processed data is normalized using the Z-score standardization method, and the normalized data is classified using the K-means clustering algorithm to obtain a sample library of wind profile indices corresponding to severe storms, heat waves, and cold fronts;

[0015] The data in the wind profile index sample library is used as training samples and input into a generative adversarial network to extract wind speed, temperature, and air pressure, and through correlation analysis and principal component analysis, the principal component whose explained variance accounts for more than a preset threshold is selected as an extreme weather attribute feature;

[0016] By using a generative adversarial network algorithm and adversarial learning between a generator and a discriminator, the data in the sample library that meets the characteristics of extreme weather attributes is expanded to obtain a first data set.

[0017] In the above scheme, the determining of a plurality of wind profile indices according to the first data set, and the determining of the corresponding wind profile indices under extreme weather conditions according to the nonlinear relationship model between the plurality of wind profile indices and meteorological parameters, specifically includes:

[0018] Obtain meteorological parameters of temperature gradient, wind speed gradient and turbulence intensity under extreme weather conditions;

[0019] Preprocessing the meteorological parameters, extracting key features from the meteorological parameters, and determining a meteorological parameter feature vector;

[0020] The random forest algorithm is used to train the meteorological

[0021] Nonlinear relationship model between parameters and wind profile index;

[0022] The nonlinear relationship model between the trained meteorological parameters and the wind profile index is used to predict the newly collected meteorological parameter data under extreme weather conditions to obtain the corresponding wind profile index.

[0023] In the above scheme, the dynamic prediction model of the wind profile index is determined according to the corresponding wind profile index under the extreme weather conditions by using the bidirectional long short-term memory network, specifically including:

[0024] Determine the extreme weather time series characteristic dataset based on the wind profile index, temperature gradient, wind speed gradient and turbulence intensity of extreme weather;

[0025] Training a bidirectional long short-term memory network according to the extreme weather time series feature data set until the training of the bidirectional long short-term memory network is completed;

[0026] Performing feature importance analysis on the extreme weather time series feature data set based on the trained bidirectional long short-term memory network, evaluating the impact classification results of each meteorological element feature on the wind profile index prediction result, and optimizing the trained bidirectional long short-term memory network based on the impact classification results;

[0027] The real-time collected temperature gradient, wind speed gradient and turbulence intensity are input into the optimized bidirectional long short-term memory network to predict the wind profile index corresponding to the current data.

[0028] In the above scheme, the real-time collection of ground wind speed data and the determination of the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index specifically include:

[0029] Obtain surface wind speed data through real-time weather stations;

[0030] Using a power law model or a logarithmic rate distribution model, respectively obtaining a first wind speed profile evaluation result and a second wind speed profile evaluation result;

[0031] The wind speed at the hub height of the wind turbine is determined by taking a weighted average of the first wind speed profile evaluation result and the second wind speed profile evaluation result.

[0032] In the above scheme, the power output curves under different wind speeds are determined according to the correspondence between the historical ground wind speed data and the historical wind turbine power output data, specifically including:

[0033] According to the preset time period, real-time collection of historical ground wind speed data and historical wind turbine power output data;

[0034] Preprocessing the collected historical ground wind speed data and the collected historical wind turbine power output data respectively;

[0035] The historical wind turbine power is grouped according to wind speed intervals, the mean, standard deviation, skewness, and kurtosis of the historical wind turbine power in each wind speed interval are determined, and the distribution characteristics of the historical wind turbine power are analyzed;

[0036] The mean value of the historical wind turbine power output data in each interval is taken as the dependent variable, and the least square method is used to perform polynomial curve fitting to obtain the power characteristic curve equation of the wind turbine in the entire wind speed range.

[0037] In the above scheme, the power output curves under different wind speeds are determined according to the correspondence between the historical ground wind speed data and the historical wind turbine power output data. After that, the method further includes:

[0038] Set different meteorological parameter threshold conditions to determine whether the current moment is in extreme weather conditions;

[0039] If the weather conditions are extreme, the theoretical power is integrated with the result of the dynamic prediction model of the wind profile index to correct the power prediction value;

[0040] If there are no extreme weather conditions, the theoretical power will be directly output as the prediction result.

[0041] The present application also proposes a wind power determination system for different weather conditions, characterized in that the system comprises: a data acquisition unit, a wind speed prediction unit, a power curve fitting unit and a wind power output unit;

[0042] The data acquisition unit is used to collect historical weather data and cluster and classify the data to obtain wind profile index data, perform data enhancement on the wind profile index data, and obtain a first data set; determine a plurality of wind profile indices according to the first data set, and determine the corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indices and meteorological parameters;

[0043] The wind speed prediction unit is used to use a bidirectional long short-term memory network to determine a dynamic prediction model of a wind profile index according to the corresponding wind profile index under the extreme weather conditions; collect ground wind speed data in real time, and determine the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index;

[0044] The power curve fitting unit is used to determine the power output curve under different wind speeds according to the corresponding relationship between the historical ground wind speed data and the historical wind turbine power output data;

[0045] The wind power output unit is used to determine the wind power corresponding to the wind speed at the hub height of the wind turbine according to the power output curves under different wind speeds.

[0046] The present application also proposes a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the following steps:

[0047] Collect historical weather data and perform cluster classification on the data to obtain wind profile index data, and perform data enhancement on the wind profile index data to obtain a first data set;

[0048] Determine a plurality of wind profile indexes according to the first data set, and determine a corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indexes and meteorological parameters;

[0049] Using a bidirectional long short-term memory network, a dynamic prediction model of a wind profile index is determined according to the corresponding wind profile index under the extreme weather conditions;

[0050] Collecting ground wind speed data in real time, and determining the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index;

[0051] According to the correspondence between historical ground wind speed data and historical wind turbine power output data, the power output curve under different wind speeds is determined;

[0052] The wind power corresponding to the wind speed at the hub height of the wind turbine generator set is determined according to the power output curves under different wind speeds.

[0053] The present application also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor in the following steps:

[0054] Collect historical weather data and perform cluster classification on the data to obtain wind profile index data, and perform data enhancement on the wind profile index data to obtain a first data set;

[0055] Determine a plurality of wind profile indexes according to the first data set, and determine a corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indexes and meteorological parameters;

[0056] Using a bidirectional long short-term memory network, a dynamic prediction model of a wind profile index is determined according to the corresponding wind profile index under the extreme weather conditions;

[0057] Collecting ground wind speed data in real time, and determining the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index;

[0058] According to the correspondence between historical ground wind speed data and historical wind turbine power output data, the power output curve under different wind speeds is determined;

[0059] The wind power corresponding to the wind speed at the hub height of the wind turbine generator set is determined according to the power output curves under different wind speeds.

[0060] The embodiment of the present invention has the following beneficial effects: firstly, historical weather data is collected and clustered to obtain wind profile index data, and data enhancement is performed on the wind profile index data to obtain a first data set; a plurality of wind profile indices are determined according to the first data set, and the corresponding wind profile indices under extreme weather conditions are determined according to the nonlinear relationship models between the plurality of wind profile indices and meteorological parameters; a dynamic prediction model of the wind profile index is determined according to the corresponding wind profile indices under extreme weather conditions using a bidirectional long short-term memory network; ground wind speed data is collected in real time, and the wind profile index is determined according to the dynamic prediction model of the wind profile index. The wind speed at the hub height of the turbine set; according to the correspondence between the historical ground wind speed data and the historical wind turbine power output data, the power output curve under different wind speeds is determined; according to the power output curve, the wind power corresponding to the wind speed at the hub height of the wind turbine set is determined; this method collects and clusters historical weather data, combined with data enhancement of the wind profile index, to construct a wind power prediction framework suitable for different weather conditions. The framework uses a nonlinear relationship model to determine the wind profile index under extreme weather conditions, and establishes a dynamic prediction model of the wind profile index through a bidirectional long short-term memory network (BiLSTM). By collecting ground wind speed data in real time and combining it with a dynamic prediction model, the wind speed at the hub height of the wind turbine set can be accurately predicted, and then the corresponding wind power can be determined according to the correspondence between historical wind speed and power. This method significantly improves the accuracy of wind power prediction, especially the prediction ability under extreme weather conditions, and provides strong support for the operation optimization and scheduling decision-making of wind farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] in:

[0063] Figure 1 The figure is a flow chart of a method for determining wind power in different weather conditions in one embodiment. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0065] The purpose of the terms used herein is only to describe specific embodiments and is not intended to be limiting of the present invention. When used herein, the singular forms "a", "an" and "said / the" are also intended to include the plural forms, unless the context clearly indicates otherwise, and it should also be understood that the terms "consisting of" and / or "comprising", when used in this specification, determine the presence of the features, integers, steps, operations, elements and / or parts, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, parts and / or groups. When used herein, the term "and / or" includes any and all combinations of the relevant listed items.

