A wind speed correction method and system for wind farms based on dynamic spatiotemporal modeling

Abstract

This invention relates to the field of wind power generation technology and discloses a wind speed correction method and system based on dynamic spatiotemporal modeling. The method includes: obtaining the overall power curve of the wind farm and constructing a multimodal training dataset by obtaining the overall wind speed over historical periods; inputting the dataset into a convolutional neural network to extract local features, and then inputting it into a long short-term memory network to calculate the correlation of features at each time step and obtain attention weights, ultimately obtaining a global feature representation; setting two multilayer perceptron branches for wind speed prediction correction, obtaining predicted values ​​for ordinary weather and extreme weather; constructing correction curves for each sector and obtaining correction curve prediction values; and obtaining the final corrected wind speed value through weighted fusion of the predicted values ​​for ordinary weather, extreme weather, and correction curves. This application improves correction accuracy and robustness, and enhances the interpretability and applicability of the model.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and more specifically, to a wind farm wind speed correction method and system based on dynamic spatiotemporal modeling. Background Technology

[0002] With the continuous expansion of wind power generation, the accuracy of wind power forecasting is crucial for ensuring the safe operation of the power grid and the friendly integration of new energy sources. Due to the coupling effects of meteorological system disturbances, terrain influences, and eddy currents, wind speed exhibits highly nonlinear and non-stationary characteristics in both time and space. This makes it difficult to meet the accuracy requirements of wind farm operation and scheduling by directly relying on numerical weather prediction models. Therefore, intelligent correction of forecasted wind speeds has become a key step in improving the reliability and accuracy of wind power forecasting.

[0003] Most existing correction methods are based on neural networks to model the mapping relationship between meteorological elements and wind speed errors, but they still have significant shortcomings in modeling the dynamic changes in wind speed, characterizing the spatial heterogeneity of wind direction, and responding to extreme weather. For example, patent CN119067269A discloses an integrated wind farm prediction wind speed correction method and system, which realizes the correction output of forecast wind speed by constructing a multi-layer perceptron structure and fusing multi-dimensional input information such as wind speed, wind direction, and temperature. While this method improves correction accuracy to some extent, it still has some problems: the scheme fails to utilize the long-term dependence and nonlinear variation trend of wind speed time series, and cannot effectively cope with scenarios with drastic wind speed fluctuations; the scheme does not explicitly model the spatial differences corresponding to different wind directions in the wind farm, and cannot effectively identify local disturbances or uneven wind speed distribution under the dominance of wind direction; under extreme weather conditions, such as thunderstorms, strong winds, or sudden changes in airflow, the existing model does not set a dedicated prediction path, resulting in a decrease in model accuracy during critical periods; it lacks a physically inspired explicit correction mechanism, such as a correction curve strategy that fits the historical statistical relationship by dividing the sector by wind direction, resulting in weak model interpretability, susceptibility to changes in the distribution of training data, and insufficient stability.

[0004] Therefore, it is necessary to design a wind speed correction method and system for wind farms based on dynamic spatiotemporal modeling to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a wind speed correction method and system for wind farms based on dynamic spatiotemporal modeling, aiming to solve the problems of poor dynamic response capability and low adaptability to extreme weather in the current technology.

[0006] In one aspect, this invention proposes a wind speed correction method for wind farms based on dynamic spatiotemporal modeling, comprising:

[0007] The measured power curves of each wind turbine generator in the wind farm are obtained, and the measured power curves are preprocessed to obtain the power curve of the entire farm.

[0008] The wind speed for the entire field during historical periods is obtained from the overall power curve and correlated with the wind direction over time. At the same time, meteorological features for the corresponding time periods are collected to construct a multimodal training dataset.

[0009] The multimodal training dataset is input into a convolutional neural network to extract local features of wind speed spatiotemporal data, and the local features are input into a long short-term memory network. An attention mechanism is introduced into the feature sequence output by the long short-term memory network to calculate the correlation of features at each time step and obtain attention weights. The time step features are weighted and converged according to the attention weights to obtain a global feature representation.

[0010] Two multilayer perceptron branches are set up. The global feature representation is input into the corresponding multilayer perceptron branch to perform wind speed prediction correction and obtain the prediction value of ordinary weather branch and the prediction value of extreme weather branch.

[0011] The multimodal training dataset is divided into several levels of circular sectors according to the overall wind direction. Correction curves for each sector are constructed based on the relationship between historical weather forecast wind speed and the overall wind speed. The weather forecast wind speed for the target time period is input into the correction curves of the sector where the target wind direction is located and its adjacent sectors to obtain the sector correction results. The correction curve prediction value is obtained by fusing the sector correction results of the adjacent sectors through inverse distance weighting.

[0012] The final wind speed correction value is obtained by weighted fusion of the general weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value.

[0013] Furthermore, when preprocessing the measured power curve to obtain the overall power curve, the process includes:

[0014] The measured power curves of each wind turbine generator set are normalized at intervals along the wind speed axis to unify the wind speed grid. Linear interpolation is performed on the missing power data at the corresponding wind speed points and duplicate or outlier points are removed. Boundary constraints are set at the wind speed boundary according to the model with the largest rated wind speed and boundary value interpolation is performed to ensure the continuity of the curves, thus obtaining the optimized power curves of each unit.

[0015] At each uniform wind speed point, the optimized power values ​​are added together, and the summation is obtained by traversing the wind speed grid to obtain the overall power curve.

[0016] Furthermore, based on the overall power curve, the historical wind speed for each time period is obtained and correlated with the overall wind direction over time. Simultaneously, meteorological characteristics for the corresponding time periods are collected. When constructing a multimodal training dataset, the following steps are included:

[0017] The total power of the wind farm is obtained within a historical period, and a monotonic correspondence between power and wind speed is established based on the total power curve. The total wind speed at each moment is obtained on the wind speed grid based on reverse lookup and piecewise linear interpolation. The total wind direction representing the wind farm is obtained, the angle data is de-looped and quantified using sine and cosine encoding, and time-aligned with the total wind speed at a uniform sampling step size.

[0018] Meteorological features are collected on the same time axis as historical periods and normalized. The meteorological features include wind speed, wind direction, temperature, humidity, air pressure and boundary layer height.

[0019] Samples are constructed based on a sliding time window. The meteorological features within the window, the overall wind direction, and derived features are used as inputs. The overall wind speed corresponding to the forward step length of the window is used as a label to form the multimodal training dataset. Weather labels for ordinary weather and extreme weather are labeled according to wind force level thresholds. The derived features include wind vector components, wind shear index, wind direction turning rate, boundary layer stability index, and boundary layer height change rate.

[0020] Furthermore, when inputting the multimodal training dataset into a convolutional neural network to extract local features of wind speed spatiotemporal data, the process includes:

[0021] The multimodal training dataset is stacked in the time dimension according to the sliding time window, and the meteorological features and the overall wind direction are concatenated in the feature dimension to form a one-dimensional time series tensor.

[0022] A one-dimensional temporal convolutional neural network is used, with causal padding in the first layer and multi-scale convolutional kernels and optional dilated convolutions in the middle layers. The kernel length of the multi-scale convolutional kernels is one or more of 3, 5, and 7.

[0023] After each convolutional layer, nonlinear activation is sequentially applied, and deep degradation is suppressed through residual connections, ultimately outputting the local features.

[0024] Furthermore, when calculating the correlation of features at each time step and obtaining attention weights, and then weighting and converging the features at each time step according to the attention weights to obtain a global feature representation, the process includes:

[0025] The local features are input into a stacked long short-term memory network to model long-term dynamic dependencies and obtain a sequence of hidden states.

[0026] The relevance score is calculated based on the attention mechanism, and the attention weight is obtained by Softmax normalization.

[0027] The time-step features are weighted and converged according to the attention weights to obtain a context vector, and the context vector is then nonlinearly projected to obtain the global feature representation.

[0028] Furthermore, two multilayer perceptron branches are set up, and the global feature representation is input into the corresponding multilayer perceptron branch for wind speed prediction correction. When obtaining the predicted values ​​for ordinary weather and extreme weather, the process includes:

[0029] When setting up two multilayer sensor branches, there are two branches: one for general weather and one for extreme weather.

[0030] Each of the multilayer perceptron branches includes at least two fully connected layers and a nonlinear activation layer, with random deactivation set between layers to suppress overfitting; during the training phase, the parameters of the corresponding multilayer perceptron branch are updated only based on the weather label while the other multilayer perceptron branch is frozen; during the prediction phase, the global feature representation is fed forward into the two multilayer perceptron branches respectively to simultaneously generate the predicted values ​​of the normal weather branch and the extreme weather branch, and quantile loss is used as a robust loss function to constrain the training error.

[0031] Furthermore, when dividing the multimodal training dataset into several levels of circular sectors according to the overall wind speed, and constructing the correction curve for each sector based on the relationship between historical weather forecast wind speed and the overall wind speed, the process includes:

[0032] Using the overall wind direction as an angle variable, the wind direction circumference is divided into circular sectors of equal width, and the outer extension of the sector is divided into several rings according to the wind speed level to form a hierarchical structure.

[0033] Within each circular sector, the correction curve is obtained by performing monotonic constrained regression fitting with historical weather forecast wind speed as the independent variable and the overall wind speed as the dependent variable.

[0034] Furthermore, when inputting the forecasted wind speed for the target time period into the correction curves corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain the sector correction results, and obtaining the correction curve prediction value by inverse distance weighted fusion of the sector correction results of the adjacent sectors, the process includes:

[0035] The target circular sector and its adjacent circular sectors are determined according to the overall wind direction during the target time period. The meteorological forecast wind speed during the target time period is substituted into the corresponding correction curve to obtain multiple sector correction results. The inverse distance weight is calculated based on the angle between the central axis of each sector and the overall wind direction. The sector correction results are weighted and summed to obtain the predicted value of the correction curve.

