Ultra-short-term wind power prediction method based on digital twinning

Abstract

The invention provides an ultra-short-term wind power prediction method based on digital twinning, belongs to the technical field of wind power prediction, and establishes an ultra-short-term wind power prediction model based on digital twinning aiming at the problem that wind power prediction does not consider physical characteristics of a fan and meteorological characteristics of a wind power plant completely. According to the model, the reliability of prediction is improved by constructing a digital physical model of a wind power plant, and high-precision power prediction is realized by using digital twinning. A physical information neural network fused with a fan physical model is introduced into the model to process a data set, a digital twinborn prediction model comprising a digital physical model and a parallel Ctransform-BiGRU model is constructed, and spatial-temporal characteristics of data are captured by using a deep learning module. The prediction process is coupled with the actual physical scene of the wind power plant by using a digital physical model, so that the prediction precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power prediction, in particular to an ultra-short-term wind power prediction method based on digital twinning. BACKGROUND

[0002] Under the background of the "double carbon" goal, new energy power generation has gradually become the main body of the new power system. Wind power generation is a renewable energy technology that converts wind energy into electrical energy. Wind energy, as a renewable and clean energy source, has great development space and application prospects. Due to the high randomness and large fluctuation of wind energy, the operating state of the wind turbine is greatly affected by weather factors such as wind speed, temperature, and rainfall. Therefore, wind power is unstable and uncertain, and large-scale wind power integration will affect the stability of the power system. Therefore, accurate prediction of wind power can effectively improve the stability of the power system and reduce the loss of new energy power generation.

[0003] A large number of research work has been carried out at home and abroad for wind power prediction technology. The existing power prediction methods include physical methods, mathematical statistics methods, and artificial intelligence methods. The physical method is to construct a physical model based on the working principle of the wind turbine, and to predict based on the wind power physical formula and the physical information parameters of the station. This method requires a high level of relevant physical professional knowledge, and the predictor needs to be proficient in the professional knowledge of multiple physical disciplines such as electricity and mechanics. Moreover, due to the high complexity of the prediction model established by the physical method, the model establishment and analysis process can become very complex. A prediction method based on computational fluid dynamics and wind energy extraction delay detection is proposed in the prior art, which constructs a wind field extrapolation model of the wind farm to realize short-term wind power prediction. In order to solve the problem of poor wind power prediction accuracy under extreme weather conditions, a probability prediction method based on physical information reinforcement learning is proposed in the prior art, which establishes an analytical physical expression model of wind power output under extreme events. The mathematical statistics method uses a large amount of historical data to extract data features for prediction by using statistical principles such as autoregressive analysis, exponential smoothing, and moving average. A dynamic exponential time smoothing method is used in the prior art to predict wind power, which reduces the short-term wind power prediction error by constructing a quantile neural network model and dynamically updating parameters. A prediction method based on multiple outlier smoothing transformation autoregressive model is proposed in the prior art, which studies the multiple mechanisms of wind power time series fluctuation.

[0004] With the increase of new energy grid connection and data complexity type, the traditional method has been unable to meet the prediction scene of a large number of historical data and multi-source input data. The power prediction method based on artificial intelligence represented by deep learning has gradually become the main research method of power prediction. The main working principle of the power prediction method based on artificial intelligence is to use convolutional neural network (CNN), recurrent neural network (RNN), gate recurrent unit (GRU) and transformer combined with attention mechanism and other deep learning models to predict power based on a large number of historical power data and meteorological data. A multi-channel CNN-GRU wind power prediction model is proposed in the prior art, which uses CNN and GRU models to extract global and local features of data and time sequence features in historical power output. A multi-task learning prediction model based on Transformer-LSTM is proposed in the prior art, which combines the parallel processing and long dependence capturing ability of Transformer and the sequence processing ability of LSTM to predict wind power, and introduces multi-task learning to improve the accuracy of wind power prediction.

[0005] The above artificial intelligence method has achieved remarkable results in wind power prediction, but there are still some problems. On the one hand, these artificial intelligence methods are mainly based on historical power data and historical meteorological data for prediction, without considering the geographical features of the wind farm and the physical characteristics of the wind turbine itself, resulting in insufficient coupling of the model with the actual physical space and affecting the accuracy of the prediction results. On the other hand, the artificial intelligence method mainly uses deep learning and mathematical methods to process and predict data, lacks specific consideration of the identification of actual station meteorological change characteristics and the running state of the wind turbine, resulting in deviation of the prediction results.

[0006] In order to cope with these challenges, the present application proposes a super short-term wind power prediction method based on digital twinning to solve the above problems by establishing a wind power prediction system framework based on digital twinning. SUMMARY

[0007] Therefore, the purpose of the present application is to provide a super short-term wind power prediction method based on digital twinning to at least solve the above problems.

[0008] The technical solution adopted by the first aspect of the present application is as follows: The super short-term wind power prediction method based on digital twinning comprises the following steps: Constructing a wind power prediction system comprising a physical layer, a data processing layer, a digital twinning layer and a fusion output layer; The physical layer collects historical power data, meteorological data, wind turbine physical parameters and geographic information data of the wind farm. The wind turbine physical parameters include hub height, blade diameter, cut-in and cut-out wind speeds and rated power. In the data processing layer, the collected data is preprocessed, including filling missing values ​​with KNN interpolation and identifying and correcting power anomalies using the Physical Information Neural Network (PINN), wherein PINN introduces wind speed-power coupling relationship and Bates' law as physical constraints. The digital twin layer performs power prediction based on preprocessed data using both a digital physical model and a parallel Ctransformer-BiGRU deep learning model. The digital twin layer fuses the prediction results of the digital physical model and the parallel Ctransformer-BiGRU model through weighted fusion to obtain the final power prediction value; The prediction results are fused from the output layer and analyzed based on the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. Perform model performance evaluation and optimization.

[0009] The technical solution adopted in the second aspect of this invention is as follows: A digital twin-based wind power prediction system for implementing the method proposed in the first aspect includes: Data acquisition module: Collects historical power data, meteorological data, wind turbine physical parameters, and geographic information data of the wind farm; Data processing module: includes data cleaning submodule, feature fusion submodule and PINN correction submodule; Digital twin prediction module: includes a digital physical model submodule and a parallel Ctransformer-BiGRU model submodule; Fusion output module: used to weight and fuse the two types of prediction results, and output the final power prediction value and performance evaluation report.