[0066] Alternative embodiments of the present invention are described in detail below, however, in addition to these detailed descriptions, the present invention may also have other embodiments.

[0067] like Figure 1 As shown, in one embodiment, a method for determining wind power in different weather conditions is provided, and the method for determining wind power in different weather conditions includes steps S101 to S106, which are described in detail as follows:

[0068] S101, collecting historical weather data and clustering and classifying the data to obtain wind profile index data, performing data enhancement on the wind profile index data to obtain a first data set;

[0069] By collecting historical weather data and performing clustering and classification, we can more effectively identify the impact of different weather types on wind power. The application of data enhancement technology helps to increase the generalization ability of the model and improve its adaptability to different weather conditions. Among them, wind profile index data is an important parameter for evaluating the change of wind speed with height. Through in-depth analysis and utilization of it, this solution can more accurately reflect the change of wind speed in the vertical direction, thereby improving the accuracy of wind power prediction.

[0070] In some embodiments, collecting historical weather data and clustering and classifying the historical weather data, obtaining wind profile index data, and performing data enhancement on the wind profile index data to obtain a first data set specifically includes:

[0071] The historical weather data was cleaned and missing values ​​and outliers were removed. The processed data were normalized using the Z-score standardization method. The normalized data were classified using the K-means clustering algorithm to obtain a sample library of wind profile indices corresponding to severe storms, heat waves, and cold fronts.

[0072] The data in the wind profile index sample library is used as training samples and input into the generative adversarial network to extract wind speed, temperature, and air pressure. Through correlation analysis and principal component analysis, the principal component with an explained variance ratio exceeding the preset threshold is selected as the extreme weather attribute feature;

[0073] Using the generative adversarial network algorithm, through adversarial learning between the generator and the discriminator, the data in the sample library that meets the characteristics of extreme weather attributes is expanded to obtain the first data set.

[0074] Specifically, historical weather data are preprocessed to remove missing values ​​and outliers, and the Z-score standardization method is used to normalize the data. Then, the K-means clustering algorithm is used to classify the data. The number of clusters k is set to 5, and the clustering effect is evaluated by calculating the silhouette coefficient and SSE (sum of squared errors) to obtain the optimal clustering result. A wind profile index sample library of extreme weather types such as severe storms, heat waves, and cold fronts is formed. Feature engineering is performed on the data in the sample library to extract 20 key attributes such as wind speed, temperature, and air pressure. Through correlation analysis and principal component analysis (PCA) and other methods, the top 10 principal components with an explained variance of more than 80% are selected as extreme weather attribute features.

[0075] In some embodiments, by generating an adversarial network algorithm, data in a wind profile index sample library of different extreme weather types are used as training samples. Through adversarial learning between the generator and the discriminator, the amount of extreme weather wind profile index data approaching different extreme weathers can be increased to enrich the training samples.

[0076] Specifically, the data in the extreme weather wind profile index sample library are preprocessed, including data cleaning, outlier processing and normalization, and key features such as wind speed gradient and wind direction change rate are extracted through feature engineering as input for generative adversarial network (GAN) training.

[0077] Preferably, a GAN algorithm is adopted, and the processed wind profile index data is used as a training sample. Through the adversarial learning of the generator and the discriminator, more realistic extreme weather wind profile index data are generated to enrich the training samples. According to the data distribution characteristics of the extreme weather wind profile index sample library, the network structure of the generator is designed, and a multi-layer convolutional neural network (CNN) is adopted as the main network of the generator. Through operations such as upsampling and transposed convolution, the random noise vector and the conditional vector are mapped to wind profile index data similar to the real sample. The conditional vector contains attribute information such as extreme weather type and wind speed range. By splicing it with the noise vector, conditional generation is achieved.

[0078] Furthermore, a discriminator network is constructed, which also adopts the CNN structure. Through downsampling and convolution operations, the input wind profile index data is mapped to a scalar value, which represents the probability that the input data is a real sample. During the training process, the generator and the discriminator conduct adversarial learning through the minimum and maximum game method. The generator generates realistic wind profile index data as much as possible to deceive the discriminator; while the discriminator distinguishes between real samples and generated samples as accurately as possible.

[0079] Among them, the WassersteinGAN (WGAN) training framework is used, the discriminator outputs the approximate value of the Wasserstein distance, and the generator loss function is the average output of the discriminator for the generated samples. At the same time, the gradient penalty (GradientPenalty) regularization method is introduced to prevent the gradient from disappearing or exploding. The Adam optimization algorithm is adopted, and appropriate hyperparameters such as learning rate and batch size are set. The generator and discriminator are trained alternately until a equilibrium state is reached.

[0080] Finally, the quality of the generated wind profile index data is evaluated, and indicators such as InceptionScore (IS) and FréchetInceptionDistance (FID) are used to quantify the authenticity and diversity of the generated data. The IS indicator measures the quality of the generated data by comparing the KL divergence of the conditional probability distribution and the marginal probability distribution of the generated samples on the pre-trained Inception network. The FID indicator evaluates the authenticity of the generated data by calculating the Fréchet distance between the generated samples and the real samples in the feature space of the intermediate layer of the Inception network. Combined with manual judgment and expert review, high-quality generated samples are screened out and added to the original training set to construct an enhanced extreme weather wind profile index sample library. The enhanced sample library is used to train the extreme weather prediction model. The performance of the models trained with the original training set and the enhanced training set on the test set is compared by cross-validation and other methods to evaluate the effectiveness of the generated data.

[0081] In some embodiments, the data in the extreme weather wind profile index sample library is first preprocessed: the Z-score standardization method is used to normalize the wind speed, wind direction and other features, and the data is scaled to a range of a mean of 0 and a standard deviation of 1. The wind speed time series is then decomposed at multiple scales using wavelet transform, and the wind speed gradient features at different frequency scales are extracted. The wind direction change rate is calculated to obtain a 20-dimensional key feature vector. Next, the DCGAN (Deep Convolutional GAN) architecture is used to design the generator and discriminator networks. The generator uses a 4-layer transposed convolutional network, and the number of output channels of each layer is 512, 256, 128, and 1, respectively. The transposed convolution The kernel size is 4x4, the stride is 2, and BatchNorm and ReLU activation functions are used after the first three layers. The discriminator uses a 4-layer convolutional network with 64, 128, 256, and 1 output channels in each layer. The convolution kernel size is 3x3 and the stride is 2. BatchNorm and LeakyReLU activation functions are used after the first three layers. The last layer uses the Sigmoid function to output probability values. During the training process, the WGAN-GP loss function is used. The discriminator loss includes the difference between the Wasserstein distance estimates of the real samples and the generated samples, as well as the gradient penalty term with a penalty coefficient of 10. The generator loss is the average output of the discriminator for the generated samples.