[0036] Furthermore, when obtaining the final wind speed correction value through a weighted fusion of the general weather branch forecast value, the extreme weather branch forecast value, and the correction curve forecast value, the process includes:

[0037] Performance indices are calculated for the general weather branch forecast, the extreme weather branch forecast, and the correction curve forecast, and subjective and objective weights are determined based on the performance indices. Weight normalization and non-negativity constraints are used to obtain combined weights, where the sum of all weights in the combined weights is 1. The final wind speed correction value is calculated based on the combined weights, the general weather branch forecast, the extreme weather branch forecast, and the correction curve forecast.

[0038] Compared with existing technologies, the advantages of this invention are as follows: By acquiring and preprocessing the measured power curves of each wind turbine generator in a wind farm, the overall power curve of the wind farm is obtained. This allows for the acquisition of the overall wind speed through inversion and time-series pairing with the overall wind direction, constructing a multimodal training dataset encompassing meteorological characteristics, and achieving global dynamic perception of the wind farm's operating status. A convolutional neural network is used to extract local features from the spatiotemporal wind speed data, and a long short-term memory network is used to capture its time-dependent characteristics. An attention mechanism is introduced into the output sequence to weight and converge the importance of each time step, enhancing the model's ability to model wind speed trends. Two multilayer perceptron branches are set up for ordinary weather and extreme weather, enabling the learning of prediction strategies under different meteorological conditions, thus improving the response to extreme weather events. Spatially, the training dataset is divided into circular sectors based on the overall wind direction. Correction curves for each sector are constructed based on the relationship between historical forecast wind speed and measured wind speed. An inverse distance weighting strategy is used to fuse the correction results of neighboring sectors, generating physically meaningful correction curve prediction values. By integrating general weather forecasts, extreme weather forecasts, and correction curve forecasts, the final wind speed correction value is obtained, which improves the correction accuracy and robustness, enhances the interpretability and engineering applicability of the model, and makes up for the shortcomings of existing technologies in spatiotemporal dynamic modeling, wind direction zoning response, and extreme case adaptation.

[0039] On the other hand, this application also provides a wind farm wind speed correction system based on dynamic spatiotemporal modeling, used to apply the above-mentioned wind farm wind speed correction method based on dynamic spatiotemporal modeling, including:

[0040] The acquisition unit is configured to acquire the measured power curves of each wind turbine generator in the wind farm, and to preprocess the measured power curves to obtain the power curves of the entire farm.

[0041] The association unit is configured to obtain the wind speed of the entire field for a historical period based on the overall power curve, and to associate it with the wind direction over time. At the same time, it collects meteorological features of the corresponding period and constructs a multimodal training dataset.

[0042] The first processing unit is configured to input the multimodal training dataset into a convolutional neural network to extract local features of wind speed spatiotemporal data, input the local features into a long short-term memory network, introduce an attention mechanism into the feature sequence output by the long short-term memory network, calculate the correlation of features at each time step and obtain attention weights, and perform weighted aggregation of the time step features according to the attention weights to obtain a global feature representation.

[0043] The second processing unit is configured to set up two multilayer perceptron branches, input the global feature representation into the corresponding multilayer perceptron branches to perform wind speed prediction correction, and obtain the prediction value of ordinary weather branch and the prediction value of extreme weather branch.

[0044] The third processing unit is configured to divide the multimodal training dataset into several levels of circular sectors according to wind direction, construct correction curves for each sector based on the relationship between historical forecast wind speed and actual wind speed, input the meteorological forecast wind speed of the target time period into the correction curves corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain sector correction results, and obtain correction curve prediction values ​​by fusing the sector correction results of the adjacent sectors through inverse distance weighting.

[0045] The correction unit is configured to obtain the final wind speed correction value by weighted fusion of the general weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value.

[0046] It is understandable that the wind speed correction method and system for wind farms based on dynamic spatiotemporal modeling described above have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0047] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0048] Figure 1 A flowchart of a wind farm wind speed correction method based on dynamic spatiotemporal modeling provided in an embodiment of the present invention;

[0049] Figure 2 This is a functional block diagram of a wind farm wind speed correction system based on dynamic spatiotemporal modeling, provided in an embodiment of the present invention. Detailed Implementation

[0050] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] In traditional wind power forecasting systems, the correction process for wind speed forecasts often neglects the coupling effect of the spatiotemporal dynamics of wind speed. The nonlinear fluctuations and long-term dependencies of wind speed sequences are difficult to fully model using traditional neural networks, leading to increased prediction errors when wind speed changes rapidly or periodic disturbances exist. Simultaneously, the spatial heterogeneity of wind direction within wind farms is not effectively characterized, and the differences in local wind speed distributions corresponding to different wind directions under the same meteorological conditions are not explicitly modeled, causing the correction model to lag in response to sudden changes in wind direction or local eddies. Furthermore, abrupt changes in meteorological elements under extreme weather conditions differ from those in conventional weather models, but existing models do not design independent prediction paths for such scenarios, limiting the model's generalization ability under conditions of strong convection or boundary layer instability.

[0052] For example, in wind farms deployed in complex terrain areas, there is a systematic deviation between the regional wind speed data provided by the meteorological forecasting system and the actual anemometer observations at the wind farm. When the prevailing wind direction is affected by the flow around the mountain, forming a local vortex, the actual wind speed of each turbine within the wind farm exhibits spatial differentiation. Traditional correction models use a unified fully connected network structure to process multi-dimensional meteorological inputs, which cannot effectively capture the nonlinear relationship between wind speed and meteorological elements within different wind direction sectors. During severe convective weather, the sudden drop in boundary layer height exacerbates the vertical wind shear, but the model does not establish a dedicated feature extraction channel, causing the correction results to lag behind actual wind speed changes.

[0053] If the above problems are not addressed, the wind farm power prediction system will be unable to adapt to the dynamic characteristics of wind speed under complex terrain and variable weather conditions, leading to deviations between the power prediction curve and grid dispatch instructions. Long-term accumulated prediction errors may cause voltage fluctuations at the wind farm's grid connection point, increasing the burden on power system frequency regulation. In extreme weather events, uncorrected wind speed prediction deviations may cause turbines to operate at overspeed or exceed reactive power limits, triggering protection devices and impacting the wind farm's operational economy and equipment safety.

[0054] For this, please refer to Figure 1 As shown, this application proposes a wind speed correction method for wind farms based on dynamic spatiotemporal modeling, including:

[0055] S100: Obtain the measured power curves of each wind turbine generator in the wind farm, and preprocess the measured power curves to obtain the power curve of the entire farm.

[0056] S200: Obtain the wind speed of the entire field during historical periods based on the overall power curve, and correlate it with the wind direction over time. At the same time, collect meteorological characteristics of the corresponding time periods to construct a multimodal training dataset.

[0057] S300: Input the multimodal training dataset into the convolutional neural network to extract local features of wind speed spatiotemporal data, and input the local features into the long short-term memory network. Introduce an attention mechanism into the feature sequence output by the long short-term memory network, calculate the correlation of features at each time step and obtain attention weights, and perform weighted aggregation of time step features according to the attention weights to obtain global feature representation.

[0058] S400: Set up two multilayer perceptron branches, input the global feature representation into the corresponding multilayer perceptron branches to perform wind speed prediction correction, and obtain the prediction values ​​of ordinary weather branch and extreme weather branch.

[0059] S500: The multimodal training dataset is divided into several levels of circular sectors according to the overall wind direction. Based on the relationship between historical weather forecast wind speed and overall wind speed, correction curves for each sector are constructed. The weather forecast wind speed for the target time period is input into the correction curves of the sector where the target wind direction is located and its adjacent sectors to obtain the sector correction results. The correction curve prediction value is obtained by fusing the sector correction results of adjacent sectors through inverse distance weighting.

[0060] S600: The final wind speed correction value is obtained by weighted fusion of the forecast values ​​of ordinary weather branch, extreme weather branch, and correction curve.

[0061] Specifically, dynamic spatiotemporal modeling refers to extracting local features of wind speed spatiotemporal data through convolutional neural networks and combining them with long short-term memory networks to capture long-term dependencies in time series. This can be achieved by using one-dimensional temporal convolutional kernels and causal padding to process time-dimensional data, and suppressing deep degradation through residual connections. This feature effectively captures the dynamic changes in wind speed across time and space, addressing the shortcomings of existing methods in modeling scenarios with drastic wind speed fluctuations. The attention mechanism involves calculating the correlation of features at each time step in the feature sequence output by the long short-term memory network, and weighting and converging the features using attention weights. This can be achieved by using Softmax to normalize the correlation score and generate a context vector. This feature adaptively focuses on information from key time steps, enhancing the model's ability to represent the non-stationary characteristics of wind speed. The multilayer perceptron branch refers to setting up two independent prediction paths for ordinary and extreme weather conditions. This can be achieved by constructing the branch structure using fully connected layers and random deactivation techniques, and selectively updating parameters based on weather labels during the training phase. This feature allows for the design of differentiated prediction strategies for different weather conditions, improving the correction robustness under extreme weather scenarios. The circular sector correction curve refers to dividing the wind direction into hierarchical sectors and constructing a monotonic regression relationship between the forecast wind speed and the overall wind speed. Specifically, this can be achieved by dividing the wind direction circumference into equal-angled widths and fusing the results of adjacent sectors using inverse distance weighting. This feature can explicitly model the impact of spatial heterogeneity of wind direction on wind speed distribution, enhancing the interpretability and physical consistency of the correction process.