[0010] Compared with the prior art, the beneficial effects of the present invention are: To address the challenge of improving the prediction accuracy of wind farm power generation, this invention proposes a digital twin wind power prediction model that combines a digital physical model and a Ctransformer-BiGRU deep learning model. This model integrates actual physical parameters and geographical information of the wind farm, incorporating wind power physics principles and the Bates agreement, thus enhancing its coupling capability with the real physical space and its adaptability to real-time meteorological changes at the farm. Furthermore, by combining the CNN-Transformer and BiGRU models, the local feature extraction capability of CNN and the long sequence modeling capability of Transformer are combined to achieve accurate wind power prediction, further enhanced by the TPE Bayesian hyperparameter optimization method.

[0011] This invention establishes a digital twin-based ultra-short-term wind power prediction framework to achieve high-accuracy wind power prediction. The invention uses KNN interpolation and a PINN model incorporating wind power physical information to process the data, significantly improving the quality of the input data and enhancing the model's connection to the real physical space. At the model level, this invention proposes an ultra-short-term wind power prediction model based on digital twins. This method effectively improves the prediction accuracy of wind power by fusing a digital physical model and a parallel Ctransformer-BiGRU model. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the overall process of an ultra-short-term wind power prediction method based on digital twins proposed in an embodiment of the present invention.

[0014] Figure 2 This is a schematic diagram of the overall structure of an ultra-short-term wind power prediction system based on digital twins proposed in another embodiment of the present invention.

[0015] Figure 3 This is a schematic diagram of the digital physical model structure of the ultra-short-term wind power prediction method based on digital twins proposed in an embodiment of the present invention.

[0016] Figure 4 This is a schematic diagram of the BiGRU structure of the ultra-short-term wind power prediction method based on digital twin proposed in an embodiment of the present invention.

[0017] Figure 5This is a schematic diagram of the TPE-Ctransformer-BiGRU model structure of the ultra-short-term wind power prediction method based on digital twin proposed in an embodiment of the present invention.

[0018] Figure 6 This is a heatmap diagram illustrating the correlation analysis in the computational analysis of the ultra-short-term wind power prediction method based on digital twins proposed in an embodiment of the present invention.

[0019] Figure 7 This is a schematic diagram of the model training loss function curve of the ultra-short-term wind power prediction method based on digital twin proposed in an embodiment of the present invention.

[0020] Figure 8 This is a schematic diagram of the RMSE search results for the hyperparameter 'a' in an ultra-short-term wind power prediction method based on digital twins proposed in an embodiment of the present invention.

[0021] Figure 9 This is a schematic diagram of the model power prediction curve of the ultra-short-term wind power prediction method based on digital twin proposed in an embodiment of the present invention.

[0022] Figure 10 is a schematic diagram of the prediction results of the ultra-short-term wind power prediction method based on digital twins proposed in an embodiment of the present invention during periods of strong wind (a) and no wind (b).

[0023] Figure 11 This is a schematic diagram of the ablation experiment power prediction curve of the ultra-short-term wind power prediction method based on digital twin proposed in an embodiment of the present invention.

[0024] Figure 12 This is a schematic diagram of the power prediction curve in a comparative experiment of the ultra-short-term wind power prediction method based on digital twin proposed in an embodiment of the present invention.

[0025] Figure 13 This is a comparative experimental performance diagram of an ultra-short-term wind power prediction method based on digital twins proposed in an embodiment of the present invention. Detailed Implementation

[0026] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0027] Reference Figure 1 An embodiment of the present invention provides an ultra-short-term wind power prediction method based on digital twins, the method comprising the following steps: Construct a wind power prediction system comprising a physical layer, a data processing layer, a digital twin layer, and a fusion output layer; The physical layer collects historical power data, meteorological data, wind turbine physical parameters and geographic information data of the wind farm. The wind turbine physical parameters include hub height, blade diameter, cut-in and cut-out wind speeds and rated power. In the data processing layer, the collected data is preprocessed, including filling missing values ​​with KNN interpolation and identifying and correcting power anomalies using the Physical Information Neural Network (PINN), wherein PINN introduces wind speed-power coupling relationship and Bates' law as physical constraints. The digital twin layer performs power prediction based on preprocessed data using both a digital physical model and a parallel Ctransformer-BiGRU deep learning model. The digital twin layer fuses the prediction results of the digital physical model and the parallel Ctransformer-BiGRU model through weighted fusion to obtain the final power prediction value; The prediction results are fused from the output layer and analyzed based on the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. Perform model performance evaluation and optimization.

[0028] The data preprocessing also includes feature extraction and fusion: The input features were screened using Pearson and Spearman correlation analysis, retaining features with high power correlation. By fusing low-correlation features, new features including x-axis wind speed, cosine wind direction, wet-bulb temperature, and thermodynamic index are generated. Introducing time-scale features, including year, month, day, hour, and season information, enhances the model's ability to capture periodic changes.

[0029] The construction of the digital physical model specifically includes: The model is based on the wind turbine's theoretical power generation formula, and the specific formula is as follows:

[0030] in, Indicates air density, Indicates the area swept by the wind turbine. Indicates wind speed. Indicates the wind energy utilization coefficient; By introducing an air density correction term:

[0031] in, This indicates the density of air at standard atmospheric pressure. This refers to the actual altitude, which is related to the latitude and longitude of the wind farm. Atmospheric characteristic altitude, It is the acceleration due to gravity. Molar mass of air and These are the ideal gas constant and the standard atmospheric temperature, respectively. Combining the correction term for air density with the formula for theoretical wind turbine power generation, the theoretical power generation of wind power is obtained as follows:

[0032] Given that the known cut-in and cut-out wind velocities of the wind turbines at the power station are 3 m / s and 22 m / s respectively, and based on the theoretical power curve of the power station, the wind turbines reach their maximum power of 1.5 MW when the wind speed v reaches 10 m / s. A piecewise function representation of this is constructed for the digital physical model.

[0033] in, This is the actual air density. Indicates the area swept by the wind turbine. Indicates wind speed. Indicates the wind energy utilization coefficient. The rated power of a single fan is defined as follows: when the wind speed is less than 3 m / s, the fan is in a stopped state; when the wind speed is greater than 3 m / s and less than 10 m / s, the fan output power satisfies Bates' Law; when the wind speed is greater than 10 m / s and less than 22 m / s, the fan output power is 1.5 MW, the rated power of a single fan; when the wind speed is greater than 22 m / s, the fan is in a cut-off state to protect the equipment, and the power is 0.