[0082] The Adam optimization algorithm was used, with a learning rate of 0.0002, Beta1 of 0.5, a batch size of 64, and 200 epochs of training. In each epoch, the discriminator was updated 5 times and the generator was updated 1 time. During the training process, a batch of samples was generated every 50 epochs, and the IS and FID indicators were calculated. The IS indicator was calculated using the pre-trained Inception-v3 network, and the FID indicator was extracted using the pre-trained ResNet-50 network. Finally, the generated samples with IS indicators greater than 4.0 and FID indicators less than 0.1 were selected and added to the original training set to obtain the enhanced extreme weather wind profile index sample library, which was used to train and generate extreme weather prediction models based on the LSTM algorithm. Through 5-fold cross validation, the mean absolute error (MAE) and root mean square error (RMSE) of the models before and after enhancement on the wind speed prediction task were compared. It was found that the MAE and RMSE of the enhanced model were reduced by 0.5m / s and 0.8m / s, respectively, which proved the effectiveness of data enhancement.

[0083] S102, determining a plurality of wind profile indices according to the first data set, and determining a corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indices and meteorological parameters;

[0084] By constructing a nonlinear relationship model between the wind profile index and meteorological parameters and determining the wind profile index under extreme weather conditions based on it, this scheme provides a scientific basis for wind power prediction under extreme weather conditions, which is of great significance for improving the operating stability and safety of wind farms under extreme conditions.

[0085] In some embodiments, determining a plurality of wind profile indices according to the first data set, and determining a corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indices and meteorological parameters, specifically includes:

[0086] Obtain meteorological parameters of temperature gradient, wind speed gradient and turbulence intensity under extreme weather conditions;

[0087] Preprocess the meteorological parameters, extract key features from the meteorological parameters, and determine the meteorological parameter feature vectors;

[0088] The random forest algorithm is used to train the nonlinear relationship model between meteorological parameters and wind profile index with the meteorological parameter feature vector as input;

[0089] The nonlinear relationship model between the trained meteorological parameters and the wind profile index is used to predict the newly collected meteorological parameter data under extreme weather conditions to obtain the corresponding wind profile index.

[0090] Specifically, meteorological parameters such as temperature gradient, wind speed gradient and turbulence intensity under extreme weather conditions can be obtained through multi-source data collection channels such as meteorological observation stations, meteorological radars, and meteorological satellites to form an extreme weather meteorological parameter data set. Extreme weather conditions can be determined based on the thresholds of meteorological elements, such as wind speed exceeding 30m / s, temperature below -20℃ or above 40℃, etc. There is a physical correlation between various meteorological parameters. For example, the temperature gradient affects the atmospheric stability, which in turn affects the wind speed and turbulence intensity.

[0091] Preferably, the collected meteorological parameter data are preprocessed, including data cleaning, outlier processing, missing value filling, etc., to ensure the integrity and accuracy of the data. Outlier processing can use the 3σ principle or box plot method, and missing value filling can use mean filling or K nearest neighbor filling, etc., and then the data format conversion and standardization processing, such as Z-score standardization or Min-Max normalization, are performed to facilitate subsequent feature extraction and model training. According to the change characteristics of the wind profile index under extreme weather conditions, the key features in the meteorological parameter data are extracted, such as performing wavelet transform on the temperature and wind speed data to extract the change characteristics at different scales.

[0092] The turbulence intensity data is then Fourier transformed to extract frequency domain features, etc. The feature combination that best reflects the changing law of the wind profile index is selected to construct a meteorological parameter feature vector. The random forest algorithm is used with the meteorological parameter feature vector as input and the wind profile index as output to train and establish a nonlinear relationship model between meteorological parameters and wind profile index. The random forest constructs and integrates multiple decision trees through bootstrap sampling and random feature selection, and has good nonlinear fitting ability and generalization performance. During the training process, cross-validation and other methods are used, and optimization algorithms such as grid search or random search are used to find the optimal combination of model hyperparameters, such as the number of decision trees, the maximum depth of the tree, the minimum number of samples for node splitting, etc., to improve the fitting ability and generalization performance of the model. The trained random forest model is used to predict the newly collected meteorological parameter data under extreme weather conditions to obtain the corresponding wind profile index estimate.

[0093] At the same time, through feature importance analysis, such as average impurity reduction or permutation test, the key meteorological parameters affecting the estimation of the wind profile index are determined to provide guidance for subsequent extreme weather warnings and prevention and control. The estimated value of the wind profile index is compared with the actual observed value, and the estimation error and evaluation indicators, such as the mean absolute error, root mean square error, determination coefficient, etc., are calculated to evaluate the estimation performance of the random forest model. The mean square error and determination coefficient of the model are calculated using cross-validation, and the bias and variance of the model are analyzed using learning curves to further optimize the model. If the estimation error exceeds the predetermined threshold, it is necessary to further adjust the feature selection, optimize the model structure, or increase the training samples, etc., and continuously iterate and improve to improve the accuracy and robustness of the wind profile index estimation.

[0094] Specifically, first, the temperature, wind speed and turbulence intensity data collected every minute by the meteorological observation station, the reflectivity data scanned every 5 minutes by the meteorological radar, and the cloud map data updated every 10 minutes by the meteorological satellite are used to comprehensively judge whether the extreme weather conditions are met. For example, when the temperature is lower than -25℃ or higher than 42℃ for 3 consecutive hours, the wind speed exceeds 35m / s for 10 consecutive minutes, and the turbulence intensity EDR index exceeds 0.5 for 5 consecutive minutes, it is judged as extreme weather. Then, an outlier detection algorithm, such as IsolationForest or One-ClassSVM, is used to identify and remove abnormal data points. For missing values, the nearest neighbor interpolation method is used to fill them. Next, the meteorological element data is Min-Max normalized to scale the value range to the [0, 1] interval. Then, the temperature and wind speed data are wavelet transformed, and the Daubechies4 wavelet basis function is selected to decompose them into 5 scales to extract each scale. The statistical characteristics of wavelet coefficients such as mean, variance, skewness and kurtosis under the scale were calculated. The turbulence intensity data were fast Fourier transformed to extract the spectrum energy characteristics of three frequency bands: 0-0.1Hz, 0.1-1Hz and 1-10Hz. Finally, a 15-dimensional meteorological parameter feature vector was obtained. The random forest algorithm was used for modeling. Each decision tree randomly selected 80% of the samples from the training set using the Bootstrap method. At each node, 30% of all the features were randomly selected for optimal splitting. 500 decision trees were generated for ensemble prediction. The grid search method was used to optimize the random forest hyperparameters. The search space included the maximum tree depth [10, 20, 30], the minimum number of node samples [2, 5, 10], and the minimum number of leaf node samples [1, 2, 4]. The optimal parameter combination was determined by 5-fold cross validation. The model performance was evaluated on the test set. The mean absolute error (MAE) was 0.15, the root mean square error (RMSE) was 0.22, and the determination coefficient (R 2 ) is 0.85.

[0095] S103, using a bidirectional long short-term memory network, determining a dynamic prediction model of a wind profile index according to corresponding wind profile indexes under extreme weather conditions;

[0096] Using a bidirectional long short-term memory network (BiLSTM) as a dynamic prediction model can make full use of the time series characteristics of wind power data, while capturing relevant information at previous and subsequent time points to improve the accuracy of the prediction; in addition, BiLSTM also has a certain degree of adaptive ability and can adapt to changes in wind speed under different weather conditions.

[0097] A bidirectional long short-term memory network is used, with the wind profile index, temperature gradient, wind speed gradient and turbulence intensity of extreme weather as input, to capture the time series characteristics of extreme weather and establish a dynamic prediction model for the wind profile index of extreme weather, including: according to the definition and standard of extreme weather, extracting historical data of extreme weather events from the meteorological database, including time series data of multiple meteorological elements such as wind profile index, temperature gradient, wind speed gradient and turbulence intensity, to form an extreme weather time series feature data set, preprocessing the extreme weather time series feature data, using linear interpolation, spline interpolation and other methods to fill missing values, using Z-score or box plot and other methods to detect and filter outliers, and normalizing the data, scaling the characteristics of each meteorological element to the same scale range, dividing the processed time series data into time windows of fixed length, each time window containing meteorological element observations of multiple time steps as the input sequence of the model.