[0062] This application combines dynamic spatiotemporal modeling, a bi-branch prediction mechanism, and a physics-inspired sector correction strategy. It extracts spatiotemporal features using convolutional neural networks and long short-term memory networks, focuses on key time step information using an attention mechanism, employs independent branches to handle both ordinary and extreme weather conditions, constructs explicit correction curves by combining wind direction sector division, and finally achieves high-precision wind speed correction through multimodal data fusion. This method achieves synergistic optimization in spatiotemporal dynamic modeling, extreme weather adaptability, and wind direction spatial heterogeneity characterization, effectively improving the accuracy and stability of wind speed correction.

[0063] The working process and principle of this application are as follows: Measured power curves of each wind turbine generator in the wind farm are obtained, and these curves are preprocessed to obtain the overall power curve of the wind farm. The preprocessing process includes operations such as curve normalization, interpolation, and boundary constraints to ensure data consistency and continuity.

[0064] The historical wind speed for each time period is obtained from the overall power curve, and a time-series correlation is established with the overall wind direction. Simultaneously, meteorological characteristics for the corresponding time periods, such as wind speed, wind direction, temperature, and humidity, are collected to construct a multimodal training dataset. This step aligns and fuses data from different sources over time, providing input for subsequent deep learning models.

[0065] A multimodal training dataset is input into a convolutional neural network to extract local features from the spatiotemporal data of wind speed. Convolutional neural networks can effectively capture local patterns and short-term dependencies in time-series data. The extracted local features are then input into a long short-term memory network to model long-term dynamic dependencies.

[0066] An attention mechanism is introduced into the feature sequence output by the Long Short-Term Memory (LSTM) network to calculate the correlation of features at each time step and obtain attention weights. This step enables the model to dynamically focus on the importance of different time steps. The time-step features are then weighted and converged according to the attention weights to obtain a global feature representation, thereby capturing the key information of the entire sequence.

[0067] Two branches of a multilayer perceptron are configured, one for wind speed prediction correction in ordinary weather and the other for extreme weather. The global feature representation is input into the corresponding multilayer perceptron branch to obtain the prediction values ​​for ordinary weather and extreme weather. This dual-branch structure enables specialized predictions for different weather conditions, improving the model's adaptability.

[0068] Simultaneously, the multimodal training dataset is divided into several levels of circular sectors according to the overall wind direction. Based on the relationship between historical weather forecast wind speed and overall wind speed, correction curves are constructed for each sector. This physics-inspired method can capture local characteristics under different wind directions.

[0069] In the forecast phase, the forecast wind speed for the target time period is input into the correction curves corresponding to the sector containing the target wind direction and its adjacent sectors to obtain the sector correction results. The correction curve prediction values ​​are then obtained by fusing the sector correction results of adjacent sectors using inverse distance weighting. This method considers both spatial continuity and locality.

[0070] The final wind speed correction value is obtained by weighted fusion of the general weather branch forecast, the extreme weather branch forecast, and the correction curve forecast. This application comprehensively considers the advantages of different forecasting strategies to improve the overall forecast accuracy.

[0071] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0072] The measured power curve data of each wind turbine generator in the wind farm is collected through the SCADA system. These curves are preprocessed, including wind speed axis normalization, linear interpolation of power data, outlier removal, and boundary constraints, to obtain the power curve of the entire wind farm.

[0073] Based on the overall power curve, a power-wind speed correspondence is established, and the overall wind speed for historical periods is obtained through reverse lookup and interpolation. Simultaneously, overall wind direction data is acquired, and de-looping and sine / cosine encoding are performed. Meteorological characteristics from the same period are collected, including wind speed, wind direction, temperature, humidity, air pressure, and boundary layer height.

[0074] When constructing the multimodal training dataset, a sliding time window method is used. The meteorological features within the window, the overall wind direction, and derived features are used as inputs, and the overall wind speed corresponding to the window's forward look step size is used as the label. Derived features include wind vector components, wind shear index, etc.

[0075] Training data is input into a one-dimensional temporal convolutional neural network, and local features are extracted using multi-scale convolutional kernels. The feature sequence output by the convolutional network is input into a long short-term memory network to model long-term dependencies. An attention mechanism is applied to the output sequence of the long short-term memory network, and the attention weights at each time step are calculated and weighted to obtain a global feature representation.

[0076] Two multilayer perceptron branches are set up for ordinary weather and extreme weather, each containing multiple fully connected layers and nonlinear activation functions. The global feature representation is input into the two branches to obtain wind speed predictions for ordinary weather and extreme weather, respectively.

[0077] Simultaneously, the wind direction circle is divided into multiple equal-angle sectors, and a correction curve is fitted based on historical data within each sector. During forecasting, the corresponding sector and adjacent sectors are determined according to the wind direction for the target time period. The forecasted wind speed is input into the correction curves of these sectors, and the predicted value of the correction curve is obtained through inverse distance weighted fusion.

[0078] Finally, the predicted values ​​from the general weather branch, the extreme weather branch, and the correction curve are weighted and fused to obtain the final wind speed correction value. The weighting coefficients are dynamically adjusted based on the historical performance indicators of each prediction branch.

[0079] Through the above-described scheme, this application can effectively capture the spatiotemporal dynamics of wind speed and the spatial heterogeneity of wind direction. The combination of convolutional neural networks and long short-term memory networks improves the modeling ability for nonlinear fluctuations and long-term dependencies in wind speed sequences, reducing prediction errors when wind speed changes rapidly or periodic disturbances exist. The introduction of an attention mechanism enables the model to dynamically focus on features at key time steps, improving its responsiveness to abrupt changes.

[0080] A physics-inspired circular sector partitioning method, combined with correction curves fitted from historical data, effectively characterizes the spatial heterogeneity of wind direction within a wind farm. This method can identify local wind speed distribution differences corresponding to different wind directions, improving the model's response speed to sudden changes in wind direction or local eddies.

[0081] The dual-branch forecast structure, which establishes independent forecast paths for ordinary and extreme weather, improves the model's adaptability under different meteorological conditions. In particular, the model's generalization ability is enhanced under extreme weather conditions such as severe convection or boundary layer instability.

[0082] In summary, this application improves the accuracy and reliability of wind speed correction, providing a more accurate wind speed input for wind farm power prediction.

[0083] Some of the solutions described above in this application propose obtaining the overall power curve of the wind farm through measured power curves to support subsequent wind speed correction. However, in practice, the wind speed axis resolution of the measured power curves varies among different wind turbine models, resulting in missing data, duplication, or outliers. Furthermore, the curves are discontinuous in the wind speed boundary regions due to differences in the rated wind speeds of different turbine models. These problems directly affect the accuracy of the overall power curve, thereby reducing the reliability of the subsequent wind speed correction results.

[0084] This application further proposes a method for preprocessing measured power curves to obtain the overall power curve, including: normalizing the measured power curves of each wind turbine along the wind speed axis to unify the wind speed grid; performing linear interpolation on missing power data at corresponding wind speed points and removing duplicate or outlier points; setting boundary constraints at the wind speed boundary according to the turbine model with the highest rated wind speed and performing boundary value interpolation to ensure curve continuity, thereby obtaining the optimized power curve of each turbine; and summing the optimized power values ​​at each unified wind speed point and summing them along the wind speed grid to obtain the overall power curve.

[0085] Interval normalization maps discrete wind speed points from different models to standardized grid nodes by setting a uniform wind speed grid spacing, eliminating differences in the original data resolution. Linear interpolation fills in missing power data in the normalized wind speed grid nodes by linearly fitting the power values ​​of adjacent wind speed points, ensuring data integrity. Outlier removal identifies and removes data points exceeding a reasonable range of power variation relative to wind speed by setting a threshold. Boundary constraints are based on the model with the highest rated wind speed, forcibly setting the rated power value of that model at the upper and lower boundaries of the wind speed grid to avoid data conflicts between different models. Boundary value interpolation uses quadratic spline interpolation to generate a continuous curve within the boundary constraints, ensuring a smooth transition of the power curve throughout the field.

[0086] Specifically, during the interval normalization process, the wind speed axis of the measured power curves of each unit is uniformly divided into grid nodes with intervals of 0.5 m / s, covering the effective wind speed range of all models. For each grid node, if a unit has no measured data at that node, the power values ​​of the two adjacent wind speed points are used for linear interpolation calculation to fill the gap. When multiple power measurements exist at the same wind speed point, the data point with the smallest deviation from the average power is retained, and the rest are discarded as duplicates. At the upper and lower boundaries of the wind speed grid, for example, when the boundary corresponding to the model with the maximum rated wind speed is 25 m / s, the power curves of all models are truncated in the range above 25 m / s, and a continuous curve is generated in the range of 24.5-25 m / s using cubic polynomial interpolation. After the optimization of each unit is completed, the power values ​​of all units at the same wind speed grid node are superimposed to form the overall power curve. For example, at the 15 m / s wind speed node, if the optimized powers of the three units are 3.2 MW, 2.8 MW, and 3.0 MW respectively, then the overall power is 9.0 MW. This process provides high-precision input for subsequent whole-field wind speed calculations by eliminating data noise, standardizing data formats, and enforcing boundary continuity.

[0087] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0088] When preprocessing the measured power curves to obtain the overall power curve, the measured power curves of each wind turbine are first normalized along the wind speed axis to unify the wind speed grid. Specifically, 0.5 m / s can be selected as the uniform wind speed interval, and the wind speed points of the original measured power curves are interpolated onto this unified grid.