[0034] The construction of the parallel Ctransformer-BiGRU model includes: Spatial features of meteorological data are extracted using a CNN module, including convolutional and pooling layers; Use the Transformer module to capture long-term dependencies in time series, including multi-head self-attention mechanisms and feedforward networks; Use the BiGRU module to capture the short-term bidirectional fluctuation characteristics of time series. The CNN-Transformer and BiGRU are connected in parallel, and their outputs are fused through feature concatenation.

[0035] Also includes: The TPE algorithm was used to optimize the hyperparameters of the parallel Ctransformer-BiGRU model. The optimized parameters included learning rate, training epochs, batch size, number of CNN units, number of Transformer heads, dimension of the feedforward network, number of BiGRU units, and dropout rate. The TPE algorithm finds the optimal combination of hyperparameters by maximizing the expected improvement of the EI function.

[0036] The formula for fusing the prediction results is:

[0037] Among them, P A P is the predicted value of the digital physics model. B denoted as the predicted value of the deep learning model, and 'a' as the fusion hyperparameter, which is determined by the random forest algorithm in the interval [0,2] to minimize the RMSE between the fusion result and the true value.

[0038] An early stopping mechanism is used during model training, and training is stopped when the validation set loss function no longer decreases within 20 consecutive rounds. The Adam optimizer is used to optimize the model parameters, and the loss function is the mean absolute error (MAE).

[0039] The specific formula for calculating the Pearson and Spearman correlation analysis is as follows: Pearson correlation analysis calculation formula:

[0040] in, To delay time, The Pearson correlation coefficient is used. n For data length, The power at the time to be predicted, Before the predicted time Power at any moment and This is the corresponding average value; Spearman correlation analysis calculation formula:

[0041] Where n is the number of samples. This represents the positional difference of the i-th data pair; The formula for calculating the Pearson-Spearman bicorrelation analysis is obtained by combining Pearson correlation analysis and Spearman correlation analysis: .

[0043] For example, the system of the present invention mainly consists of three main modules: a physical layer, a data processing layer, and a digital twin layer. The physical layer represents the actual physical objects of the wind farm, including historical power data, meteorological data, and wind turbine equipment and sensors. The data processing layer mainly includes a data cleaning model, a feature extraction model, and a PINN data correction layer. The digital twin layer includes a digital twin prediction model composed of a digital physical model and a deep learning model, as well as a data output model.

[0044] The physical layer refers to the actual physical objects in the real world, such as information on wind turbine components, geographical information about wind farms, wind power data, and meteorological data. The physical layer collects and stores the necessary data and parameters. Wind turbine component information includes hub height, blade diameter, and cut-in / cut-out wind speeds. Geographical information about wind farms includes latitude and longitude, land area, and the number of wind turbines.

[0045] (1) (2) (3) As shown in the above formula, For the first j Historical data from each sampling point, including meteorological characteristics. Actual power data of wind farms Among them, meteorological characteristics W This includes hub height, wind speed, wind direction, temperature, wind farm air pressure, and air humidity. Wind power data. P This represents the power generation capacity of the wind turbine at each sampling point.

[0046] The data processing layer serves as the connection channel between the physical layer and the digital twin layer, enabling the transmission of wind farm-related data to the deep learning load forecasting model based on the digital twin. This layer consists of a data cleaning layer, a feature fusion layer, and a PINN data correction layer. The data cleaning layer fills in missing values ​​and removes outliers from the original data, the feature fusion layer extracts features, reduces dimensionality, and performs multi-scale fusion on the original input features, and the PINN data correction layer implements physical correction of the input data by introducing a Physical Information Neural Network (PINN).

[0047] The digital twin layer constructs a digital twin model and uses it to predict wind power. To accurately reflect the actual characteristics of wind farms, it is necessary to consider the physical characteristics and geographical features of wind farms and extract the inherent relationships in relevant data. By constructing a digital physical model that can reflect the physical operation mechanism of wind power generation and a parallel Ctransformer-BiGRU model based on deep learning, hidden spatiotemporal features are extracted. A digital twin fusion formula is constructed to combine the above models, realizing wind power prediction that considers the physical characteristics of wind turbines and the spatiotemporal features of historical data.

[0048] The prediction results are fused from the output layer and analyzed based on the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. The model performance was evaluated and optimized. The prediction effect of the model was evaluated by analyzing the error evaluation index, and the performance of the proposed digital twin-based prediction model was further evaluated based on ablation experiments and comparative experiments.

[0049] For data cleaning in data preprocessing, the raw data often contains outliers and missing values ​​due to the frequent impact of extreme weather, turbine start-up and shutdown, and errors in data acquisition equipment during wind turbine operation. Directly transmitting outlier data to the digital twin layer will prevent the model from accurately capturing the characteristics of the raw data. Therefore, data cleaning is necessary for the data transmitted from the physical layer to remove outliers and missing values. This method uses KNN interpolation to fill in missing values ​​in the original data. It automatically calculates and fills in missing values ​​based on the nearest neighbor values. Outliers are replaced by the average of the time steps before and after the outlier. The KNN interpolation coefficient is set to n=30, meaning missing values ​​are replaced by the average of the nearest 30 time steps. The basic idea of ​​KNN interpolation is to calculate the similarity between the data point to be interpolated and the training set data points, based on the category or value of the n nearest neighbors. The core lies in the selection of the similarity metric. The Euclidean distance between the points to be interpolated is calculated using equation (4), where... The point to be interpolated. Let j be the j-th feature of the missing point. Let be the j-th feature of the i-th nearest neighbor. Fill in the gaps using the average of the k nearest neighbor values, as shown in equation (5), where... The value to fill in is n, where n is the number of adjacent elements. The value of the i-th point (4) (5) For outliers in the original data, a Physical Information Neural Network (PINN) is used to integrate the wind turbine output characteristic calculation formula and relevant physical parameters of the wind farm. This allows for the determination of the coupling relationship between wind speed v and wind power P, as well as the wind turbine's cut-in wind speed. Cut-off wind speed Information such as the rated power of a single wind turbine is integrated into a physical information neural network to construct a wind power outlier correction model. The PINN model consists of a deep neural network, and its characteristic is that a physical information term, i.e. the physical law it follows, is added to the loss function. When training the model, it is necessary to minimize not only the data error, but also the physical information error to ensure that the prediction results conform to the physical characteristics of wind power generation.