[0098] A bidirectional long short-term memory network model was built, and the input layer, forward LSTM layer, backward LSTM layer, fully connected layer and output layer were set. The input layer received the extreme weather time series feature data. The forward LSTM layer and the backward LSTM layer respectively contained multiple hidden units. Nonlinear activation functions such as ReLU were used, and the dropout regularization method was used to prevent overfitting. The hyperparameters such as the number of hidden units, the number of layers, the dropout rate, and the length of the time window of the LSTM layer were optimized by grid search and other methods. The model structure and parameter combination with the best performance on the validation set were selected. The existing extreme weather time series feature data were used to train the bidirectional long short-term memory network model by supervised learning, and the mean square error was used as the loss function.

[0099] The Adam optimization algorithm is used to iteratively update the model parameters. By setting an appropriate batch size and number of iterations, the convergence speed and generalization performance of the model are controlled. The training process is monitored. According to the loss function value and evaluation indicators on the validation set, the model structure and hyperparameters are adjusted in time to avoid overfitting or underfitting. The cross-validation method is used to evaluate the prediction performance of the bidirectional long short-term memory network model. The extreme weather time series feature data are divided into training set, validation set and test set. The root mean square error, mean absolute error, mean relative error and determination coefficient between the predicted value and the true value of the wind profile index are calculated on the test set to comprehensively judge the prediction accuracy and robustness of the model. The trained bidirectional long short-term memory network model is used for feature importance analysis. Global sensitivity analysis methods such as variance decomposition and Sobol index are used to quantitatively evaluate the influence of various meteorological element characteristics on the wind profile index prediction results and the intensity of interaction, revealing the key driving factors and internal laws of changes in extreme weather wind profile index, and providing decision-making basis for extreme weather warning and prevention and control. The optimized bidirectional long short-term memory network model is applied to the real-time dynamic prediction of extreme weather wind profile index. The temperature gradient, wind speed gradient, turbulence intensity and other data are collected in real time through meteorological observation equipment, and input into the model for prediction, so as to obtain the change trend of wind profile index in different time scales such as 1 hour, 3 hours and 6 hours in the future.

[0100] In some embodiments, a dynamic prediction model of a wind profile index is determined based on the corresponding wind profile index under extreme weather conditions using a bidirectional long short-term memory network, specifically including:

[0101] Determine the extreme weather time series characteristic dataset based on the wind profile index, temperature gradient, wind speed gradient and turbulence intensity of extreme weather;

[0102] Train the bidirectional long short-term memory network according to the extreme weather time series feature data set until the training of the bidirectional long short-term memory network is completed;

[0103] The feature importance analysis of the extreme weather time series feature data set was performed based on the trained bidirectional long short-term memory network, the impact classification results of each meteorological element feature on the wind profile index prediction results were evaluated, and the trained bidirectional long short-term memory network was optimized based on the impact classification results;

[0104] The real-time collected temperature gradient, wind speed gradient and turbulence intensity are input into the optimized bidirectional long short-term memory network to predict the wind profile index corresponding to the current data.

[0105] Specifically, first, the extreme weather event data from 2000 to 2020 were extracted from the meteorological database, including time series data such as wind profile index, temperature gradient, wind speed gradient and turbulence intensity every minute, totaling 10 million records. The Pearson correlation coefficient and Spearman rank correlation coefficient were used to analyze the correlation between various meteorological elements, and the redundant variables with correlation coefficients less than 0.1 were eliminated. Then, the local outlier factor (LOF) algorithm was used to detect outliers, and the observation records with LOF scores greater than 2 were marked as outliers. The KNN interpolation method was used to correct the outliers, and the data were Mi n-Max normalization processing is performed to scale the value range to the interval [0, 1]. Then, a sliding time window is used to slice the time series data. Each window contains 60 time steps, that is, 1 hour of observation data. The sliding step length between adjacent windows is 10 minutes. A bidirectional LSTM model is built with an input layer dimension of 4, corresponding to 4 meteorological elements. The forward and backward LSTM layers each contain 2 layers, each with 128 hidden units. The tanh activation function and a dropout rate of 0.2 are used. The fully connected layer contains 256 neurons, using the ReLU activation function. The output layer uses a linear activation function to predict the wind profile index value at the next moment.

[0106] The mean square error (MSE) loss function and Adam optimizer were used. The initial learning rate was 0.001, and the learning rate was decayed to 0.1 times the original value every 10 epochs. The batch size was 256, and 100 epochs were trained. The model performance was evaluated by 5-fold cross validation. The mean square error MSE on the test set was 0.02, the mean absolute error MAE was 0.05, the mean absolute percentage error MAPE was 2%, and the determination coefficient R 2 The Sobol method is used to analyze the global sensitivity of the model, and it is found that the Sobol first-order exponents of temperature gradient and wind speed gradient are 0.35 and 0.28 respectively, which have the greatest impact on the prediction results of wind profile index. The trained model can also be used to predict the newly collected meteorological data. When the predicted value of wind profile index in the next hour exceeds 1.5, the orange warning is automatically triggered, and the warning information is issued to the public through mobile phone text messages and TV scrolling subtitles.

[0107] S104, collecting ground wind speed data in real time, and determining the wind speed at the hub height of the wind turbine according to a dynamic prediction model of the wind profile index;

[0108] In some embodiments, the ground wind speed data is collected in real time, and the wind speed at the hub height of the wind turbine is determined according to the dynamic prediction model of the wind profile index, specifically including:

[0109] Obtain surface wind speed data through real-time weather stations;

[0110] Using a power law model or a logarithmic rate distribution model, respectively obtaining a first wind speed profile evaluation result and a second wind speed profile evaluation result;

[0111] The wind speed at the hub height of the wind turbine is determined by taking a weighted average of the first wind speed profile evaluation result and the second wind speed profile evaluation result.

[0112] Specifically, 10 real-time weather stations are deployed in the wind farm, each of which is equipped with 4 wind speed sensors, which are installed at heights of 10 meters, 30 meters, 50 meters and 70 meters respectively. The sampling frequency is set to once per minute, the wind speed measurement range is 0-70m / s, and the accuracy is ±0.3m / s. The wind speed sensor is connected to the data collector through the RS485 bus. The data collector packages the wind speed data and timestamp, longitude and latitude information and uploads them to the time series database (such as InfluxDB) of the wind farm every minute. The database is indexed according to the weather station number and timestamp to facilitate subsequent data query and analysis. The original wind speed data of each weather station is cleaned by MapReduce, and invalid data exceeding the physical threshold of 0-70m / s is eliminated. Then, the Tukey algorithm is used to detect outliers in the wind speed data, that is, data exceeding 1.5 times the difference between the first quartile and the third quartile is regarded as an outlier. The linear interpolation method before and after is used for correction to remove the influence of the outlier. According to the relative position relationship between the hub height of the wind turbine and each weather station.

[0113] The inverse distance weighted interpolation method (IDW) is used to calculate the measured wind speed at a height of 80 meters where the wind turbine is located, as well as the predicted value of the extreme weather wind profile index at the corresponding moment, and it is used as a reference for the subsequent wind speed profile model estimation. The wind speed at a height of 10 meters is used as the benchmark value, and the wind speed at a height of 80 meters is estimated by the power law model. When the extreme weather wind profile index is less than 0.2, the power index is 0.14, and when the wind profile index is greater than 0.6, the power index is 0.22. In other cases, the power index is calculated by the linear interpolation formula of 0.14+(0.22-0.14)×(wind profile index-0.2) / (0.6-0.2). At the same time, the logarithmic rate distribution model is used to estimate the wind speed at a height of 80 meters. The surface roughness length and zero plane displacement in the model parameters are obtained according to the surface classification information table, and are corrected in combination with the measured wind speed profile data.