[0089] Secondly, linear interpolation is performed on the missing power data at the corresponding wind speed points. For example, if power data is missing at a certain wind speed point, it is filled by linear interpolation of the power values ​​from two adjacent valid wind speed points. At the same time, duplicate or outlier points, such as data points whose power values ​​deviate significantly from the normal range, are removed.

[0090] Furthermore, boundary constraints are set at the wind speed boundary based on the turbine model with the highest rated wind speed. Assuming the turbine model with the highest rated wind speed in the wind farm is 15 m / s, the right boundary of the overall power curve is set to 15 m / s. Boundary value interpolation is then performed to ensure curve continuity; cubic spline interpolation can be used to achieve a smooth transition. Through these steps, the optimized power curves for each turbine are obtained.

[0091] Finally, the optimized power values ​​are summed at each uniform wind speed point, and the summation is performed by traversing the wind speed grid. For example, for each wind speed point at a 0.5 m / s interval, the optimized power values ​​of all units at that wind speed are accumulated to obtain the overall power curve.

[0092] Through the above technical solution, this application achieves standardized processing and integration of the measured power curves of each wind turbine generator set. This eliminates data discrepancies between different units, fills in missing data, removes outliers, and ensures the continuity and smoothness of the curves. Furthermore, by accumulating the optimized power curves of each unit to obtain the overall power curve of the entire field, a reliable data foundation is provided for subsequent wind speed correction. This processing method improves the accuracy and representativeness of the overall power curve, contributing to enhanced accuracy and reliability of wind speed correction.

[0093] In some of the above-mentioned schemes in this application, when obtaining the overall power curve by preprocessing the measured power curve and constructing a multimodal training dataset, the model training data lacks an effective representation of the dynamic change law of wind speed due to the failure to fully integrate the complex relationship between the actual operation data of the wind farm and meteorological characteristics, which affects the accuracy of subsequent predictions.

[0094] This application further proposes to acquire the total power of the wind farm within a historical period and establish a monotonic correspondence between power and wind speed based on the total power curve. The total wind speed at each moment is calculated on the wind speed grid based on reverse lookup and piecewise linear interpolation. The total wind direction representing the wind farm is acquired, and the angle data is de-looped and quantified using sine and cosine encoding, and time-aligned with the total wind speed using a unified sampling step size. Meteorological features on the same time axis as the historical period are collected and normalized. These meteorological features include wind speed, wind direction, temperature, humidity, air pressure, and boundary layer height. Samples are constructed based on a sliding time window, using the meteorological features within the window, the total wind direction, and derived features as inputs. The total wind speed corresponding to the forward look step size of the window is used as a label to form a multimodal training dataset. Weather labels for ordinary and extreme weather are labeled according to wind force level thresholds. Derived features include wind vector components, wind shear index, wind direction turning rate, boundary layer stability index, and boundary layer height change rate.

[0095] Among these methods, reverse lookup and piecewise linear interpolation establish a monotonic relationship between power and wind speed, ensuring the physical consistency of wind speed throughout the field; sine and cosine coding converts wind direction angles into continuous values, avoiding the circular discontinuity of angle data; normalization eliminates differences in meteorological feature dimensions, improving model convergence efficiency; sliding time windows convert time-series data into supervised learning samples, capturing the temporal dependence of wind speed changes; and derived features enhance data representation capabilities through physical derivation, such as wind vector components being decomposed into east-west and north-south components, wind shear index reflecting the vertical wind speed gradient, and boundary layer stability index calculated based on temperature and boundary layer height.

[0096] Specifically, the overall power curve provides a mapping relationship between the wind farm's overall power and wind speed. The corresponding overall wind speed on the wind speed grid is determined through reverse lookup, and piecewise linear interpolation fills in missing values ​​between discrete wind speed points, ensuring a continuous and smooth wind speed sequence. The overall wind direction undergoes de-looping to eliminate abrupt changes between 0° and 360°, and sine and cosine encoding converts angles into sine and cosine values ​​to avoid model bias due to angle periodicity. Meteorological feature normalization uses maximum-minimum scaling to ensure features of different dimensions are within the same numerical range. A sliding time window divides historical time-series data into multiple samples, with the window length and forward step size set according to the wind speed variation cycle; for example, a window length of 6 hours and a forward step size of 1 hour. In derived features, the wind vector component decomposes wind direction and wind speed using trigonometric functions, the wind shear index is calculated based on the wind speed difference at different heights, and the boundary layer stability index jointly characterizes the atmospheric stability state through the temperature gradient and the boundary layer height change rate. Weather labels are categorized based on wind speed thresholds; for example, wind speeds below 17.2 m / s are considered normal weather, while speeds above this threshold are considered extreme weather. These labels are used for subsequent branch training. Thus, the multimodal training dataset, by fusing wind farm operation data, meteorological observation data, and physically derived features, provides the model with multidimensional dynamic information input, improving wind speed correction accuracy.

[0097] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0098] The system acquires the total power over a historical period and establishes a monotonic correlation between power and wind speed based on the total power curve. It then calculates the total wind speed at each moment on a wind speed grid using reverse lookup and piecewise linear interpolation. The system obtains the total wind direction representing the wind farm, de-loops the angle data, and quantifies it using sine and cosine encoding. This data is then time-aligned with the total wind speed using a uniform sampling step.

[0099] Meteorological features were collected and normalized along the same timeline as historical periods. These meteorological features included wind speed, wind direction, temperature, humidity, air pressure, and boundary layer height.

[0100] Samples are constructed using a sliding time window. Meteorological features within the window, along with overall wind direction and derived features, are used as input. The overall wind speed corresponding to the window's forward-looking step length is used as the label, forming a multimodal training dataset. Weather labels for ordinary and extreme weather are then labeled based on wind force level thresholds. Derived features include wind vector components, wind shear index, wind direction turning rate, boundary layer stability index, and boundary layer height change rate.

[0101] Specifically, a 6-hour sliding time window can be used to construct samples with a 10-minute sampling step. For meteorological characteristics, wind speed and direction can use forecast values ​​at a height of 10 meters, temperature and humidity can use forecast values ​​at a height of 2 meters, air pressure can use sea-level air pressure, and boundary layer height can use the output value of the forecast model. Normalization can be performed using the maximum-minimum normalization method.

[0102] The derived features are calculated as follows: the wind vector components are obtained by decomposing the wind speed and wind direction using trigonometric functions; the wind shear index is calculated using the ratio of wind speed at 10 meters to 80 meters; the wind direction turning rate is calculated as the change in wind direction between two adjacent moments; the boundary layer stability index can be calculated using the Richardson number; and the boundary layer height change rate is calculated as the change in boundary layer height between two adjacent moments.

[0103] Weather labels can be defined using the Beaufort wind scale, with winds of force 7 and above defined as extreme weather. When constructing samples, a 36-hour forecast lead time can be used, meaning the sliding window contains data from 24 historical moments, and the label represents the wind speed for the next 12 moments.

[0104] Through the aforementioned technical solutions, this application constructs a training dataset containing multimodal meteorological features, derived features, and label information, providing rich input information for subsequent dynamic spatiotemporal modeling. The time-series features of wind speed are captured by constructing a sliding time window. Simultaneously, derived features are introduced to enhance the characterization of the physical properties of the wind field. Furthermore, the division of weather labels provides the model with prior information to distinguish between ordinary and extreme weather. This multi-dimensional, multi-scale data construction method helps improve the model's ability to learn wind speed variation patterns, thereby enhancing the accuracy and robustness of wind speed correction.

[0105] In some of the solutions described above in this application, the multimodal training dataset contains high-dimensional spatiotemporal features and complex meteorological elements. Directly inputting it into a traditional neural network makes it difficult to effectively extract local spatiotemporal correlation patterns, resulting in the model being unable to accurately capture the local abrupt changes and periodic fluctuations in the dynamic changes of wind speed.

[0106] This application further proposes stacking the multimodal training dataset in the time dimension using a sliding time window and concatenating the meteorological features and the overall wind direction in the feature dimension to form a one-dimensional temporal tensor; using a one-dimensional temporal convolutional neural network, with causal padding in the first layer and multi-scale convolutional kernels and optional dilated convolutions in the middle layers, the kernel length of the multi-scale convolutional kernels being one or more of 3, 5, and 7; after each convolutional layer, nonlinear activation is set sequentially and deep degradation is suppressed through residual connections, finally outputting local features.

[0107] In this architecture, a one-dimensional temporal tensor transforms time-series data into a fixed-length input structure using a sliding window, and integrates multimodal features through concatenation to achieve cross-modal information fusion. Causal padding ensures that convolutional operations rely only on historical data, avoiding the leakage of future information. Multi-scale convolutional kernels capture local features of short, medium, and long periods using different kernel lengths, while dilated convolutions expand the receptive field by adjusting the dilation rate, enhancing the detection capability for sparse events. Residual connections propagate gradients through skip paths, mitigating the gradient vanishing problem in deep network training.

[0108] Specifically, after the multimodal training dataset is segmented by a sliding window, meteorological features and wind direction data are concatenated along the feature dimension to form a one-dimensional temporal tensor, where each time step contains a multi-dimensional feature vector. Causal padding involves zero-padding at the beginning of the convolutional layers to ensure the output length matches the input and depends only on the left-hand input. Multi-scale convolutional layers are used in parallel with kernels of lengths 3, 5, and 7 to extract local patterns across different time spans. For example, a kernel of length 3 captures abrupt changes between adjacent time points, while a kernel of length 7 identifies fluctuations in periodic intervals. Dilated convolutions expand the temporal coverage without increasing the number of parameters by sampling the input data at intervals; for example, with a dilation rate of 2, a kernel of length 5 actually covers 11 time steps. The feature maps output from each convolutional layer are activated by ReLU and then added to the input through residual connections to suppress degradation in deep networks. The final output local features contain multi-scale spatiotemporal patterns, providing a foundation for subsequent long short-term memory network modeling of long-term dependencies.