[0050] Feature extraction and fusion in data preprocessing, particularly input data feature selection, is a crucial step in power prediction, directly impacting model prediction performance and efficiency. Raw data often contains redundant and interfering information; directly using all features for training leads to excessively long training times and reduced feature extraction capabilities. Wind power generation is influenced by various meteorological factors, such as wind speed, temperature, and humidity. The selected raw input features include time, wind speed, wind direction, temperature, humidity, and atmospheric pressure. Correlation analysis is performed on the cleaned feature columns. Here, Pearson & Spearman correlation analysis is used for feature selection. Based on the results, appropriate input features are chosen to achieve feature dimensionality reduction. The relevant calculation formulas are as follows: (6) (7) (8) Equation (6) is the formula for Pearson correlation analysis, where: For delay time; Here, n is the Pearson correlation coefficient; n is the data length. The power at the moment to be predicted; Before the predicted time Power at any moment; and This represents the corresponding average value. Equation (7) is the formula for Spearman correlation analysis, where n is the number of samples. Let be the position difference of the i-th data pair. Equation (8) is a combination of Pearson & Spearman bicorrelation analysis.

[0051] Reference Figure 3 For the digital physics model, the power generation of the wind farm is modeled digitally. It is known that the theoretical power generation of the wind turbine is related to physical quantities such as wind speed, rotor swept area, and air density. The specific calculation formula is as follows: (9) In the formula, Indicates air density, Indicates the area swept by the wind turbine. Indicates wind speed. The wind energy utilization coefficient is represented by , and according to Bates theory, its maximum value is 0.593. For the air density in equation (9)... Its expression under standard atmospheric pressure is shown in the following formula: (10) In equation (10), m represents the gas mass and V represents the gas volume. Represents the pressure of a gas. Represents the molar mass of air. Represents the ideal gas constant. This represents the temperature of an ideal gas. Since the values ​​of physical quantities such as air pressure change at different altitudes, the above equation is corrected as follows: (11) In equation (11), This indicates the air density at standard atmospheric pressure, which is approximately 1.225 kg / m³. This represents the actual altitude, which is related to the latitude and longitude of the wind farm. The atmospheric characteristic altitude is approximately 8435 meters. This is the acceleration due to gravity. The molar mass of air is approximately 0.02895 kg / mol. and Here, denoted by the ideal gas constant and standard atmospheric temperature, respectively, are used. Combining the corrected air density and the theoretical power generation formula for wind turbines, the theoretical power generation of wind power is obtained as follows: (12) The data used in this invention comes from wind turbine #9 of the Guotou Santanghu Wind Farm in Xinjiang, my country. The wind turbine's coordinates are 93:06:53.29907E, 44:04:15.73557N, with an altitude of 1378.79m. The actual air density at the wind turbine location was calculated. =1.22497kg / m3, therefore the digital physical model of wind turbine power generation is corrected as shown in equation (13).

[0052] (13) The main physical quantities in equation (13) include the blade diameter, wind speed, and operating efficiency of the wind farm. It is known that the cut-in and cut-out wind speeds of the wind turbine at this site are 3 m / s and 22 m / s, respectively. According to the theoretical power curve of the site, the wind turbine reaches its maximum power of 1.5 MW when the wind speed v reaches 10 m / s. Since the wind turbine operation must satisfy Bates' law, the cut-in and cut-out wind speeds and Bates' law are selected as constraints. Equation (13) is selected as the digital physical model, and the blade diameter and wind speed are selected as input feature data for physical model construction. Since the wind turbine has a minimum cut-in wind speed to drive the blade rotation and protective measures are taken to ensure the long-term operation of the wind turbine, the digital physical model of equation (13) is modified to a piecewise function representation of equation (14).

[0053] (14) In the formula This is the actual air density. Indicates the area swept by the wind turbine. Indicates wind speed. Indicates the wind energy utilization coefficient. This refers to the rated power of a single fan. Specifically, when the wind speed is less than 3 m / s, the fan is in a stopped state; when the wind speed is greater than 3 m / s but less than 10 m / s, the fan output power satisfies Bates' Law; when the wind speed is greater than 10 m / s but less than 22 m / s, the fan output power is 1.5 MW (the rated power of a single fan); when the wind speed is greater than 22 m / s, the fan is cut off to protect the equipment, and the power is 0.

[0054] This preliminary wind power prediction model is based on TPE-Ctransformer-BiGRU. TPE (Tree-structured Parzen Estimator) optimizes the model's hyperparameters using Bayesian optimization to efficiently search for the optimal hyperparameter combination. CTransformer combines the advantages of Convolutional Neural Networks (CNN) and Transformers. CNN effectively extracts the spatial features of wind farms and captures the spatial correlation of wind turbines; Transformer, with its self-attention mechanism, captures the long-term dependencies of time series data and handles the periodicity and volatility of the data. BiGRU focuses on handling short-term fluctuations and bidirectional information in time series data, considering both positive and negative time dependencies through its bidirectional structure, enabling rapid response to instantaneous power changes. The parallel structure of CTransformer and BiGRU allows them to complement each other, modeling the data from different perspectives, thus more comprehensively capturing the complex characteristics of wind power data and improving prediction accuracy and robustness.

[0055] The CNN-Transformer model primarily consists of an encoder and a decoder, connected by a multi-head attention mechanism and a feedforward neural network. Through parallel processing capabilities and positional encoding, this model can process sequential data in parallel, thereby improving processing efficiency. The Transformer introduces an attention mechanism to achieve fast parallelism, overcoming the slow training of recurrent neural networks. Due to its unique architecture and capabilities, it has been used in time series analysis. Its core application in time series analysis lies in its self-attention mechanism, which effectively captures long-term dependencies in time series data.

[0056] Reference Figure 4 Bi-GRU is a special variant of GRU, consisting of two GRU models with the same output but opposite information transmission directions. The Bi-GRU model can simultaneously consider information from past and future moments, fully exploiting the temporal characteristics of power data and improving data utilization and model prediction accuracy. Based on the working principle of the GRU model, a Bi-GRU model is constructed by setting positive and negative order GRU layers. for Time-order GRU model output. for The output of the negative-order GRU model at time step 1, and the calculation formula of the BiGRU model can be expressed as: (15) (16) (17) In the formula: and The input weight matrix and are respectively the forward-order GRU model input weight matrix and Output the weight matrix at each time step; and The input weight matrix and are respectively the negative-order GRU model input weight matrix and Output the weight matrix at each time step; and Output the weight matrices for the positive-order GRU model and the negative-order GRU model, respectively.

