Wind power plant wind speed correction method and system based on dynamic space-time modeling

Abstract

The invention relates to the technical field of wind power generation, and discloses a wind power plant wind speed correction method and system based on dynamic space-time modeling, and the method comprises the steps: obtaining a whole power curve, obtaining the whole wind speed of a historical period, and constructing a multi-modal training data set; inputting a convolutional neural network to extract local features, inputting a long-short-term memory network, calculating the correlation of each time step feature, obtaining an attention weight, and finally obtaining global feature representation; setting two multi-layer perceptron branches to carry out wind speed prediction correction to obtain a common weather branch prediction value and an extreme weather branch prediction value; constructing a correction curve of each sector and obtaining a correction curve prediction value; and according to the common weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value, carrying out weighted fusion to obtain a final wind speed correction value. According to the method, the correction precision and robustness are improved, and the interpretability and applicability of the model are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and in particular to a method and system for correcting wind speed in a wind farm based on dynamic spatiotemporal modeling. Background Art

[0002] With the continuous expansion of wind power generation, the accuracy of wind power forecasts is crucial for ensuring safe grid operation and enabling the friendly integration of renewable energy sources. Because wind speed is affected by factors such as meteorological disturbances, terrain, and eddy currents, it exhibits highly nonlinear and non-stationary characteristics in both time and space. Consequently, wind speed forecasts directly derived from numerical weather forecast models are unable to meet the accuracy requirements for wind farm operations. Therefore, intelligent correction of wind speed forecasts has become a key step in improving the reliability and accuracy of wind power forecasts.

[0003] Most existing correction methods use neural networks to model the mapping relationship between meteorological factors and wind speed errors. However, these methods still have significant shortcomings in modeling the dynamic changes in wind speed, characterizing the spatial heterogeneity of wind direction, and responding to extreme weather events. For example, patent CN119067269A discloses a method and system for correcting wind speed forecasts for integrated wind farms. This method achieves a corrected output for wind speed forecasts by constructing a multi-layer perceptron structure and integrating multi-dimensional input information such as wind speed, wind direction, and temperature. Although this method has improved the correction accuracy to a certain extent, some problems still exist: the scheme fails to utilize the long-term dependence and nonlinear change 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 meteorological conditions, such as thunderstorms, strong winds or sudden changes in airflow, the existing model does not set a special prediction path, resulting in a decrease in the accuracy of the model during critical periods; there is a lack of explicit correction mechanisms based on physics inspiration, such as the correction curve strategy based on wind direction sector division and historical statistical relationship fitting. The model has weak interpretability, is easily affected by changes in the distribution of training data, and lacks stability.

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

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

[0006] In one aspect, the present invention proposes a wind speed correction method for a wind farm based on dynamic spatiotemporal modeling, comprising: Obtaining a measured power curve of each wind turbine generator set in the wind farm, and preprocessing the measured power curve to obtain a power curve for the entire farm; The entire wind speed of the historical period is obtained according to the entire power curve, and a time series correlation is performed with the entire wind direction. At the same time, the meteorological characteristics of the corresponding period are collected to construct a multimodal training data set; Inputting the multimodal training data set into a convolutional neural network to extract local features of the wind speed spatiotemporal data, and inputting the local features into a long short-term memory network, introducing an attention mechanism into the feature sequence output by the long short-term memory network, calculating the correlation of the features of each time step and obtaining attention weights, and performing weighted aggregation on the time step features according to the attention weights to obtain a global feature representation; Two multi-layer perceptron branches are set, and the multi-layer perceptron branches corresponding to the global feature representation input are used to perform wind speed prediction correction to obtain a normal weather branch prediction value and an extreme weather branch prediction value; Dividing the multimodal training data set into a plurality of circular sectors according to the entire wind direction, constructing a correction curve for each sector based on the relationship between the historical weather forecast wind speed and the entire wind speed, inputting the weather forecast wind speed for the target period into the correction curve corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain a sector correction result, and obtaining a correction curve prediction value 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 normal weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value.