[0109] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0110] The multimodal training dataset is stacked in a sliding time window along the time dimension, and the meteorological features and overall wind direction are concatenated in the feature dimension to form a one-dimensional temporal tensor. A one-dimensional temporal convolutional neural network is used, with causal padding in the first layer and multi-scale convolutional kernels and optional dilated convolutions in the middle layers. The kernel lengths of the multi-scale convolutional kernels are 3, 5, and 7. Nonlinear activations are applied sequentially after each convolutional layer, and deep degradation is suppressed through residual connections, ultimately outputting local features.

[0111] Specifically, the sliding time window is set to a length of 24 hours with a step size of 1 hour. The shape of the one-dimensional temporal tensor is (batch_size, 24, num_features), where num_features is the total dimension of meteorological features and overall wind direction. The convolutional neural network contains three convolutional layers. The first layer uses causal padding with a kernel size of 3. The second and third layers use kernels with kernel sizes of 5 and 7, respectively. Each convolutional layer is followed by a ReLU activation function, and the input is directly added to the output through residual connections. After the final convolutional layer, local features are obtained through global average pooling.

[0112] Through the above technical solutions, this application can effectively extract local features from spatiotemporal wind speed data. By employing multi-scale convolutional kernels, the model can simultaneously capture wind speed variation patterns at different time scales. Causal imputation ensures that the model uses only current and past information for prediction, avoiding information leakage. Residual connections help alleviate the vanishing gradient problem, enabling deeper networks to be trained more effectively. This design allows the model to more accurately characterize the spatiotemporal correlation of wind speed, thereby improving the accuracy of wind speed correction.

[0113] In some of the schemes described above in this application, when modeling long-term dynamic dependencies using long short-term memory networks, the contribution of features at different time steps in the hidden state sequence to the current prediction is not effectively quantified. This may result in key time step features being averaged out, failing to fully capture key fluctuation patterns in the dynamic changes of wind speed, affecting the ability of global feature representation to characterize complex spatiotemporal relationships, and thus reducing the accuracy of wind speed correction.

[0114] This application further proposes to model long-term dynamic dependencies by stacking local feature inputs into a long short-term memory network to obtain a hidden state sequence; calculate the relevance score based on the attention mechanism and obtain the attention weights through Softmax normalization; weight and converge the time-step features according to the attention weights to obtain the context vector, and obtain the global feature representation by nonlinear projection of the context vector.

[0115] Among them, the stacked long short-term memory network adopts a multi-layer structure to enhance the sequence modeling ability. Each time step in the hidden state sequence corresponds to a feature vector. The attention mechanism measures the relevance by calculating the dot product score between the query vector and the key vector of each time step. The score is converted into attention weights of probability distribution after exponential operation and normalization. The weighted convergence process generates a context vector by linearly combining the feature vectors of each time step according to the weights. The nonlinear projection uses fully connected layers and activation functions to perform dimensionality reduction or dimensionality increase transformation on the context vector.

[0116] Specifically, the feature vector at each time step in the hidden state sequence is mapped to a query vector, a key vector, and a value vector. After calculating the dot product score of the query vector and all key vectors, the score is converted into attention weights using the Softmax function. The value vectors at each time step are weighted and summed to obtain the context vector, which integrates the contributions of different time steps in the sequence. The nonlinear projection layer maps the context vector to the target dimension space, forming a global feature representation containing information from key time steps. Through the attention weight allocation mechanism, the model can adaptively focus on historical time step features that are more relevant to the current prediction, suppressing noise interference. At the same time, the nonlinear projection enhances the nonlinear expressive power of the features, enabling the global feature representation to more accurately characterize the spatiotemporal evolution of wind speed, providing a robust feature foundation for subsequent wind speed correction.

[0117] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0118] The hidden state sequence is obtained by modeling long short-term memory (LSTM) networks that stack local feature inputs to form long-term dynamic dependencies. Specifically, a two-layer LSM network is used, with each layer containing 128 hidden units. The input of the first LSM network is the local feature sequence, and its output serves as the input of the second layer. The output of the second LSM network is the hidden state sequence.

[0119] The relevance score is calculated based on an attention mechanism, and the attention weights are obtained through Softmax normalization. Further, an additive attention mechanism is employed, using a feedforward neural network to calculate the relevance score between the query vector and the hidden state at each time step. The query vector is obtained from the hidden state at the last time step through a linear transformation. The attention weights are obtained by normalizing the relevance score using the Softmax function.

[0120] The context vector is obtained by weighted aggregation of time-step features based on attention weights, and then a global feature representation is obtained by nonlinear projection of the context vector. Alternatively, the context vector is obtained by weighted summation of the hidden state sequence using attention weights. This context vector is then subjected to a nonlinear transformation through a single-layer feedforward neural network to obtain the final global feature representation.

[0121] Through the above technical solution, this application can effectively capture long-term dependencies in wind speed time series and adaptively focus on information at important time steps through an attention mechanism. The global feature representation comprehensively considers the dynamic changes in historical wind speed data, improving the accuracy and robustness of wind speed correction. Simultaneously, the attention mechanism enhances the model's interpretability, facilitating the analysis of key time points affecting wind speed prediction.

[0122] In some of the schemes mentioned above in this application, the multilayer perceptron structure may cause conflict in the direction of model parameter updates during training due to the difference in the distribution of ordinary weather and extreme weather samples. The sparsity of extreme weather samples makes it difficult for the model to fully learn its unique patterns, thereby affecting the accuracy of wind speed prediction under extreme weather conditions.

[0123] This application further proposes setting up two multilayer perceptron branches, including a general weather multilayer perceptron branch and an extreme weather multilayer perceptron branch; each multilayer perceptron branch includes at least two fully connected layers and nonlinear activation, and random deactivation is set between layers to suppress overfitting; during the training phase, the parameters of the corresponding multilayer perceptron branch are updated only based on the weather label while the other multilayer perceptron branch is frozen; during the prediction phase, the global feature representation is fed forward into the two multilayer perceptron branches respectively to simultaneously generate the prediction values ​​of the general weather branch and the extreme weather branch; quantile loss is used as a robust loss function to constrain the training error.

[0124] The fully connected layers of the ordinary weather branch and the extreme weather branch use the same network depth and width, but the initialization parameters are independent; the random inactivation rate is set between 0.2 and 0.5 and dynamically adjusted during training; the quantile loss function selects three quantiles of 0.05, 0.5 and 0.95 to calculate the weighted error; the weather labels are divided according to the wind force level threshold, and when the measured wind speed exceeds 20 m / s, it is marked as an extreme weather sample.

[0125] Specifically, during the training phase, when the weather label of the input sample is "normal weather," only the fully connected layer parameters of the normal weather branch are updated with gradients, while the parameters of the extreme weather branch remain frozen; conversely, for extreme weather samples, only the parameters of the extreme weather branch are updated. The two branches learn the wind speed error distribution patterns under normal and extreme weather modes through independent optimization processes. During the prediction phase, the global feature representation is simultaneously input into both branches, each outputting its corresponding predicted value. The quantile loss function enhances the model's robustness to abnormal wind speed values ​​by minimizing the deviation between the predicted and true values ​​at multiple quantiles. Random deactivation randomly disables some neurons in the hidden layer, forcing the network to learn redundant feature representations and avoiding overfitting to specific weather patterns. This technical solution effectively mitigates learning interference between different weather patterns and improves prediction stability under extreme weather conditions through branch decoupling and parameter isolation mechanisms.

[0126] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0127] Two multilayer perceptron branches are configured: one for general weather and one for extreme weather. Each branch consists of three fully connected layers and nonlinear activations, with random deactivation applied between layers to suppress overfitting. During training, parameters are updated only for the corresponding branch based on weather labels, while the other branch is frozen. During prediction, the global feature representation is fed forward into both branches to simultaneously generate predictions for both general and extreme weather conditions. Quantile loss is used as a robust loss function to constrain the training error.

[0128] Specifically, the three fully connected layers of the multilayer perceptron branch for ordinary weather have 128, 64, and 32 neurons, respectively, while the three fully connected layers of the multilayer perceptron branch for extreme weather have 64, 32, and 16 neurons, respectively. Both branches use ReLU as the non-linear activation function, with a random deactivation layer with a dropout rate of 0.3 set between the first and second layers. During training, the corresponding branch is selected for backpropagation based on the weather label of the sample, while the parameters of the unselected branches remain unchanged. During prediction, the global feature representation is simultaneously input into both branches to obtain two prediction results. The quantile loss function is set to 0.5 quantile, i.e., median regression, to enhance the model's robustness to outliers.

[0129] Through the above technical solutions, this application can construct specialized prediction models for both ordinary and extreme weather, improving the model's adaptability and prediction accuracy under different weather conditions. Simultaneously, by employing a quantile loss function, the model's robustness to outliers is enhanced, improving the reliability of the prediction results. Furthermore, by selectively updating parameters during the training phase, interference from irrelevant samples is avoided, improving the model's generalization ability.

[0130] In some of the solutions mentioned above in this application, a method for constructing correction curves by dividing circular sectors was proposed. However, in scenarios with spatial heterogeneity and uneven wind speed distribution dominated by wind direction, the existing methods have failed to effectively combine wind speed level and wind direction angle for hierarchical division, resulting in insufficient adaptability of the correction curves in local areas and difficulty in accurately capturing the dynamic relationship between different wind speed ranges and wind direction sectors.