[0057] This invention employs a tree-structured parzen estimator (TPE) algorithm for hyperparameter optimization. The predictive performance of the Ctransformer-BiGRU model is affected by parameters such as the number of network units, the number of network layers, the learning rate, and the dropout rate; therefore, hyperparameter tuning is necessary. The TPE model uses two density functions. , To define .

[0058] (18) Is The loss is less than the threshold Observation density at time The density is formed by the remaining observations. TPE selects the expected improvement (EI) as the acquisition function, as shown in Equation (19).

[0059] (19) in Represents the conditional probability distribution. Represents the marginal probability distribution of the target value. Let represent the prior probability distribution of the hyperparameters. , representing a certain quantile in the TPE algorithm, is used to divide the two density functions and calculate the prior probability distribution. As shown in equation (20).

[0060] (20) Therefore, the expression for the acquisition function EI (19) can be transformed into equation (20). From the above equation, it can be seen that in order to maximize EI, it is necessary to ensure that at point... The probability of falling is high and The probability is low, so TPE is used. The algorithm evaluates the selected candidates and, after a certain number of iterations, returns the candidate with the highest EI. , (twenty one) (twenty two)

[0061] Reference Figure 5, based on the TPE-Ctransformer-BiGRU model structure, it consists of a TPE optimization model, a convolutional neural network (CNN), a Transformer model, and a bidirectional gated recurrent unit (BiGRU) model. The main structure of this model includes two parallel model channels: a CNN-Transformer series model and a series model composed of multiple layers of bidirectional gated recurrent units. Then, the TPE optimization model is used to optimize the hyperparameters of the Ctransformer-BiGRU model. In wind power prediction, meteorological data such as wind speed and wind direction has spatial correlation. The CNN extracts the local features hidden in the input data through convolutional layers and pooling layers, and captures the spatial features hidden in historical data through convolutional operations. At the same time, the convolutional neural network can gradually reduce the dimensionality of high-dimensional input data while retaining key information, which helps to reduce computational complexity. The self-attention mechanism of the Transformer allows the model to focus on the entire sequence information at each time step, enabling the model to better understand global time dynamics. This model processes long sequence data through the self-attention mechanism, captures long-distance time dependencies such as seasonal changes and daily cycle changes. At the same time, compared with traditional recurrent neural network models, the Transformer model can process data in parallel, greatly improving the operation efficiency of the model. The CNN-Transformer part of this model can process information in both spatial and time dimensions simultaneously. By combining the local feature extraction ability of the CNN and the long sequence modeling ability of the Transformer, it can more accurately predict the future value of wind power.

[0062] For the digital twin fusion model, in order to utilize the advantages of the digital physical model and the parallel Ctransformer-BiGRU model, the prediction results of the two models can be linearly combined, and the digital twin fusion result is used as the final predicted wind power. The prediction result of the digital physical model is defined as PA, the deep learning prediction result is defined as PB, and the final fusion prediction result is PD. The final prediction result is as shown in Equation (23): (23) In the above formula, a is the digital twin fusion hyperparameter, and its value range is 0 < a < 2. Given the actual wind turbine power generation value PT of the wind farm, the random forest algorithm is used to optimize the digital twin parameters in combination with Equation (23).

[0063] The specific example analysis is as follows: For the experimental platform and data sources, this invention uses PyCharm 2024.2.4 (Community Edition) as the development platform and Anaconda3 as the integrated development environment for the model. A deep learning framework based on Tensorflow-2.10 and Keras-2.10.0 was built using Python 3.9 to create an ultra-short-term wind power prediction model. Wind power data and meteorological data from July 2023 to July 2024 for one year from the No. 9 wind turbine of the Guotou Santanghu Wind Farm in Xinjiang, my country, were selected as the dataset for simulation verification to fully evaluate the effectiveness of the proposed method. The selected data has a time granularity of 1 minute, totaling 520,000 data points, with a training set, validation set, and test set ratio of 8:1:1. Experimental equipment: Windows 11 operating system, 13th Gen Intel(R) Core(TM) i7-13700KF CPU, NVIDIA GeForce RTX 4060 Ti GPU, and 32GB of memory.

[0064] For the data processing results, the selected data underwent data preprocessing. First, KNN interpolation and PINN data correction were used to preprocess the original data. Then, Pearson & Spearman correlation analysis, feature extraction, and fusion were performed on the processed data. The processing results are as follows: The results of the meteorological feature correlation analysis of the input data are as follows: Figure 6 As shown, analysis Figure 6 It is evident that data columns with high correlation to power generation include wind speed, temperature, humidity, and blade pitch angle. Data such as wind turbine status and pressure have low correlation with power generation; therefore, meteorological features with high correlation, such as wind speed and temperature, are selected as initial input features. Feature fusion is performed on features such as wind direction, wind turbine status, and pressure to obtain feature data with higher correlation. The new features obtained from feature fusion are shown in Table 1.

[0065] Table 1 Feature fusion results

[0066] Besides meteorological characteristics, time-scale features are also closely related to wind power output. The time-scale feature matrix, closely related to the time-period features of wind power output, can provide systematic time-series data for model training. Time features include year, season, month, date, hour, and minute. These features reflect the time-period and seasonal patterns of the wind power system. Specifically, changes in the year reflect long-term trends and changes in the system environment; seasonal changes and months reflect changes in wind speed; and dates and hours reflect periodic features such as daily temperature and wind direction changes.

[0067] The final model input features are shown in Table 2. By using the meteorological features and features generated by fusing the original data to generate highly correlated features and time features as input features to the model, the model's feature extraction capability is improved. The model can more accurately capture the fluctuation of wind power generation and the periodic changes of wind speed in the environment, which is crucial for improving the accuracy of wind power prediction models.

[0068] Table 2 Model Input Features

[0069] For model parameter selection, this invention uses the TPE hyperparameter optimization method under the Optuna optimization framework to optimize the key hyperparameters of the parallel Ctransformer-BiGRU model. The working principle of this method is to iteratively update the required hyperparameters of the model within the given hyperparameter search space, aiming to minimize the model training loss function and ultimately select the optimal hyperparameters for output. The deep learning part of the model selected in this invention mainly consists of a parallel Ctransformer-BiGRU model, where the number of layers in both the CNN and BiGRU models is set to 2. The main hyperparameters include global hyperparameters and individual hyperparameters of each model. The specific hyperparameters and search range settings are shown in Table 3 below. Table 3 Hyperparameter Search Space

[0070] Based on the parameter ranges in the table above, hyperparameter optimization was performed on the deep learning part of the proposed model. The TPE model was used for 100 parameter optimizations, and the optimal hyperparameter results are shown in Table 4. The model training epochs were set to 80, and the learning rate was set to 0.0373. The input batch size was set to 512. The number of units in the CNN layers was 256 and 192, respectively. The number of multi-head attention heads in the transformer layer was set to num_heads=3, the feedforward network dimension was ff_dim=64, and the BiGRU model had 2 layers with 32 and 192 units, respectively.