[0114] Using the fuzzy logic reasoning method, the power law model estimation value and the logarithmic rate distribution model estimation value are weighted averaged according to the extreme weather wind profile index. When the index is less than 0.2, the power law model weight is 0.8 and the logarithmic law model weight is 0.2. When the index is greater than 0.6, the power law model weight is 0.2 and the logarithmic law model weight is 0.8. In other cases, the weight is calculated by the membership function. The wind speed estimation result at a height of 80 meters is compared with the rated wind speed of 12m / s and the cut-out wind speed of 25m / s of the wind turbine. When the estimated wind speed exceeds the cut-out wind speed for 3 consecutive minutes, a shutdown command is sent to the EMS system of the wind farm control center. When the estimated wind speed is lower than the rated wind speed for 10 consecutive minutes, a power reduction command is sent to the EMS system. The EMS system executes the command through the unit SCADA system to adjust the blade angle and generator torque, and at the same time updates the meteorological parameters of the wind power forecasting system, and corrects the results of the rolling forecast power for the next 4 hours for the day-ahead and real-time scheduling of wind power.

[0115] S105, determining power output curves under different wind speeds according to the correspondence between historical ground wind speed data and historical wind turbine power output data;

[0116] The power output curve under different wind speeds is determined by the corresponding relationship of historical data, which provides a reliable basis for the prediction of wind power. This method based on actual data can more accurately reflect the actual operating conditions of wind turbines and improve the reliability of prediction results.

[0117] Based on the wind speed data under extreme weather conditions, the power characteristic curve of the wind turbine is verified and adjusted. Representative wind turbines are selected in the wind farm, and wind speed sensors and power measurement equipment are installed at different heights of the turbine tower to collect real-time power output data of the wind turbines under different wind speed conditions. According to the meteorological conditions and wind resource characteristics of the area where the wind farm is located, the time period for data collection is reasonably set. For example, seasons with large wind speed changes are selected for key data collection to obtain operating data within the full wind speed range. The data sampling frequency is set to 1 minute, and data collection is continued for more than 3 months to form a wind speed-power data set. The collected wind speed-power data is preprocessed to eliminate data under abnormal conditions such as the start-up and shutdown process of the wind turbine and fault maintenance, and to extract effective steady-state operation data.

[0118] The moving average filtering method is used to remove high-frequency oscillation noise in wind speed and power data and improve the smoothness of the data. The power data is grouped according to the wind speed range. The wind speed range is selected as 0.5m / s, namely 0-0.5m / s, 0.5-1m / s, 1-1.5m / s, etc. The power data in each wind speed range is statistically analyzed, and the mean, standard deviation, skewness, kurtosis and other statistical quantities of the power data in each wind speed range are calculated. The distribution characteristics of the power data are analyzed, and outliers exceeding plus or minus 3 times the standard deviation are removed to improve the data quality. The median of the wind speed range is used as the independent variable, and the mean of the power data in the range is used as the dependent variable. The least squares method is used to perform polynomial curve fitting to obtain the power characteristic curve equation of the wind turbine in the entire wind speed range. The root mean square error (RMSE) and determination coefficient (R 2 ), evaluate the fitting accuracy, if RMSE exceeds 50kW or R 2 If it is lower than 0.9, try to increase the polynomial order or use piecewise function fitting to improve the curve fitting effect, obtain the historical wind speed data under extreme weather conditions (such as typhoons, cold waves, etc.) in the area where the wind farm is located, and extract the actual power output data of the wind turbines in the corresponding period.

[0119] The actual power output data is compared with the theoretical output power of the power characteristic curve under the same wind speed conditions, and the degree of deviation between the actual output power and the theoretical value is calculated. When the deviation exceeds 20% and lasts for more than 1 hour, it is judged that the power generation efficiency of the wind turbine has decreased under extreme weather conditions, and the correction procedure of the power characteristic curve is triggered, including multiplying the correction coefficient by the correction coefficient on the basis of the original characteristic curve according to the deviation ratio between the actual output power and the theoretical value. When the wind speed is less than the cut-in wind speed, the correction coefficient is 0; when the wind speed is greater than the cut-in wind speed and less than the rated wind speed of 12m / s, the correction coefficient is the ratio of the actual output power to the theoretical value; when the wind speed is greater than the rated wind speed and less than the cut-out wind speed of 25m / s, the correction coefficient decays linearly with the wind speed, and when it decays to the cut-out wind speed, the correction coefficient is the correction coefficient under the rated wind speed. The correction coefficient is 50% of the coefficient; when the wind speed is greater than the cut-out wind speed, the correction coefficient keeps the value at the cut-out wind speed unchanged, and the corrected wind turbine power characteristic curve is embedded in the wind turbine intelligent monitoring system to calculate the theoretical output power in real time, compare it with the actual power and analyze the deviation. When the deviation exceeds 20% for 5 consecutive minutes and parameters such as wind speed and pitch angle are normal, an alarm message is sent to the operator on duty to prompt that the unit may have hidden dangers of failure and further diagnosis is required. At the same time, the corrected power characteristic curve can also be applied to the short-term wind power prediction system. The system adopts a long short-term memory (LSTM) neural network model, combined with the future wind speed data of the meteorological forecast and the corrected power characteristic curve, to make a rolling prediction of the wind power output in the next 24 hours, and update the prediction result every 15 minutes.

[0120] Specifically, 10 representative 2.0MW wind turbines were selected in the wind farm. A wind speed sensor was installed at the hub center and blade tip of each turbine, with a measuring range of 0-70m / s and an accuracy of ±0.3m / s. A power measurement device was installed at the front end of the nacelle, with a measuring range of 0-2500kW and an accuracy of ±0.5%. The data collection period was set to 1 year, focusing on the typhoon-prone season from July to September and the cold wave-prone season from December to February. The sampling frequency was 1 minute. The collected raw data was cleaned to remove invalid data caused by turbine start-up and shutdown, maintenance, and measurement values ​​exceeding the physical threshold. The wind speed and power series were smoothed using the Holt-Winters exponential smoothing method, with smoothing parameters α and β taking 0.2 and 0.1, and the wind speed data were divided into 140 intervals at intervals of 0.5m / s.

[0121] The mean, standard deviation, skewness and kurtosis of the power data in each interval are calculated, and the upper and lower bounds of the outliers of the power data in each interval are calculated. The data points beyond the upper and lower bounds are removed, and the median of each wind speed interval and the corresponding power mean are used as sample points. The least squares method is used to fit the 9th order polynomial to obtain the power characteristic curve of the wind turbine. The RMSE is 23.5kW, and R 2 The data of wind speed and wind turbine power in the corresponding period are extracted. The actual output power in typhoon weather is 25% lower than the theoretical value on average, and the actual output power in extremely cold weather is 18% lower than the theoretical value on average. When the deviation lasts for more than 1.5 hours, the characteristic curve correction program is started, with the rated wind speed of 12m / s as the boundary. When the wind speed is lower than 12m / s, the correction coefficient is the ratio of the actual output power to the theoretical value. When the wind speed is higher than 12m / s, the correction coefficient decays linearly from the ratio at the rated wind speed to 50% at the cut-out wind speed of 25m / s. The corrected characteristic curve is embedded in the unit SCADA system, and the actual power is compared with the theoretical power at a period of 5 minutes. When the deviation exceeds 20% and lasts for a period of 15 minutes, the correction coefficient is used as the correction factor. When the wind speed observation values ​​and actual power of 30 consecutive days are used as samples, 70% are randomly selected as training sets and 30% as test sets. The LSTM model contains 2 hidden layers, 64 neurons in each layer, a learning rate of 0.001, an optimization algorithm Adam, and a loss function mean square error. After 100 Epochs of training, the average relative error of the power prediction of the test set is reduced to less than 8%.