[0131] This application further proposes to divide the wind direction circumference into circular sectors of equal width using the overall wind direction as the angle variable, and to divide the sector extension into several rings according to the wind speed level to form a hierarchical structure; within each circular sector, a correction curve is obtained by monotonic constrained regression fitting with historical weather forecast wind speed as the independent variable and overall wind speed as the dependent variable.

[0132] The wind direction circle is divided into multiple sectors with the same angle. The angle range of each sector is set according to actual needs, for example, each sector covers 30 degrees. The outer extension of the sector is further divided into multiple rings according to wind speed levels, such as dividing wind speed into levels of 0-5 m / s, 5-10 m / s, and 10-15 m / s. Within each sub-region combining a sector and a ring, an ordinal regression or monotonic spline fitting method is used to establish a mapping relationship between the forecast wind speed and the overall wind speed, ensuring that the correction curve changes monotonically with the input wind speed. The sector division and ring setting form a two-dimensional grid structure, with each grid corresponding to a specific wind direction range and wind speed interval, allowing the correction curve to simultaneously consider differences in wind direction angle and changes in wind speed intensity.

[0133] Specifically, when constructing the correction curve, the sector is first determined based on the wind direction angle in historical data, and then the corresponding annular zone level is determined based on the forecast wind speed value. Within the same sector and annular zone, historical forecast wind speed is used as the input variable, and the overall wind speed is used as the target variable. A constrained regression algorithm is used to fit the monotonic relationship between the two to generate the correction curve for each sub-region. For example, in the annular zone with a northeast wind direction of 30 degrees and a wind speed of 5-10 m / s, piecewise linear regression is used to force the slope to remain non-negative, ensuring that the overall wind speed prediction value increases synchronously when the forecast wind speed increases. This combination of hierarchical division and monotonic constraints can effectively adapt to the non-uniform characteristics of wind speed distribution under different wind directions, improving the local adaptability and physical consistency of the correction curve.

[0134] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0135] Using the overall wind direction as the angular variable, the wind direction circumference is divided into circular sectors of equal width. Furthermore, the outer perimeter of each sector is further divided into several rings according to wind speed levels, forming a hierarchical structure. Specifically, the wind direction circumference is divided into eight sectors with a width of 45 degrees. Each sector is further divided into three rings, corresponding to low, medium, and high wind speed ranges, respectively.

[0136] Within each circular sector, a correction curve is obtained by performing monotonic constrained regression fitting with historical weather forecast wind speed as the independent variable and the overall wind speed as the dependent variable. Furthermore, an ordinal-preserving regression algorithm is used to fit the historical data within each sector to ensure the monotonicity of the correction curve. Thus, a dedicated correction curve can be obtained for each sector and wind speed interval combination.

[0137] For example, for the medium wind speed range within a 45-degree northbound sector, corresponding data points of historical weather forecast wind speeds and actual wind speeds are collected. These data points are then fitted using an ordinal-preserving regression algorithm to obtain a monotonically increasing correction curve. This curve can be used to map the forecast wind speeds within the sector to more accurate estimates of actual wind speeds.

[0138] Through the above technical solutions, this application achieves refined correction based on wind direction and wind speed. By considering the spatial heterogeneity of wind direction, the correction curve can better capture the local topographic effects and wind field characteristics under different wind directions. Simultaneously, the hierarchical correction structure improves the model's adaptability to different wind speed ranges. This physically inspired explicit correction mechanism enhances the model's interpretability and stability, making the correction results more reliable and easier to understand. Furthermore, this method constructs correction curves through historical statistical relationships, reducing dependence on large amounts of training data and improving the model's robustness under data sparsity conditions.

[0139] In some of the above-mentioned schemes in this application, when the multimodal training dataset is divided into several levels of circular sectors according to the overall wind direction and the correction curve of each sector is constructed, only the correction curve of the sector where the target wind direction is located is considered for wind speed prediction. The influence of adjacent sectors on the target wind direction is not considered. As a result, when there are natural fluctuations in wind direction or measurement errors, the correction result of a single sector is prone to deviate from the actual wind speed distribution, and the prediction stability is insufficient.

[0140] This application further proposes to determine the target circular sector corresponding to the overall wind direction during the target period and its adjacent circular sectors, substitute the meteorological forecast wind speed for the target period into the corresponding correction curve to obtain multiple sector correction results; calculate the inverse distance weight based on the angle between the central axis of each sector and the overall wind direction, and obtain the correction curve prediction value by weighted summation of the sector correction results.

[0141] Specifically, when determining the target circular sector and its adjacent sectors, matching is performed within the pre-divided circular sectors based on the angle value of the overall wind direction. Adjacent sectors include two sectors that are angularly adjacent to the target sector. When substituting into the correction curve, the correction curve for each sector is constructed based on the monotonic constrained regression relationship between historical weather forecast wind speed and overall wind speed, ensuring that the corresponding correction result can be directly output after inputting the weather forecast wind speed. When calculating the inverse distance weight, the absolute value of the angle between the central axis of each sector and the overall wind direction is used as the distance metric, and the weight coefficient is calculated using the inverse distance weighting formula, with the weight being inversely proportional to the angle. When performing weighted summation, the correction result of each sector is multiplied by its corresponding inverse distance weight and then summed to obtain the comprehensive correction curve prediction value.

[0142] Specifically, after obtaining the overall wind direction during the target time period, the target sector is determined within a pre-divided circular sector based on this angle value, and its two adjacent sectors on the left and right are selected as adjacent sectors. The forecast wind speed for the target time period is input into the correction curves corresponding to the target sector and its two adjacent sectors to the left and right, respectively, to obtain the correction results for three sectors. The absolute value of the angle between the central axis of each sector and the overall wind direction is calculated. For example, if the central axis of the target sector is 30 degrees and the overall wind direction is 28 degrees, the angle is 2 degrees; if the central axis of the adjacent left sector is 15 degrees, the angle is 13 degrees; and if the central axis of the adjacent right sector is 45 degrees, the angle is 17 degrees. The weights are calculated using an inverse distance weighting formula. For example, the weight coefficient is 1 / (1+angle). After normalization, the weight of the target sector is 0.65, the weight of the left adjacent sector is 0.22, and the weight of the right adjacent sector is 0.13. The correction results of the three sectors are multiplied by their corresponding weights and then summed to obtain the final predicted value of the correction curve. By incorporating correction results from adjacent sectors and performing weighted fusion based on spatial proximity, the prediction error caused by angle deviation in a single sector is effectively reduced, thereby improving the robustness and accuracy of the correction results.

[0143] As a preferred embodiment, the specific implementation of this application is as follows: When the overall wind direction during the target time period is 90 degrees, the target circular sector corresponding to this wind direction is first determined. Assuming the wind direction circumference is divided into 16 sectors of equal width, the target sector corresponds to the range of 86.25 degrees to 93.75 degrees. Adjacent sectors are defined as two sectors adjacent to the boundary of the target sector, corresponding to the areas of 78.75 degrees to 86.25 degrees and 93.75 degrees to 101.25 degrees. The meteorological forecast wind speed for the target time period is input into the correction curve corresponding to the target sector, and simultaneously into the correction curves of the two adjacent sectors, resulting in correction results for three sectors. The angle between the central axis of each sector and the 90-degree wind direction is calculated; for example, if the central axis of the target sector is 90 degrees, the central axes of the adjacent sectors are 82.5 degrees and 97.5 degrees, respectively. An inverse distance weight is calculated based on the absolute value of the angle, where the weight is inversely proportional to the absolute value of the angle. The correction results of the three sectors are weighted and summed to obtain the predicted value of the correction curve.

[0144] Through the above technical solution, this application effectively reduces the impact of local wind direction shifts on wind speed correction. By fusing the correction results of adjacent sectors, it suppresses boundary abrupt errors that may be caused by a single sector division. This scheme can adaptively balance the historical statistical patterns of adjacent wind direction areas, enhance the model's robustness to wind direction measurement errors and short-term fluctuations, and improve the spatial continuity of the correction results.

[0145] In some of the solutions mentioned above in this application, a method is proposed to divide circular sectors according to the overall wind direction and construct correction curves. However, when there are fluctuations in wind direction or measurement errors, the overall wind direction during the target period may deviate from the central axis of the sector, resulting in the correction curve of a single sector failing to accurately reflect the actual wind speed distribution, and the sudden changes in wind speed between adjacent sectors are not effectively smoothed out.

[0146] This application further proposes to determine the target circular sector corresponding to the overall wind direction during the target period and its adjacent circular sectors, substitute the meteorological forecast wind speed for the target period into the corresponding correction curve to obtain multiple sector correction results; calculate the inverse distance weight based on the angle between the central axis of each sector and the overall wind direction, and obtain the correction curve prediction value by weighted summation of the sector correction results.

[0147] The determination of adjacent sectors is based on the equal-angle width of the wind direction circumference division. Two adjacent sectors are set, located on the left and right sides of the target sector. The inverse distance weight calculation uses the cosine square function of the angle difference, with the weight coefficient proportional to the square of the cosine of the angle. During the fusion of sector correction results, when the target wind direction is located at the boundary between two adjacent sectors, a weight allocation threshold of 15 degrees is set. When the angle between the target wind direction and the central axis of any sector exceeds 45 degrees, only the two sectors with the highest weights are retained for calculation.