[0071] Table 4. Search results for hyperparameters

[0072] Model training was performed using the hyperparameters in Table 4. The Adam optimizer was selected for model optimization, with the early stopping mechanism parameter set to patience=20. The mean absolute error (MAE) was chosen as the loss function during training. The training set loss and validation set loss curves during training are shown below. Figure 7 As shown in the analysis, the training set loss (train_loss) gradually stabilizes after 20 training iterations. The validation set loss (val_loss) gradually stabilizes after 30 training iterations. At 60 training iterations, both loss curves approach convergence. The hyperparameters of the digital twin fusion model are optimized based on the principle that the prediction error is minimized when the root mean square error (RMSE) between the fusion result and the true power is minimized. The hyperparameter 'a' is set to a range of [0,2] with 10000 interval points. The optimization results are as follows: Figure 8 As shown.

[0073] Finally, the random forest algorithm was used to find that when the hyperparameter 'a' was 0.4364, the RMSE was minimized to 0.0641, indicating that the fusion model achieved the best prediction performance. Analysis Figure 8 It can be seen that when the value of 'a' is less than 0.4364, the RMSE gradually decreases until it reaches its minimum value of 0.0641 when a = 0.4364. Specifically, when a = 0.1, the RMSE value is 0.073, which is greater than the minimum value of 0.0641. When the value of 'a' is greater than 0.4364, the RMSE gradually increases until a = 2.0, at which point the RMSE = 0.22, which is much greater than the minimum value. Therefore, the hyperparameter 'a' should be set to 0.4364.

[0074] For performance evaluation and analysis, since the power generation of wind turbines can reach zero, the mean absolute percentage error (MAPE) commonly used in power prediction becomes meaningless. Therefore, this invention selects mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. The evaluation index is used as a primary indicator to characterize the model's prediction accuracy from multiple perspectives. The calculation formula for the evaluation index is as follows: (twenty four) (25) (26) In the formula: , and These represent the actual wind power output, the average value of the actual wind power output, and the predicted value, respectively. To predict the number of data points in the validation set; Number the predicted point sequence. Figure 9 To predict the power curves, curve "true" represents the actual power curve, curve "Ctransformer-BiGRU" represents the future power predicted by the deep learning model, curve "Physical" represents the power curve predicted by the digital physical model, and curve "Proposed Model" represents the wind power prediction curve based on digital twins. It can be seen that the deep learning method performs better in fitting the overall trend and has an advantage in capturing the curve's changing process. This model learns prediction patterns by analyzing a large amount of historical data, possessing the ability to handle large amounts of data and capture complex nonlinear relationships, and performing excellently in handling rapidly changing data and recognizing patterns. When predicting extreme power values ​​caused by strong winds, the physical method significantly outperforms the deep learning method. This is because the digital physical model integrates the physical characteristics of the wind farm and environmental factors, making the model more tightly coupled with the actual wind farm, thus better capturing power values ​​caused by sudden changes in physical information. The digital twin method combines the advantages of both the digital physical model and the deep learning model, exhibiting significant advantages in feature extraction based on large amounts of historical data and in a deep understanding of the physical laws governing wind power.

[0075] Table 5 shows the evaluation indicators for the prediction results of this invention. The table evaluates the prediction results of the digital physical model, deep learning model, and digital twin model, using the selected evaluation indicators: MAE, RMSE, and R. 2 The prediction accuracy of the digital physical model reached 92.2%, with a mean absolute error and root mean square error of 0.0501MW and 0.1007MW, respectively. This indicates that combining the actual physical information of the wind farm and the actual physical model of the wind turbine improves the model's prediction performance. Meanwhile, the prediction accuracy of the parallel Ctransformer-BiGRU model based on deep learning reached 94.07%, which is significantly higher than that of the digital physical model R. 2The accuracy improved by 2.2%, with MAE reaching 0.0602MW and RMSE reaching 0.0879MW, representing a 20.4% improvement and a 14.6% reduction compared to the digital physical model. This deep learning model incorporates a bidirectional gated recurrent unit (BiGRU) and a Transformer model, offering efficient time series processing and powerful feature extraction capabilities. It can process data in parallel, exhibiting excellent flexibility, adaptability, and generalization ability. The digital twin model's prediction accuracy reached 96.85% compared to the former two, representing improvements of 5.03% and 2.87% respectively, while MAE and RMSE decreased by 28.9% and 27.1% respectively compared to the standalone Ctransformer-BiGRU.

[0076] Table 5 Evaluation Indicators for Model Prediction Results

[0077] Figure 10(a) and (b) show the predicted waveforms for windy and calm periods, respectively. Analysis of waveform (b) shows that the wind turbine power generation is relatively stable during calm periods. During this stage, the three models shown in the figure all exhibit high prediction accuracy and relatively small prediction errors. Analysis of Figure (a) shows that when strong winds occur at the wind farm, the rapid changes in wind speed and direction lead to increased power generation, but also cause extreme and unstable values ​​in the power curve. In this situation, the predicted_power model of the digital twin prediction model significantly outperforms the other two models, indicating that the digital twin model, which incorporates physical information, can better capture meteorological changes and real physical spatial variations.

[0078] For the ablation experiments, to verify the effectiveness of each module component of the proposed digital twin-based wind power prediction model and its contribution to model performance improvement, the effects of each module on model performance improvement were compared and verified by reducing the number of model modules. Five comparison scenarios were set up in this invention: a) The digital twin-based wind power prediction model, i.e., the proposed digital twin model integrating a digital physical model and a parallel Ctransformer-BiGRU. b) Eliminating the digital physical model part of the digital twin model, retaining only the parallel Ctransformer-BiGRU model. c) Eliminating the BiGRU model, retaining only the Ctransformer model. d) Eliminating the transformer model of the Ctransformer model, retaining only the CNN model. These scenarios were used to compare and verify the performance improvement of the digital physical model, BiGRU model, and Transformer model. The ablation experiments selected mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (CQD). () is used as a model evaluation metric. The waveforms of the model's prediction results are compared as follows: Figure 11 As shown in Table 6, the model evaluation metrics are as follows.