[0122] In some embodiments, according to the correspondence between historical ground wind speed data and historical wind turbine power output data, determining the power output curves at different wind speeds specifically includes:

[0123] According to the preset time period, real-time collection of historical ground wind speed data and historical wind turbine power output data;

[0124] Pre-process the collected historical ground wind speed data and historical wind turbine power output data respectively;

[0125] The historical wind turbine power is grouped according to the wind speed range, the mean, standard deviation, skewness and kurtosis of the historical wind turbine power in each wind speed range are determined, and the distribution characteristics of the historical wind turbine power are analyzed;

[0126] The mean value of historical wind turbine power output data in each interval is taken as the dependent variable, and the least squares method is used to perform polynomial curve fitting to obtain the power characteristic curve equation of the wind turbine in the entire wind speed range.

[0127] In some embodiments, according to the correspondence between historical ground wind speed data and historical wind turbine power output data, power output curves under different wind speeds are determined, and then the method further includes:

[0128] Set different meteorological parameter threshold conditions to determine whether the current moment is in extreme weather conditions;

[0129] If the weather conditions are extreme, the theoretical power is integrated with the results of the dynamic prediction model of the wind profile index to correct the power prediction value;

[0130] If there are no extreme weather conditions, the theoretical power will be directly output as the prediction result.

[0131] Determine whether the current moment is in extreme weather conditions. If so, merge the theoretical power with the results of the extreme weather wind profile index dynamic prediction model to correct the power prediction value; if not, directly output the theoretical power as the prediction result. Set different meteorological parameter threshold conditions according to the extreme weather level and impact time in the meteorological warning information. For example, a red typhoon warning corresponds to a 10-minute average wind speed ≥32.7m / s and a duration ≥2 hours, and an orange cold wave warning corresponds to a daily minimum temperature ≤-12℃ and a drop of ≥10℃ within 48 hours. If any threshold condition is met, it is determined that the current moment is in extreme weather conditions. If the current moment is determined to be non-extreme weather conditions, the theoretical power prediction value of the wind farm is directly output as the final power prediction result, written into the wind power prediction system database, and the power prediction curve and deviation statistical indicators are displayed in real time through the Grafana visualization platform.

[0132] If the current moment is judged to be an extreme weather condition, the wind speed, temperature, and pressure profile forecast data for the area in the next 4 hours are obtained from the reanalysis database of the European Centre for Medium-Term Weather Forecasts (ECMWF), and interpolated with the measured data of the automatic weather station to generate a 5km×5km gridded data set within the wind farm. The data set is input into the dynamic prediction model of the extreme weather wind profile index. The model uses a multivariate long short-term memory (MV-LSTM) neural network algorithm to extract the spatiotemporal evolution characteristics of the wind profile index under extreme weather conditions by training historical data for a total of 5 years from 2015 to 2020. The model input features include hourly wind speed profiles, temperature profiles, and pressure profiles, as well as the wind profile of the previous moment. The wind profile index is outputted in 15-minute increments over the next 4 hours. The model uses the root mean square error (RMSE) as the loss function, the Adam optimizer, the learning rate is 0.001, the batch size is 64, and 100 epochs are trained. The RMSE on the test set is less than 0.2, which meets the business accuracy requirements. The theoretical power forecast value of the wind farm under extreme weather conditions is weightedly fused with the correction coefficient output by the dynamic prediction model of the wind profile index. The correction coefficient is dynamically adjusted according to the deviation between the wind profile index and the rated value. The greater the deviation, the higher the weight. The sigmoid function is used to map the wind profile index deviation to between 0 and 1 to obtain the correction coefficient weight. The greater the weight, the greater the impact of extreme weather conditions on the wind farm power.

[0133] When the wind profile index returns to the normal range (such as 2.5-4.0), the correction coefficient weight drops to 0, and the fusion formula is corrected power = theoretical power × (1-weight) + theoretical power × correction coefficient × weight, that is, the theoretical power is mainly used under normal weather conditions, and the corrected power is mainly used under extreme weather conditions. Once three consecutive 10-minute time windows meet the extreme weather conditions criteria, the extreme weather wind profile index dynamic prediction model and power prediction correction program are automatically activated. At the same time, extreme weather warning prompts are sent to wind farm operators, production technology departments, and power grid dispatching departments via SMS, emails, and pop-up windows. The warning information includes extreme weather type, level, start and end time, impact range, and expected power deviation, etc. It is updated every 15 minutes, and the recipient is required to confirm the receipt within 10 minutes after receiving the information. If the extreme weather conditions last for more than 2 hours, the warning information will be automatically pushed to the wind farm production management system, triggering the emergency plan process until the extreme weather conditions are lifted and the warning signal is automatically revoked.

[0134] Specifically, 10 Vaisala WXT530 automatic weather stations are deployed in the wind farm. Each is equipped with wind speed, wind direction, temperature, air pressure, humidity and precipitation sensors. The measurement ranges are 0-60m / s, 0-360°, -52-60℃, 500-1100hPa, 0-100%RH and 0-200mm / h, respectively. The accuracy is ±3%, ±3°, ±0.3℃, ±0.5hPa, ±3%RH and ±5%, respectively. The data is uploaded to the wind farm SCADA system through the RS485 bus every minute, and communicated with the Meteorological Bureau through the IEC104 protocol. The MICAPS warning release system is connected to parse the extreme weather warning message in XML format, extract key fields such as disaster type, level, scope and effective time, and use Pandas to perform moving median filtering on the wind speed and temperature series for the original data collected by the meteorological station. The window size is 10, the median = (X+X) / 2, and the outliers that exceed the physical threshold are cleaned. The upper limit of wind speed is 60m / s, the lower limit is 0.4m / s, the upper limit of temperature is 45℃, and the lower limit is -40℃. The statistical indicators are aggregated into 10-minute granularity, including average wind speed, standard deviation, maximum gust, average temperature, etc.

[0135] Determine whether extreme weather conditions are triggered, such as orange typhoon warning (10-minute average wind speed ≥ 24.5m / s), cold wave blue warning (48-hour minimum temperature ≤ 4℃), etc. If so, send a POST request to the API interface of the European Center for Medium-Term Weather Forecasts ECMWF to obtain a 0.125°×0.125° gridded dataset for the next 4 hours, which contains wind speed, temperature and pressure profiles (100-1000hPa). Perform interpolation, clipping and merging operations on the dataset to generate a coverage wind power The 5km×5km gridded profile data of the wind field is reshaped into a four-dimensional tensor of [4, 12, 5, 5] as the input of the MV-LSTM model, with a hidden layer dimension of 64 and a loop step size of 4. The wind profile index sequence for the next 4 hours is outputted in 16 time steps of 15 minutes each. The four dimensions represent the future predicted time step (4 hours is decomposed into 15-minute intervals, a total of 16 steps), the meteorological parameter category in each time step (which may include multiple dimensions such as wind speed, temperature, and air pressure), and the spatial grid. The grid is divided into 5km×5km, and the Sigmoid function is used to map the wind speed index to generate a correction coefficient. The Sigmoid transformation can convert continuous index values ​​into probability-like weights in the interval (0, 1). By calculating (index-3.0) / 0.5 and then applying the Sigmoid function, the impact degree of different wind speed segments can be effectively distinguished. Combining this correction coefficient with the theoretical power curve of the wind farm, the power generation efficiency curve adjusted according to the real-time and predicted meteorological conditions can be obtained. When the index deviation exceeds 1.0 and lasts for 3 time steps, the warning information is pushed to 5 on-duty mobile phone numbers, 3 department mailboxes, and the large screen of the dispatch center through the Twilio SMS gateway, SMTP mail server, and Alertmanager alarm management tool. The content includes extreme weather type, warning level (such as orange), impact time (such as the next 2-4 hours), expected theoretical power deviation (such as -20%), etc., and the warning event in JSON format is POSTed to the Redis message queue.

[0136] S106. Determine the wind power corresponding to the wind speed at the hub height of the wind turbine generator according to the power output curves under different wind speeds.