[0007] Furthermore, when preprocessing the measured power curve to obtain the entire field power curve, the method includes: The measured power curves of each wind turbine generator set are normalized on the wind speed axis to unify the wind speed grid. Missing power data are linearly interpolated at the corresponding wind speed points and duplicate or abnormal points are eliminated. Boundary constraints are set at the wind speed boundary according to the maximum rated wind speed model and boundary value interpolation is performed to ensure curve continuity, thereby obtaining the optimized power curve of each unit. The optimized power values ​​are added at each uniform wind speed point, and the sum is traversed along the wind speed grid to obtain the entire field power curve.

[0008] Furthermore, the entire wind speed of the historical period is obtained according to the entire power curve, and a time series correlation is performed with the entire wind direction. At the same time, the meteorological characteristics of the corresponding period are collected to construct a multimodal training data set, including: Obtaining the entire field power within a historical period and establishing a monotonic correspondence between power and wind speed based on the entire field power curve, calculating the entire field wind speed at each moment on the wind speed grid based on reverse lookup and piecewise linear interpolation; obtaining the entire field wind direction representing the wind farm, de-ringing the angle data and digitizing it with sine and cosine coding, and time-aligning it with the entire field wind speed at a uniform sampling step; Collecting meteorological characteristics on the same time axis as the historical period and normalizing the meteorological characteristics, the meteorological characteristics including wind speed, wind direction, temperature, humidity, air pressure and boundary layer height; Samples are constructed based on a sliding time window. The meteorological characteristics within the window, the entire wind direction, and the derived characteristics are used as input. The entire wind speed corresponding to the window forward step is used as a label to form the multimodal training dataset. Weather labels for normal weather and extreme weather are annotated according to the wind force level threshold. The derived characteristics include wind vector component, wind shear index, wind direction turning rate, boundary layer stability index, and boundary layer height change rate.

[0009] Furthermore, when the multimodal training data set is input into a convolutional neural network to extract local features of the wind speed spatiotemporal data, the method includes: The multimodal training data set is stacked according to the sliding time window in the time dimension and the meteorological features and the entire wind direction are channel-joined in the feature dimension to form a one-dimensional time series tensor; A convolutional neural network using one-dimensional temporal convolutions, causal padding in the first layer, and multi-scale convolution kernels with optional dilated convolutions in the middle layers. The kernel length of the multi-scale convolution kernel is one or more of 3, 5, and 7. Nonlinear activation is set in sequence after each convolutional layer and deep layer degradation is suppressed through residual connection, and finally the local features are output.

[0010] Furthermore, the correlation of the features of each time step is calculated and the attention weight is obtained. The time step features are weighted and aggregated according to the attention weight to obtain the global feature representation, including: Inputting the local features into a stacked long short-term memory network to model long-term dynamic dependencies and obtain a hidden state sequence; Calculating the relevance score based on the attention mechanism and obtaining the attention weight through Softmax normalization; The time step features are weightedly aggregated according to the attention weight to obtain a context vector, and the context vector is nonlinearly projected to obtain the global feature representation.

[0011] Furthermore, two multi-layer perceptron branches are set, and the multi-layer perceptron branches corresponding to the global feature representation input are used to perform wind speed prediction correction, and the prediction values ​​of the normal weather branch and the extreme weather branch are obtained, including: When setting two multi-layer perceptron branches, including a normal weather multi-layer perceptron branch and an extreme weather multi-layer perceptron branch; Each of the multi-layer perceptron branches includes at least two fully connected layers and nonlinear activation, and random inactivation is set between the layers to suppress overfitting; in the training phase, only the parameters of the corresponding multi-layer perceptron branch are updated based on the weather label, and the other multi-layer perceptron branch is frozen; in the prediction phase, the global feature representation is forward-inputted into the two multi-layer perceptron branches respectively to simultaneously generate the prediction value of the normal weather branch and the prediction value of the extreme weather branch, and the quantile loss is used as the robust loss function to constrain the training error.

[0012] Furthermore, dividing the multimodal training data set into a plurality of circular sectors according to the entire wind field, and constructing a correction curve for each sector based on the relationship between the historical meteorological forecast wind speed and the entire wind field speed includes: The wind direction circle is divided into the circular sectors of equal angular width using the entire wind direction as an angle variable, and the sector extension is divided into a plurality of ring belts according to the wind speed level to form a hierarchical structure; In each of the circular sectors, monotonic constrained regression fitting is performed with the historical weather forecast wind speed as the independent variable and the entire wind speed as the dependent variable to obtain the correction curve.