[0079] Table 6 Comparison of Ablation Experiments

[0080] Analysis and comparison Figure 11 As shown in Table 6, compared to the digital twin model, Ctransformer-BiGRU, by removing the digital physical model, lacks coupling with the actual physical space and awareness of wind farm physical parameters, resulting in a significant performance decrease. MAE and RMSE increased by 40.65% and 37.13%, respectively. Compared to the digital twin model proposed in this invention, the prediction accuracy of Ctransformer-BiGRU decreased by 2.88%. Comparing the prediction curves of Ctransformer-BiGRU and the digital twin model in Figure 10, the prediction results of the digital twin model are closer to the true values, demonstrating the advantage of the digital physical model in the model.

[0081] Compared to the digital twin model, the Ctransformer model lacks a digital physical model and a parallel BiGRU model. It doesn't integrate information from the physical model, limiting its ability to deeply understand the physical model and site characteristics. The absence of the BiGRU module also restricts its ability to capture time-series features, thus reducing prediction accuracy. Compared to the digital twin model of this invention, the Ctransformer model improves the MAE by 83.64% and 62.09%, respectively, while reducing the accuracy R² by 5.29%, demonstrating the importance of the BiGRU module in capturing time-series features. Compared to the digital twin model, the CNN model reduces prediction accuracy by 11.29%. The curves in the graph show that the CNN model is significantly less accurate than the digital twin model. Compared to the Ctransformer-BiGRU and Ctransformer models, this model increases MAE by 57.47% and 20.6%, respectively, and RMSE by 54.04% and 30.32%, respectively. These results fully demonstrate the importance of the BiGRU module and the Transformer model for prediction results.

[0082] Comparison and analysis of the digital twin model with Ctransformer-BiGRU, Ctransformer, and CNN models show that the digital twin wind power prediction model, which integrates the digital physical model and the parallel Ctransformer-BiGRU model, can effectively mine the key spatiotemporal features of the data, effectively learn the wind power physical model and the real-time meteorological change information of the wind farm, significantly reduce the model error, and further improve the model's fitting degree.

[0083] For comparative experiments, to further demonstrate the superiority of the proposed digital twin-based prediction model in wind power prediction, it is necessary to compare it with different prediction models. Existing technologies use combined models of CNN, LSTM, and GRU to predict power load and wind speed in wind farms, respectively. These combined models perform well in the prediction field and have high prediction accuracy. Since the selected wind power data has significant spatiotemporal correlation, and models such as CNN, LSTM, and GRU are representative and have superior performance in processing time series data and spatial features, and the model proposed in this invention covers a bidirectional model architecture, four models—CNN-LSTM, CNN-GRU, BiLSTM, and BiGRU—were selected for model comparison experiments. By selecting commonly used spatiotemporal prediction models and their hybrid architectures for comparison experiments, the superiority of the proposed model in terms of prediction accuracy can be more intuitively demonstrated. This invention selected the prediction results from 09:00 to 12:00 on September 20, 2023. Figure 13As can be seen, for most of the time period, the predicted values ​​of the digital twin prediction model proposed in this invention are close to the actual values, and the fitting degree of the power curve trend is also high. Compared with the several models used for comparison, the digital twin model has a smaller prediction error for the actual value. The BiLSTM and CNN-LSTM models have good prediction effects on wind power trends, but still have large numerical errors. The prediction curves of the BiGRU and CNN-GRU models are close to the actual values, but there are still deviations in the location of extreme points.

[0084] Table 7 and Figure 13 The analysis of the evaluation indicators in the above comparative experiments shows that the mean absolute error (MAE) of the wind power prediction model based on digital twin proposed in this invention is reduced by 40.97%, 40.69%, 54.52%, and 32.81% compared to CNN-LSTM, CNN-GRU, BiLSTM, and BiGRU, respectively, indicating a significant improvement in model accuracy. Regarding the root mean square error (RMSE), the digital twin model shows a larger reduction compared to CNN-LSTM and BiLSTM models, decreasing by 44.07% and 50.12%, respectively, demonstrating that the digital twin model, which integrates the digital physical model and the Ctransformer-BiGRU model, offers a significant improvement in prediction accuracy compared to the LSTM model. The RMSE of the prediction results of the digital twin model is reduced by 35.19% and 39.29% compared to CNN-GRU and BiGRU, respectively, proving the advantages of the model proposed in this invention in the field of wind power prediction. In terms of the coefficient of determination (… In terms of prediction accuracy, excluding the digital twin model proposed in this invention, the best fits were achieved by the CNN-GRU model and the BiGRU model, reaching 92.49% and 91.44% respectively. This demonstrates that the GRU model has a greater advantage over LSTM in the prediction field. The digital twin model proposed in this invention achieved a prediction accuracy of 96.85%, which is an improvement of 7.71%, 4.71%, 10.91%, and 5.92% compared to CNN-LSTM, CNN-GRU, BiLSTM, and BiGRU, respectively. This indicates that the digital twin prediction model, which integrates the digital physical model, can effectively extract the physical characteristics and meteorological and geographical factors of wind farms, identify information changes at actual wind farms, and improve the prediction accuracy by integrating the spatial and temporal features extracted by CNN-Transformer and Bi-GRU.

[0085] Table 7 Evaluation Indicators for Comparative Experiments

[0086] In summary, based on the ablation and comparative experiments, the proposed digital twin-based power prediction model demonstrates superior performance in ultra-short-term wind power prediction. Ablation experiments show that the digital physics model, CNN-Transformer model, and parallel BiGRU model in the proposed model significantly improve the performance and efficiency of the prediction model. Comparative experiments show that it exhibits better prediction performance and higher prediction accuracy compared to traditional spatiotemporal prediction models such as CNN, LSTM, and GRU. The proposed model employs digital twin technology to establish a digital physics model of the wind power system, achieving real-time dynamic simulation of the wind farm. This model accurately simulates the physical processes of the wind power system, ensuring its physical rationality. The proposed model integrates CNN, Transformer, BiGRU models, and the TPE hyperparameter optimization model. CNN effectively extracts local features from wind power data and captures spatial correlations; the Transformer model uses an attention mechanism to capture long-distance dependencies and dynamically allocate weights, improving the model's efficiency and scalability; and the BiGRU model further enhances the bidirectional understanding of time series data. The TPE optimization algorithm used in this model employs an adaptive sampling strategy, which has high flexibility and scalability, and higher optimization efficiency compared to other optimization algorithms.