[0137] By combining numerical weather forecast models (such as WRF) and machine learning models (such as support vector machines), the wind speed forecast data at the hub height of each wind turbine in the wind farm in the next 24 hours is obtained, with a forecast time resolution of 15 minutes. The numerical weather forecast model simulates the evolution of large-scale weather systems by solving atmospheric motion equations; by training historical meteorological data and wind power operation data, a nonlinear mapping relationship for wind speed prediction is established, and the predicted data is written into the MySQL database of the wind farm power prediction system to provide input for subsequent theoretical power calculations. According to the number and model of the wind turbine, the power characteristic curve parameters of the corresponding unit are queried from the MongoDB database of the wind farm asset management system, including the order and coefficient of the polynomial curve, and the start and end wind speeds of the curve. The power characteristic curve of the wind turbine is obtained by collecting the historical operation data of the unit through the SCADA system, and is regularly fitted and optimized using machine learning algorithms such as neural networks to improve the accuracy and adaptability of the curve.

[0138] For each hub height wind speed prediction value at a future moment, substitute it into the polynomial equation of the wind turbine power characteristic curve, and use the Horner algorithm to efficiently calculate the value of the polynomial to obtain the theoretical power output prediction value of the wind turbine at that moment. If the wind speed prediction value is less than the cut-in wind speed of the wind turbine or greater than the cut-out wind speed, the theoretical power output is set to 0, indicating that the wind turbine is in a shutdown state.

[0139] The real-time power output data of each wind turbine is collected through the wind farm SCADA system with a data granularity of 1 minute. The data is pushed to the Kafka message queue in real time through the OPCUA protocol, and Spark Streaming is used to consume the real-time data stream in Kafka. Data cleaning and feature engineering modules are developed in Spark Streaming to perform preprocessing operations such as outlier filtering and missing value interpolation on the real-time power data, and to extract meteorological features related to power such as wind speed, wind direction, and temperature to form a wind power sample data set. The Isolation Forest algorithm is used to analyze the wind power sample data. Perform anomaly detection, identify and filter abnormal samples, improve data quality, merge the real-time wind power sample data set with the historical wind power sample data set, build an incremental wind power machine learning model, use the online gradient boosting tree algorithm, set an adaptive weight update strategy, balance the weight distribution of new and old samples, realize real-time update of model parameters, introduce the forgetting mechanism, reduce the weight of historical samples through the time decay function, adapt to the dynamic changes of wind conditions, develop an online learning module for the wind profile index model, and trigger the model parameter update process when the wind farm SCADA system collects new wind speed profile data.

[0140] The online sequential extreme learning machine algorithm is adopted, with minimizing the root mean square error (RMSE) as the objective function, and the weight matrix of the MV-LSTM model is dynamically adjusted to realize the adaptive optimization of wind profile index prediction. The wind power sample data set and the output of the wind profile index model are used as input features. The online kernel extreme learning machine algorithm is adopted. Through kernel function mapping and random feature mapping, the wind profile index model and power prediction model parameters are updated in real time using Redis cache for real-time call by the wind farm power prediction system. The control chart method is used to monitor the prediction deviation in real time. When three consecutive sample points exceed the upper and lower control limits, the abnormal diagnosis process is triggered to analyze the cause of the deviation, and the model parameters are dynamically adjusted or alarm prompts are given.

[0141] Specifically, the wind farm's SCADA system collects real-time power data of wind turbines every minute. Through the OPCUA protocol, these data are written to a Kafka topic named "power". Subsequently, the SparkStreaming program subscribes to the topic, sets a window length of 1 minute and slides every 30 seconds to process these real-time power information. The Python Pandas library is used to clean the power data in each time window to ensure that the power value falls within the reasonable physical range [0, rated power]. If an abnormal value exceeds this range, it is replaced by the mean of its two adjacent normal values. Then, the IsolationForest algorithm is used to detect anomalies in power data based on three features: wind speed, wind direction, and temperature. The sample size is set to 1000, of which 5% are abnormal samples, and the threshold for determining anomalies is set to 0.6 to further improve the data quality. On this basis, the River library is used to implement the online gradient boosting tree model, with 50 trees configured and a learning rate of 0.1. The model learns incrementally every 10 minutes based on the data of the past 3 hours. During the learning process, the weights of new and old samples decay exponentially. The attenuation coefficient is 0.9 to achieve dynamic update of the model. When the wind speed profile data is received by the SCADA system and written to InfluxDB, the update process of the dynamic sequential least squares machine learning model is triggered. The model is configured with 100 hidden layer nodes, a Sigmoid activation function, a learning rate of 0.01, a weight of 0.2 for new samples, and a weight of 0.8 for old samples to reflect the importance of the latest data. The real-time wind power data and the wind profile index prediction value are merged and stored in the list structure of Redis, and the online kernel least squares machine model is used for power prediction. , select the Gaussian kernel function, set the width to 0.5, the regularization coefficient to 0.1, and the new sample to a weight of 0.3. The model parameters are updated every 30 minutes. The comparison between the predicted value and the measured value is plotted as a line graph through PlotlyDash, and the Shewhart control chart is used to monitor the deviation between the two. The control limit is set to ±2.5 times the standard deviation. Once four consecutive data points cross the limit, the Alertmanager alarm is triggered to Grafana, and the JenkinsPipeline is started to retrain and deploy the prediction model and refresh the Redis cache.

[0142] In order to evaluate the performance of the prediction service, Locust was used to perform a stress test, simulating 1,000 users online at the same time, sending 10 requests per second, and the test duration was 1 hour. The goal was to keep the average response time no more than 200 milliseconds, the P95 response time less than 1 second, and the overall service availability to 99.9%. With the help of Python's Arrow library, a wind power forecast sequence for every 15 minutes in the next 24 hours is automatically generated every day. These data are written to the MySQL database to provide decision support for the wind farm dispatch system. Combined with the daily forecast electricity price and real-time electricity price information, the output of wind turbines and the automatic generation control (AGC) adjustment capabilities are optimized to maximize wind energy benefits. Finally, when the control chart monitors that the prediction deviation exceeds ±3 times the standard deviation for six consecutive points, the wind turbine status will be automatically marked as "abnormal" and the troubleshooting task will be dispatched to professional engineers through the operation and maintenance platform. The goal is to shorten the average fault diagnosis time from 4 hours to within 1 hour, effectively improving the fault response efficiency.

[0143] In summary, this application adopts the random forest algorithm to establish a nonlinear relationship model between meteorological parameters and wind profile index, so as to realize the estimation of wind profile index under extreme weather conditions; by adopting a bidirectional long short-term memory network, combining the wind profile index, temperature gradient, wind speed gradient and turbulence intensity of extreme weather as input, a dynamic prediction model of extreme weather wind profile index is established to capture the temporal characteristics of extreme weather; in short, the power output of wind turbines under extreme weather conditions can be accurately estimated and predicted to improve the operating efficiency and safety of wind farms, and can effectively improve the accuracy of wind power prediction under extreme weather conditions and the energy efficiency management of wind farms, thereby improving the economic benefits and system stability of wind farms.

[0144] The present application also proposes a wind power determination system for different weather conditions, the system comprising: a data acquisition unit, a wind speed prediction unit, a power curve fitting unit and a wind power output unit;

[0145] A data acquisition unit is used to collect historical weather data and cluster and classify the data, obtain wind profile index data, perform data enhancement on the wind profile index data, and obtain a first data set; determine a plurality of wind profile indices according to the first data set, and determine the corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indices and meteorological parameters;

[0146] The wind speed prediction unit is used to determine the dynamic prediction model of the wind profile index according to the corresponding wind profile index under extreme weather conditions by using a bidirectional long short-term memory network; collect ground wind speed data in real time, and determine the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index;

[0147] A power curve fitting unit is used to determine the power output curve under different wind speeds according to the corresponding relationship between the historical ground wind speed data and the historical wind turbine power output data;

[0148] The wind power output unit is used to determine the wind power corresponding to the wind speed at the hub height of the wind turbine according to the power output curves under different wind speeds.