[0013] Furthermore, the weather forecast wind speed for the target period is input into the correction curve corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain a sector correction result, and the sector correction results of the adjacent sectors are fused by inverse distance weighted method to obtain a correction curve prediction value, including: Determine the target circular sector and its adjacent circular sectors corresponding to the entire wind direction during the target period, substitute the meteorological forecast wind speed during 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 entire wind direction, and perform weighted summation of the sector correction results to obtain the predicted value of the correction curve.

[0014] Furthermore, when obtaining the final wind speed correction value based on the weighted fusion of the normal weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value, the method includes: Performance indicators are calculated for the normal weather branch prediction value, the extreme weather branch prediction value, and the correction curve prediction value, respectively, and subjective weights and objective weights are determined based on the performance indicators. Weight normalization and non-negative constraints are used to obtain a combined weight, and the sum of all weights in the combined weight is 1. The final wind speed correction value is calculated based on the combined weight and the normal weather branch prediction value, the extreme weather branch prediction value, and the correction curve prediction value.

[0015] Compared with existing technologies, the present invention has the following advantages: by obtaining the measured power curves of each wind turbine in a wind farm and preprocessing them to obtain the overall power curve, the overall wind speed is obtained by inversion and paired with the overall wind direction in time series, thus constructing a multimodal training dataset covering 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 its time-dependent characteristics are then captured through a long-short-term memory network. An attention mechanism is introduced into the output sequence to weight the importance of each time step, enhancing the model's ability to model wind speed trends. Two multi-layer perceptron branches, one for normal weather and one for extreme weather, are set up to learn prediction strategies under different meteorological conditions, improving the ability to respond to extreme weather situations. In terms of spatial structure, the training dataset is divided into circular sectors based on the overall wind direction. Correction curves are constructed for each sector based on the relationship between historical forecast wind speeds and measured wind speeds. The correction results of adjacent sectors are integrated with an inverse distance weighting strategy to generate physically meaningful correction curve predictions. By integrating the normal weather forecast values, extreme weather forecast values ​​and correction curve forecast values, 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 adaptation to extreme situations.

[0016] On the other hand, the present application also provides a wind farm wind speed correction system based on dynamic spatiotemporal modeling, which is used to apply the above-mentioned wind farm wind speed correction method based on dynamic spatiotemporal modeling, including: A collection unit is configured to obtain a measured power curve of each wind turbine generator set in the wind farm, and pre-process the measured power curve to obtain a power curve of the entire farm; a correlation unit configured to obtain the entire wind speed of a historical period according to the entire power curve, perform time series correlation with the entire wind direction, and simultaneously collect meteorological characteristics of the corresponding period to construct a multimodal training data set; A first processing unit is configured to input the multimodal training data set into a convolutional neural network to extract local features of the 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 the features of each time step and obtain attention weights, and perform weighted aggregation on the time step features according to the attention weights to obtain a global feature representation; The second processing unit is configured to set two multi-layer perceptron branches, input the global feature representation into the corresponding multi-layer perceptron branch to perform wind speed prediction correction, and obtain a normal weather branch prediction value and an extreme weather branch prediction value; a third processing unit configured to divide the multimodal training data set into a plurality of circular sectors according to wind direction, construct a correction curve for each sector based on a relationship between historical wind speed forecasts and actual wind speeds, input the meteorological forecast wind speed for a target period into the correction curve corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain a sector correction result, and fuse the sector correction results of the adjacent sectors by inverse distance weighted fusion to obtain a predicted value of the correction curve; The correction unit is configured to obtain a final wind speed correction value based on weighted fusion of the normal weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value.

[0017] It is understandable that the above-mentioned wind farm wind speed correction method and system based on dynamic spatiotemporal modeling have the same beneficial effects, which will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings: Figure 1 A flow chart of a wind farm wind speed correction method based on dynamic spatiotemporal modeling provided by an embodiment of the present invention; Figure 2 This is a functional block diagram of a wind farm wind speed correction system based on dynamic spatiotemporal modeling provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0020] In traditional existing wind power prediction systems, the correction process of meteorological forecast wind speed often ignores the coupling effect of the spatiotemporal dynamic characteristics of wind speed. The nonlinear fluctuations and long-term dependencies of wind speed series are difficult to be fully modeled by traditional neural networks, resulting in increased prediction errors when wind speed changes rapidly or there are periodic disturbances. At the same time, the spatial heterogeneity of wind direction within the wind farm is not effectively represented, and the differences in local wind speed distribution corresponding to different wind directions under the same meteorological conditions are not explicitly modeled, resulting in a delayed response of the correction model to sudden changes in wind direction or local eddies. In addition, the sudden changes in meteorological elements under extreme weather conditions are different from conventional weather patterns, but the existing models have not designed independent prediction paths for such scenarios, resulting in limited generalization capabilities of the model under conditions of severe convection or boundary layer instability.