[0148] Specifically, the angle between the target wind direction and the central axis of each sector is calculated using a vector dot product, and the square of the cosine of the calculated angle is used as the initial weight. The initial weights are normalized to ensure that the sum of the weights for each sector is 1. For example, when the angle between the target wind direction and the central axis of the main sector is 10 degrees, the main sector has a weight of 0.85, and the adjacent left and right sectors receive weights of 0.10 and 0.05 respectively. This weighting method ensures a smooth transition of the correction results from neighboring sectors to the final predicted value, effectively suppressing prediction jumps caused by wind direction measurement deviations or short-term fluctuations. Through inverse distance weighted fusion, the dominant role of the main sector correction curve is preserved, while data from adjacent sectors is used to correct prediction errors in boundary areas, improving correction stability under complex wind conditions.

[0149] As a preferred embodiment, the specific implementation of this application is as follows: In the wind speed correction process of a wind farm, for the fusion stage of the general weather branch prediction, the extreme weather branch prediction, and the correction curve prediction, the mean absolute error and root mean square error of each prediction value on the validation set are first calculated as performance indicators. Further, the analytic hierarchy process (AHP) is used to determine subjective weights, where the subjective weight of the general weather branch prediction is set to 0.4, the subjective weight of the extreme weather branch prediction is set to 0.3, and the subjective weight of the correction curve prediction is set to 0.3. Simultaneously, the objective weights of each prediction value are calculated based on the entropy weight method, where the objective weight of the general weather branch prediction is 0.35, the objective weight of the extreme weather branch prediction is 0.25, and the objective weight of the correction curve prediction is 0.4. The subjective and objective weights are linearly combined in a 1:1 ratio, and after normalization, the combined weights are 0.375, 0.275, and 0.35, respectively. Finally, the fused wind speed correction value is output by linearly weighting the combined weights and the three prediction values.

[0150] Through the above technical solution, this application effectively balances the complementarity between data-driven deep learning prediction results and physical statistics-based correction curve prediction results, solving the problem of large prediction bias in single models under complex weather conditions. By introducing a subjective and objective weight combination mechanism, it retains the qualitative judgment of model performance by expert experience while fully utilizing the statistical characteristics of data to quantitatively adjust the weight allocation, enhancing the model's adaptability in scenarios where ordinary and extreme weather alternate. This fusion method reduces the fluctuation of prediction errors caused by abnormal meteorological events and improves the continuity of wind speed correction results in the temporal dimension and the stability in the spatial dimension.

[0151] In the above embodiments, the measured power curves of each wind turbine in the wind farm are obtained and preprocessed to obtain the overall power curve. This allows for the inversion of the overall wind speed and its time-series pairing with the overall wind direction, constructing a multimodal training dataset encompassing meteorological characteristics, thus achieving global dynamic perception of the wind farm's operating status. A convolutional neural network is used to extract local features from the spatiotemporal wind speed data, and a long short-term memory network is used to capture its time-dependent characteristics. An attention mechanism is introduced into the output sequence to weight and converge the importance of each time step, enhancing the model's ability to model wind speed trends. Two multilayer perceptron branches are set up for ordinary weather and extreme weather, enabling the learning of prediction strategies under different meteorological conditions and improving the response to extreme weather events. Spatially, the training dataset is divided into circular sectors based on the overall wind direction. Correction curves for each sector are constructed based on the relationship between historical forecast wind speed and measured wind speed. An inverse distance weighting strategy is used to fuse the correction results of neighboring sectors, generating physically meaningful correction curve prediction values. By integrating general weather forecasts, extreme weather forecasts, and correction curve forecasts, the final wind speed correction value is obtained, which improves the correction accuracy and robustness, enhances the interpretability and engineering applicability of the model, and makes up for the shortcomings of existing technologies in spatiotemporal dynamic modeling, wind direction zoning response, and extreme case adaptation.

[0152] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a wind farm wind speed correction system based on dynamic spatiotemporal modeling, used to apply the above-mentioned wind farm wind speed correction method based on dynamic spatiotemporal modeling, including:

[0153] The acquisition unit is configured to acquire the measured power curves of each wind turbine generator in the wind farm and preprocess the measured power curves to obtain the power curve of the entire farm.

[0154] The association unit is configured to obtain the wind speed of the entire field for a historical period based on the overall power curve, and to correlate it with the wind direction over time. At the same time, it collects meteorological features of the corresponding period to construct a multimodal training dataset.

[0155] The first processing unit is configured to input the multimodal training dataset into a convolutional neural network to extract local features of wind speed spatiotemporal data, input the local features into a long short-term memory network, introduce an attention mechanism into the feature sequence output by the long short-term memory network, calculate the correlation of features at each time step and obtain attention weights, and perform weighted aggregation of time step features according to the attention weights to obtain a global feature representation.

[0156] The second processing unit is configured to set up two multilayer perceptron branches, input the global feature representation into the corresponding multilayer perceptron branches to perform wind speed prediction correction, and obtain the prediction values ​​of ordinary weather branch and extreme weather branch.

[0157] The third processing unit is configured to divide the multimodal training dataset into several levels of circular sectors according to wind direction, construct correction curves for each sector based on the relationship between historical forecast wind speed and actual wind speed, input the meteorological forecast wind speed of the target time period into the correction curves corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain sector correction results, and obtain correction curve prediction values ​​by fusing the sector correction results of adjacent sectors through inverse distance weighting.

[0158] The correction unit is configured to obtain the final wind speed correction value by weighted fusion of the general weather branch forecast value, the extreme weather branch forecast value and the correction curve forecast value.

[0159] Specifically, the data acquisition unit normalizes the measured power curves of multiple units using a unified wind speed grid, performs linear interpolation and boundary constraint interpolation along the wind speed axis, and sums the results along the grid to generate the overall power curve. The association unit aligns meteorological features with the overall wind direction using a sliding time window, constructs a multimodal training dataset containing derived features, and labels it with weather tags. The first processing unit uses a one-dimensional temporal convolutional neural network to extract local features, models long-term dependencies using a long short-term memory network, and introduces an attention mechanism to generate global feature representations. The second processing unit sets up two independent multilayer perceptron branches, activates the corresponding branches based on weather tags for prediction, and configures fully connected layers and randomly deactivated layers for ordinary weather and extreme weather, respectively. The third processing unit divides the data into hierarchical sectors according to wind direction to construct correction curves, and generates correction curve prediction values ​​by fusing adjacent sector results using inverse distance weighting. The correction unit calculates combined weights based on performance indicators and performs weighted fusion of the multi-branch prediction results.

[0160] Specifically, the acquisition unit receives measured power data uploaded by each wind turbine, performs grid alignment on the wind speed axis to eliminate boundary discontinuities caused by differences in rated wind speeds between different turbine models, fills in missing data points through interpolation, and finally outputs the overall power curve. The correlation unit simultaneously receives meteorological monitoring data and overall wind direction data, performs sine and cosine encoding on the wind direction to eliminate angle looping errors, constructs a multimodal input matrix based on temperature, humidity, and air pressure, and generates timestamped training samples based on a sliding window. The first processing unit converts the input matrix into a temporal tensor, extracts local wind speed fluctuation patterns through multi-scale convolutional kernels, uses residual connections to prevent gradient vanishing, and selects key time step features through an attention mechanism to form a global representation. The second processing unit dynamically freezes irrelevant branch parameters based on weather labels during the training phase, and performs parallel predictions for ordinary and extreme weather during the prediction phase, suppressing outlier interference through a quantile loss function. The third processing unit establishes a wind speed correction mapping relationship for each sector based on historical data, matches adjacent sectors based on wind direction and angle during real-time prediction, and uses inverse distance weighting to fuse spatial proximity effects. The correction unit dynamically adjusts the weight allocation by calculating the prediction error rate of each branch, and finally outputs the fused wind speed correction value. The system achieves parallel processing of data acquisition, feature extraction, multi-model prediction and physical correction through modular design. While improving the response speed to extreme weather, it balances the advantages of statistical learning and physical models through a weight fusion mechanism.

[0161] As a preferred embodiment, the solution of this application is implemented as follows: The wind speed correction system for wind farms includes an acquisition unit, an association unit, a first processing unit, a second processing unit, a third processing unit, and a correction unit. The acquisition unit obtains the measured power curves from each wind turbine generator in the wind farm and generates the overall power curve after preprocessing. The association unit deduces the overall wind speed for historical periods based on the overall power curve, aligns it with the overall wind direction in time, and integrates meteorological features to construct a multimodal training dataset. The first processing unit inputs the dataset into a convolutional neural network to extract local spatiotemporal features, models long-term dependencies through a long short-term memory network, and introduces an attention mechanism into the feature sequence to calculate the weights of each time step. After weighted aggregation, a global feature representation is formed. The second processing unit sets up two multilayer perceptron branches to perform wind speed prediction correction for ordinary weather and extreme weather, respectively. During training, the parameters of irrelevant branches are frozen, and during prediction, the predicted values ​​under the two types of weather conditions are output in parallel. The third processing unit divides the data into multi-level circular sectors based on the overall wind direction. It fits correction curves for each sector based on the relationship between historical weather forecast wind speeds and actual wind speeds. After inputting the correction curves for the corresponding and adjacent sectors with the forecast wind speed for the target time period, it uses inverse distance weighted fusion to obtain the predicted value of the correction curve. The correction unit then weights and fuses the predicted values ​​from the two multi-layer sensor branches with the predicted value of the correction curve according to combined weights to generate the final corrected wind speed value.