[0087] Reference Figure 2 Another embodiment of the present invention proposes a wind power prediction system based on digital twins, the system being used to perform wind power prediction based on digital twins as proposed in one embodiment, the system comprising: Data acquisition module: Collects historical power data, meteorological data, wind turbine physical parameters, and geographic information data of the wind farm; Data processing module: includes data cleaning submodule, feature fusion submodule and PINN correction submodule; Digital twin prediction module: includes a digital physical model submodule and a parallel Ctransformer-BiGRU model submodule; Fusion output module: used to weight and fuse the two types of prediction results, and output the final power prediction value and performance evaluation report.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for ultra-short-term wind power prediction based on digital twins, characterized in that, The method includes the following steps: Construct a wind power prediction system comprising a physical layer, a data processing layer, a digital twin layer, and a fusion output layer; The physical layer collects historical power data, meteorological data, wind turbine physical parameters and geographic information data of the wind farm. The wind turbine physical parameters include hub height, blade diameter, cut-in and cut-out wind speeds and rated power. In the data processing layer, the collected data is preprocessed, including filling missing values ​​with KNN interpolation and identifying and correcting power anomalies using the Physical Information Neural Network (PINN), wherein PINN introduces wind speed-power coupling relationship and Bates' law as physical constraints. The digital twin layer performs power prediction based on preprocessed data using both a digital physical model and a parallel Ctransformer-BiGRU deep learning model. The digital twin layer fuses the prediction results of the digital physical model and the parallel Ctransformer-BiGRU model through weighted fusion to obtain the final power prediction value; The prediction results are fused from the output layer and analyzed based on the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. Perform model performance evaluation and optimization.

2. The ultra-short-term wind power prediction method based on digital twins according to claim 1, characterized in that, The data preprocessing also includes feature extraction and fusion: The input features were screened using Pearson and Spearman correlation analysis, retaining features with high power correlation. By fusing low-correlation features, new features including x-axis wind speed, cosine wind direction, wet-bulb temperature, and thermodynamic index are generated. Introducing time-scale features, including year, month, day, hour, and season information, enhances the model's ability to capture periodic changes.

3. The ultra-short-term wind power prediction method based on digital twins according to claim 1, characterized in that, The construction of the digital physical model specifically includes: The model is based on the wind turbine's theoretical power generation formula, and the specific formula is as follows: in, Indicates air density, Indicates the area swept by the wind turbine. Indicates wind speed. Indicates the wind energy utilization coefficient; By introducing an air density correction term: in, This indicates the density of air at standard atmospheric pressure. This refers to the actual altitude, which is related to the latitude and longitude of the wind farm. Atmospheric characteristic altitude, It is the acceleration due to gravity. Molar mass of air and These are the ideal gas constant and the standard atmospheric temperature, respectively. Combining the correction term for air density with the formula for theoretical wind turbine power generation, the theoretical power generation of wind power is obtained as follows: Given that the known cut-in and cut-out wind velocities of the wind turbines at the power station are 3 m / s and 22 m / s respectively, and based on the theoretical power curve of the power station, the wind turbines reach their maximum power of 1.5 MW when the wind speed v reaches 10 m / s. A piecewise function representation of this is constructed for the digital physical model. in, This is the actual air density. Indicates the area swept by the wind turbine. Indicates wind speed. Indicates the wind energy utilization coefficient. The rated power of a single fan is defined as follows: when the wind speed is less than 3 m / s, the fan is in a stopped state; when the wind speed is greater than 3 m / s and less than 10 m / s, the fan output power satisfies Bates' Law; when the wind speed is greater than 10 m / s and less than 22 m / s, the fan output power is 1.5 MW, the rated power of a single fan; when the wind speed is greater than 22 m / s, the fan is in a cut-off state to protect the equipment, and the power is 0.

4. The ultra-short-term wind power prediction method based on digital twins according to claim 1, characterized in that, The construction of the parallel Ctransformer-BiGRU model includes: Spatial features of meteorological data are extracted using a CNN module, including convolutional and pooling layers; Use the Transformer module to capture long-term dependencies in time series, including multi-head self-attention mechanisms and feedforward networks; Use the BiGRU module to capture the short-term bidirectional fluctuation characteristics of time series. The CNN-Transformer and BiGRU are connected in parallel, and their outputs are fused through feature concatenation.

5. The ultra-short-term wind power prediction method based on digital twins according to claim 4, characterized in that, Also includes: The TPE algorithm was used to optimize the hyperparameters of the parallel Ctransformer-BiGRU model. The optimized parameters included learning rate, training epochs, batch size, number of CNN units, number of Transformer heads, dimension of the feedforward network, number of BiGRU units, and dropout rate. The TPE algorithm finds the optimal combination of hyperparameters by maximizing the expected improvement of the EI function.

6. The ultra-short-term wind power prediction method based on digital twins according to claim 1, characterized in that, The formula for fusing the prediction results is: Among them, P A P is the predicted value of the digital physics model. B denoted as the predicted value of the deep learning model, and 'a' as the fusion hyperparameter, which is determined by the random forest algorithm in the interval [0,2] to minimize the RMSE between the fusion result and the true value.

7. The ultra-short-term wind power prediction method based on digital twins according to claim 1, characterized in that, An early stopping mechanism is used during model training, and training is stopped when the validation set loss function no longer decreases within 20 consecutive rounds. The Adam optimizer is used to optimize the model parameters, and the loss function is the mean absolute error (MAE).

8. The ultra-short-term wind power prediction method based on digital twins according to claim 2, characterized in that, The specific formula for calculating the Pearson and Spearman correlation analysis is as follows: Pearson correlation analysis calculation formula: in, To delay time, The Pearson correlation coefficient is used. n For data length, The power at the time to be predicted, Before the predicted time Power at any moment and This is the corresponding average value; Spearman correlation analysis calculation formula: Where n is the number of samples. This represents the positional difference of the i-th data pair; The formula for calculating the Pearson-Spearman bicorrelation analysis is obtained by combining Pearson correlation analysis and Spearman correlation analysis: 。 9. A wind power prediction system based on digital twins for implementing the method of any one of claims 1 to 8, characterized in that, include: Data acquisition module: Collects historical power data, meteorological data, wind turbine physical parameters, and geographic information data of the wind farm; Data processing module: includes data cleaning submodule, feature fusion submodule and PINN correction submodule; Digital twin prediction module: includes a digital physical model submodule and a parallel Ctransformer-BiGRU model submodule; Fusion output module: used to weight and fuse the two types of prediction results, and output the final power prediction value and performance evaluation report.

Images