[0021] For example, in wind farms deployed in complex terrain areas, there are systematic deviations between the regional wind speed data provided by the meteorological forecast system and the actual wind tower observations at the wind farm. When the prevailing wind direction is affected by the flow around the mountain to form a local vortex, the actual wind speed of each unit within the wind farm shows spatial differentiation. The traditional correction model uses a unified fully connected network structure to process multi-dimensional meteorological inputs, and cannot effectively capture the nonlinear relationship between wind speed and meteorological elements in different wind direction sectors. When encountering severe convective weather, the sudden drop in boundary layer height leads to an increase in vertical wind speed shear, but the model does not establish a special feature extraction channel, which makes the correction result lag behind the actual wind speed change.

[0022] If these issues are not addressed, wind farm power forecasting systems will be unable to adapt to the dynamic characteristics of wind speeds in complex terrain and variable weather conditions, leading to deviations between power forecast curves and grid dispatch instructions. Long-term accumulated forecast errors can cause voltage fluctuations at the wind farm's grid connection point, increasing the burden on power system frequency regulation. During extreme weather events, uncorrected wind speed forecast deviations can cause turbines to overspeed or exceed reactive power limits, triggering protective devices and impacting wind farm economics and equipment safety.

[0023] For this, see Figure 1 As shown, this application proposes a wind farm wind speed correction method based on dynamic spatiotemporal modeling, including: S100: Obtaining a measured power curve of each wind turbine generator set in the wind farm, and preprocessing the measured power curve to obtain a power curve for the entire farm; S200: Obtain the entire wind speed of the historical period based on the entire power curve, and perform time series correlation with the entire wind direction. Simultaneously, collect meteorological characteristics of the corresponding period to construct a multimodal training dataset. S300: Input the multimodal training dataset into a convolutional neural network to extract local features of the wind speed spatiotemporal data. The local features are then 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 the features at each time step and obtain the attention weights. The time step features are weighted and aggregated according to the attention weights to obtain a global feature representation. S400: Setting two multi-layer perceptron branches, inputting the global feature representation into the corresponding multi-layer perceptron branch to perform wind speed prediction correction, and obtaining a normal weather branch prediction value and an extreme weather branch prediction value; S500: Divide the multimodal training dataset into several levels of circular sectors according to the overall wind direction. Construct a correction curve for each sector based on the relationship between the historical weather forecast wind speed and the overall wind speed. Input the weather forecast wind speed for the target period into the correction curve corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain a sector correction result. The sector correction results of the adjacent sectors are fused through inverse distance weighting to obtain a predicted value of the correction curve. S600: Obtain a final wind speed correction value based on a weighted fusion of the normal weather branch prediction value, the extreme weather branch prediction value, and the correction curve prediction value.

[0024] Specifically, dynamic spatiotemporal modeling involves extracting local features from spatiotemporal wind speed data using a convolutional neural network (CNN), combined with a long-short-term memory (LSTM) network to capture long-term dependencies within the time series. This is achieved by using a one-dimensional temporal convolution kernel and causal padding to process the temporal data, and residual connections to suppress deep layer degradation. This feature effectively captures the dynamic variations of wind speed in both temporal and spatial dimensions, addressing the inadequacy of existing methods in modeling scenarios with drastic wind speed fluctuations. The attention mechanism calculates the correlation of features at each time step within the feature sequence output by the LSTM network and aggregates these features using attention weights. This is achieved by using a softmax normalized correlation score and generating a context vector. This feature adaptively focuses on information at key time steps, enhancing the model's ability to represent non-stationary wind speed characteristics. The multi-layer perceptron branching approach involves setting up separate prediction paths for normal and extreme weather conditions. This is achieved by constructing a branching structure using fully connected layers and random dropout techniques, with parameters selectively updated based on weather labels during training. This feature enables the design of differentiated prediction strategies for different weather conditions, improving robustness to corrections in extreme weather scenarios. The circular sector correction curve divides the entire wind direction into hierarchical sectors and constructs a monotonic regression relationship between the forecast wind speed and the entire wind speed. This is achieved by dividing the wind direction circle into equal angular widths and fusing the results of adjacent sectors using inverse distance weighting. This feature explicitly models the impact of spatial heterogeneity in wind direction on the wind speed distribution, enhancing the interpretability and physical consistency of the correction process.