[0162] Through the above technical solutions, this application effectively solves the problems of insufficient spatiotemporal dynamic modeling, lack of handling of wind direction spatial heterogeneity, and weak response capability to extreme weather in existing technologies. By using multimodal data fusion and attention mechanism weighting, the ability to capture the spatiotemporal evolution law of wind speed is enhanced; by adopting sector-based correction curves and inverse distance weighting strategies, the adaptability to spatial heterogeneity under different wind directions is improved; by using a dual-branch multilayer perceptron structure, differentiated processing of ordinary and extreme weather is achieved, enhancing the prediction stability under extreme conditions; and the combined weighting mechanism of multiple prediction results further improves the robustness and generalization ability of the correction results.

[0163] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0164] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0165] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0166] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A wind speed correction method for wind farms based on dynamic spatiotemporal modeling, characterized in that, include: The measured power curves of each wind turbine generator in the wind farm are obtained, and the measured power curves are preprocessed to obtain the power curve of the entire farm. The wind speed for historical periods is obtained from the overall power curve and correlated with the wind direction over time. At the same time, meteorological features for the corresponding periods are collected to construct a multimodal training dataset. The multimodal training dataset is input into a convolutional neural network to extract local features of wind speed spatiotemporal data, and the local features are input into a long short-term memory network. An attention mechanism is introduced into the feature sequence output by the long short-term memory network to calculate the correlation of features at each time step and obtain attention weights. The time step features are weighted and converged according to the attention weights to obtain a global feature representation. Two multilayer perceptron branches are set up. The global feature representation is input into the corresponding multilayer perceptron branch to perform wind speed prediction correction, and the prediction values ​​of ordinary weather branch and extreme weather branch are obtained. The multimodal training dataset is divided into several levels of circular sectors according to the overall wind direction. Correction curves for each sector are constructed based on the relationship between historical weather forecast wind speed and the overall wind speed. The weather forecast wind speed for the target time period is input into the correction curves of the sector where the target wind direction is located and its adjacent sectors to obtain the sector correction results. The correction curve prediction value is obtained by fusing the sector correction results of the adjacent sectors through inverse distance weighting. The final wind speed correction value is obtained by weighted fusion of the general weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value. Two multilayer perceptron branches are set up. The global feature representation is input into the corresponding multilayer perceptron branch for wind speed prediction correction. When obtaining the predicted values ​​of ordinary weather branch and extreme weather branch, the following steps are taken: When setting up two multilayer sensor branches, there are two branches: one for general weather and one for extreme weather. Each of the multilayer perceptron branches includes at least two fully connected layers and nonlinear activation, and random deactivation is set between layers to suppress overfitting; during the training phase, the parameters of the corresponding multilayer perceptron branch are updated only based on the weather label while the other multilayer perceptron branch is frozen; during the prediction phase, the global feature representation is fed forward into the two multilayer perceptron branches respectively to simultaneously generate the predicted values ​​of the ordinary weather branch and the extreme weather branch; quantile loss is used as a robust loss function to constrain the training error; When dividing the multimodal training dataset into several levels of circular sectors based on the overall wind speed, and constructing correction curves for each sector based on the relationship between historical weather forecast wind speeds and the overall wind speed, the process includes: Using the overall wind direction as an angle variable, the wind direction circumference is divided into circular sectors of equal width, and the outer extension of the sector is divided into several rings according to the wind speed level to form a hierarchical structure. Within each circular sector, the correction curve is obtained by performing monotonic constrained regression fitting with historical weather forecast wind speed as the independent variable and the overall wind speed as the dependent variable.

2. The wind speed correction method for wind farms based on dynamic spatiotemporal modeling according to claim 1, characterized in that, Preprocessing the measured power curve to obtain the overall power curve includes: The measured power curves of each wind turbine generator set are normalized at intervals along the wind speed axis to unify the wind speed grid. Linear interpolation is performed on the missing power data at the corresponding wind speed points and duplicate or outlier points are removed. Boundary constraints are set at the wind speed boundary according to the model with the largest rated wind speed and boundary value interpolation is performed to ensure the continuity of the curves, thus obtaining the optimized power curves of each unit. At each uniform wind speed point, the optimized power values ​​are added together, and the summation is obtained by traversing the wind speed grid to obtain the overall power curve.

3. The wind speed correction method for wind farms based on dynamic spatiotemporal modeling according to claim 2, characterized in that, The historical wind speed for the entire field is obtained based on the overall power curve, and time-series correlation is performed with the overall wind direction. Simultaneously, meteorological characteristics for the corresponding time periods are collected. When constructing a multimodal training dataset, the following steps are included: The total power of the wind farm is obtained within a historical period, and a monotonic correspondence between power and wind speed is established based on the total power curve. The total wind speed at each moment is obtained on the wind speed grid based on reverse lookup and piecewise linear interpolation. The total wind direction representing the wind farm is obtained, the angle data is de-looped and quantified using sine and cosine encoding, and time-aligned with the total wind speed at a uniform sampling step size. Meteorological features are collected on the same time axis as historical periods and normalized. The meteorological features include wind speed, wind direction, temperature, humidity, air pressure and boundary layer height. Samples are constructed based on a sliding time window. The meteorological features within the window, the overall wind direction, and derived features are used as inputs. The overall wind speed corresponding to the forward step length of the window is used as a label to form the multimodal training dataset. Weather labels for ordinary weather and extreme weather are labeled according to wind force level thresholds. The derived features include wind vector components, wind shear index, wind direction turning rate, boundary layer stability index, and boundary layer height change rate.

4. The wind speed correction method for wind farms based on dynamic spatiotemporal modeling according to claim 3, characterized in that, When inputting the multimodal training dataset into a convolutional neural network to extract local features of wind speed spatiotemporal data, the following steps are included: The multimodal training dataset is stacked in the time dimension according to the sliding time window, and the meteorological features and the overall wind direction are concatenated in the feature dimension to form a one-dimensional time series tensor. A one-dimensional temporal convolutional neural network is used, with causal padding in the first layer and multi-scale convolutional kernels and optional dilated convolutions in the middle layers. The kernel length of the multi-scale convolutional kernels is one or more of 3, 5, and 7. After each convolutional layer, nonlinear activation is sequentially applied, and deep degradation is suppressed through residual connections, ultimately outputting the local features.

5. The wind speed correction method for wind farms based on dynamic spatiotemporal modeling according to claim 4, characterized in that, Calculating the correlation of features at each time step and obtaining attention weights, and then weighting and converging the features at each time step according to the attention weights to obtain a global feature representation, includes: The local features are input into a stacked long short-term memory network to model long-term dynamic dependencies and obtain a sequence of hidden states. The relevance score is calculated based on the attention mechanism, and the attention weight is obtained by Softmax normalization. The time-step features are weighted and converged according to the attention weights to obtain a context vector, and the context vector is then nonlinearly projected to obtain the global feature representation.

6. The wind speed correction method for wind farms based on dynamic spatiotemporal modeling according to claim 1, characterized in that, When inputting the forecasted wind speed for the target time period into the correction curves corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain the sector correction results, and then obtaining the correction curve prediction value by inverse distance weighted fusion of the sector correction results of the adjacent sectors, the process includes: The target circular sector and its adjacent circular sectors are determined according to the overall wind direction during the target time period. The meteorological forecast wind speed during the target time period is substituted into the corresponding correction curve to obtain multiple sector correction results. The inverse distance weight is calculated based on the angle between the central axis of each sector and the overall wind direction. The sector correction results are weighted and summed to obtain the predicted value of the correction curve.

7. The wind speed correction method for wind farms based on dynamic spatiotemporal modeling according to claim 6, characterized in that, When obtaining the final wind speed correction value by weighted fusion of the general weather branch forecast value, the extreme weather branch forecast value, and the correction curve forecast value, the following are included: Performance indices are calculated for the general weather branch forecast, the extreme weather branch forecast, and the correction curve forecast, and subjective and objective weights are determined based on the performance indices. Weight normalization and non-negativity constraints are used to obtain combined weights, where the sum of all weights in the combined weights is 1. The final wind speed correction value is calculated based on the combined weights, the general weather branch forecast, the extreme weather branch forecast, and the correction curve forecast.

8. A wind farm wind speed correction system based on dynamic spatiotemporal modeling, used to apply the wind farm wind speed correction method based on dynamic spatiotemporal modeling as described in any one of claims 1-7, characterized in that, include: The acquisition unit is configured to acquire the measured power curves of each wind turbine generator in the wind farm, and to preprocess the measured power curves to obtain the power curves of the entire farm. The association unit is configured to obtain the wind speed of the entire field for a historical period based on the overall power curve, and to associate it with the wind direction over time. At the same time, it collects meteorological features of the corresponding period and constructs a multimodal training dataset. The first processing unit is configured to input the multimodal training dataset into a convolutional neural network to extract local features of wind speed spatiotemporal data, input the local features into a long short-term memory network, introduce an attention mechanism into the feature sequence output by the long short-term memory network, calculate the correlation of features at each time step and obtain attention weights, and perform weighted aggregation of the time step features according to the attention weights to obtain a global feature representation. The second processing unit is configured to set up two multilayer perceptron branches, input the global feature representation into the corresponding multilayer perceptron branches to perform wind speed prediction correction, and obtain the prediction value of ordinary weather branch and the prediction value of extreme weather branch. The third processing unit is configured to divide the multimodal training dataset into several levels of circular sectors according to wind direction, construct correction curves for each sector based on the relationship between historical forecast wind speed and actual wind speed, input the meteorological forecast wind speed of the target time period into the correction curves corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain sector correction results, and obtain correction curve prediction values ​​by fusing the sector correction results of the adjacent sectors through inverse distance weighting. The correction unit is configured to obtain the final wind speed correction value by weighted fusion of the general weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value.

Images