[0149] The present application also proposes a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the following steps:

[0150] Collect historical weather data and perform cluster classification on the data to obtain wind profile index data, perform data enhancement on the wind profile index data, and obtain a first data set;

[0151] Determine a plurality of wind profile indexes according to the first data set, and determine corresponding wind profile indexes under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indexes and meteorological parameters;

[0152] Using a bidirectional long short-term memory network, a dynamic prediction model of the wind profile index is determined according to the corresponding wind profile index under extreme weather conditions;

[0153] Collect ground wind speed data in real time and determine the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index;

[0154] According to the correspondence between historical ground wind speed data and historical wind turbine power output data, the power output curve under different wind speeds is determined;

[0155] The wind power corresponding to the wind speed at the hub height of the wind turbine is determined according to the power output curves under different wind speeds.

[0156] The present application also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor in the following steps:

[0157] Collect historical weather data and perform cluster classification on the data to obtain wind profile index data, perform data enhancement on the wind profile index data, and obtain a first data set;

[0158] Determine a plurality of wind profile indexes according to the first data set, and determine corresponding wind profile indexes under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indexes and meteorological parameters;

[0159] Using a bidirectional long short-term memory network, a dynamic prediction model of the wind profile index is determined according to the corresponding wind profile index under extreme weather conditions;

[0160] Collect ground wind speed data in real time and determine the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index;

[0161] According to the correspondence between historical ground wind speed data and historical wind turbine power output data, the power output curve under different wind speeds is determined;

[0162] The wind power corresponding to the wind speed at the hub height of the wind turbine is determined according to the power output curves under different wind speeds.

[0163] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0164] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0165] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present application. It should be pointed out that, for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. The above disclosure is only a preferred embodiment of the present invention, and of course it cannot be used to limit the scope of the rights of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope covered by the present invention.

Claims

1. A method for determining wind power in different weather conditions, characterized in that: The method comprises: Collect historical weather data and perform cluster classification on the data to obtain wind profile index data, and perform data enhancement on the wind profile index data to obtain a first data set; Determine a plurality of wind profile indexes according to the first data set, and determine a corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indexes and meteorological parameters; Using a bidirectional long short-term memory network, a dynamic prediction model of a wind profile index is determined according to the corresponding wind profile index under the extreme weather conditions; Collecting ground wind speed data in real time, and determining the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index; According to the correspondence between historical ground wind speed data and historical wind turbine power output data, the power output curve under different wind speeds is determined; The wind power corresponding to the wind speed at the hub height of the wind turbine generator set is determined according to the power output curves under different wind speeds.

2. The method for determining wind power in different weather conditions according to claim 1, characterized in that: The collecting of historical weather data and clustering and classifying the data to obtain wind profile index data, and performing data enhancement on the wind profile index data to obtain a first data set specifically includes: The historical weather data is cleaned and missing values ​​and outliers are removed, and the processed data is normalized using the Z-score standardization method, and the normalized data is classified using the K-means clustering algorithm to obtain a sample library of wind profile indices corresponding to severe storms, heat waves, and cold fronts; The data in the wind profile index sample library is used as training samples and input into a generative adversarial network to extract wind speed, temperature, and air pressure, and through correlation analysis and principal component analysis, the principal component whose explained variance accounts for more than a preset threshold is selected as an extreme weather attribute feature; By using a generative adversarial network algorithm and adversarial learning between a generator and a discriminator, the data in the sample library that meets the characteristics of extreme weather attributes is expanded to obtain a first data set.

3. The method for determining wind power in different weather conditions according to claim 2, characterized in that: The determining of a plurality of wind profile indices according to the first data set, and determining a corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indices and meteorological parameters, specifically includes: Obtain meteorological parameters of temperature gradient, wind speed gradient and turbulence intensity under extreme weather conditions; Preprocessing the meteorological parameters, extracting key features from the meteorological parameters, and determining a meteorological parameter feature vector; A random forest algorithm is used to train a nonlinear relationship model between the meteorological parameter and the wind profile index by taking the meteorological parameter feature vector as input; The nonlinear relationship model between the trained meteorological parameters and the wind profile index is used to predict the newly collected meteorological parameter data under extreme weather conditions to obtain the corresponding wind profile index.

4. The method for determining wind power in different weather conditions according to claim 3, characterized in that: The method of using a bidirectional long short-term memory network to determine a dynamic prediction model of a wind profile index according to the corresponding wind profile index under the extreme weather conditions specifically includes: Determine the extreme weather time series characteristic dataset based on the wind profile index, temperature gradient, wind speed gradient and turbulence intensity of extreme weather; Training a bidirectional long short-term memory network according to the extreme weather time series feature data set until the training of the bidirectional long short-term memory network is completed; Performing feature importance analysis on the extreme weather time series feature data set based on the trained bidirectional long short-term memory network, evaluating the impact classification results of each meteorological element feature on the wind profile index prediction result, and optimizing the trained bidirectional long short-term memory network based on the impact classification results; The real-time collected temperature gradient, wind speed gradient and turbulence intensity are input into the optimized bidirectional long short-term memory network to predict the wind profile index corresponding to the current data.

5. The method for determining wind power in different weather conditions according to claim 4, characterized in that: The real-time collection of ground wind speed data and determination of the wind speed at the hub height of the wind turbine generator set according to the dynamic prediction model of the wind profile index specifically include: Obtain surface wind speed data through real-time weather stations; Using a power law model or a logarithmic rate distribution model, respectively obtaining a first wind speed profile evaluation result and a second wind speed profile evaluation result; The wind speed at the hub height of the wind turbine is determined by taking a weighted average of the first wind speed profile evaluation result and the second wind speed profile evaluation result.

6. The method for determining wind power in different weather conditions according to claim 5, characterized in that: Determining the power output curves at different wind speeds based on the correspondence between the historical ground wind speed data and the historical wind turbine power output data specifically includes: According to the preset time period, real-time collection of historical ground wind speed data and historical wind turbine power output data; Preprocessing the collected historical ground wind speed data and the collected historical wind turbine power output data respectively; The historical wind turbine power is grouped according to wind speed intervals, the mean, standard deviation, skewness, and kurtosis of the historical wind turbine power in each wind speed interval are determined, and the distribution characteristics of the historical wind turbine power are analyzed; The mean value of the historical wind turbine power output data in each interval is taken as the dependent variable, and the least square method is used to perform polynomial curve fitting to obtain the power characteristic curve equation of the wind turbine in the entire wind speed range.

7. The method for determining wind power in different weather conditions according to claim 6, characterized in that: According to the correspondence between the historical ground wind speed data and the historical wind turbine power output data, the power output curves under different wind speeds are determined, and then the method further comprises: Set different meteorological parameter threshold conditions to determine whether the current moment is in extreme weather conditions; If the weather conditions are extreme, the theoretical power is integrated with the result of the dynamic prediction model of the wind profile index to correct the power prediction value; If there are no extreme weather conditions, the theoretical power will be directly output as the prediction result.

8. A system for determining wind power in different weather conditions, characterized in that: The system comprises: a data acquisition unit, a wind speed prediction unit, a power curve fitting unit and a wind power output unit; The data acquisition unit is used to collect historical weather data and cluster and classify the data to obtain wind profile index data, perform data enhancement on the wind profile index data, and obtain a first data set; determine a plurality of wind profile indices according to the first data set, and determine the corresponding wind profile index under extreme weather conditions according to a nonlinear relationship model between the plurality of wind profile indices and meteorological parameters; The wind speed prediction unit is used to use a bidirectional long short-term memory network to determine a dynamic prediction model of a wind profile index according to the corresponding wind profile index under the extreme weather conditions; collect ground wind speed data in real time, and determine the wind speed at the hub height of the wind turbine according to the dynamic prediction model of the wind profile index; The power curve fitting unit is used to determine the power output curve under different wind speeds according to the corresponding relationship between the historical ground wind speed data and the historical wind turbine power output data; The wind power output unit is used to determine the wind power corresponding to the wind speed at the hub height of the wind turbine according to the power output curves under different wind speeds.

9. A readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

Images