[0025] This application combines dynamic spatiotemporal modeling, a dual-branch prediction mechanism, and a physically inspired sector correction strategy. It extracts spatiotemporal features through convolutional neural networks and long-short-term memory networks, uses an attention mechanism to focus on key time-step information, employs independent branches to handle both normal and extreme weather conditions, and constructs explicit correction curves based on wind direction sectorization. Ultimately, it achieves high-precision wind speed correction through multimodal data fusion. This method achieves synergistic optimization in spatiotemporal dynamics modeling, extreme weather adaptability, and the characterization of wind direction spatial heterogeneity, effectively improving the accuracy and stability of wind speed corrections.

[0026] The working process and principle of this application is to obtain the measured power curves of each wind turbine in the wind farm and preprocess these curves to obtain the power curve of the entire farm. The preprocessing process includes normalization, interpolation, and boundary constraints to ensure data consistency and continuity.

[0027] Based on the full-field power curve, the wind speed for the historical period is obtained and time-series correlated with the wind direction. Meteorological characteristics for the corresponding period, such as wind speed, wind direction, temperature, and humidity, are also collected to construct a multimodal training dataset. This step aligns the time and integrates features of data from different sources, providing input for the subsequent deep learning model.

[0028] The multimodal training dataset was fed into a convolutional neural network to extract local features from the spatiotemporal wind speed data. Convolutional neural networks are effective in capturing local patterns and short-term dependencies in time series data. These extracted local features were then fed into a long short-term memory network to model long-term dynamic dependencies.

[0029] An attention mechanism is introduced into the feature sequence output by the LSTM network. The correlation between features at each time step is calculated and attention weights are obtained. This step enables the model to dynamically focus on the importance of different time steps. Time step features are weighted and aggregated according to the attention weights to obtain a global feature representation that captures key information from the entire sequence.

[0030] Two multi-layer perceptron branches are set up, one for correcting wind speed forecasts for normal weather conditions and the other for extreme weather conditions. The global feature representation is input into the corresponding multi-layer perceptron branch to obtain prediction values ​​for the normal weather branch and the extreme weather branch. This dual-branch structure enables specialized predictions for different weather conditions, improving the model's adaptability.

[0031] Furthermore, the multimodal training dataset is divided into several levels of circular sectors based on the overall wind direction. Correction curves for each sector are constructed based on the relationship between historical weather forecast wind speeds and overall wind speeds. This physics-inspired approach captures local characteristics under different wind directions.

[0032] During the prediction phase, the forecasted wind speed for the target period is input into the correction curve corresponding to the sector in which the target wind direction is located and its adjacent sectors to obtain the sector correction results. The correction curve predictions are then derived by fusing the sector correction results of adjacent sectors using inverse distance weighting. This approach accounts for both spatial continuity and locality.

[0033] The final wind speed correction value is obtained by weighted fusion of the normal weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value. This application comprehensively considers the advantages of different prediction strategies to improve the overall prediction accuracy.

[0034] As a preferred embodiment, the solution of this application is specifically implemented as follows: The SCADA system collects the measured power curve data for each wind turbine within the wind farm. These curves are preprocessed, including normalization of the wind speed axis, linear interpolation of power data, outlier removal, and boundary constraints, to produce the power curve for the entire farm.

[0035] Based on the full-field power curve, a power-wind speed relationship is established. The full-field wind speed for the historical period is calculated through reverse lookup and interpolation. Wind direction data for the entire field is also acquired, de-ringed, and sine-cosine encoded. Meteorological characteristics for the same period are collected, including wind speed, direction, temperature, humidity, pressure, and boundary layer height.

[0036] A sliding time window approach is used to construct a multimodal training dataset. The meteorological characteristics, overall wind direction, and derived features within the window are used as input, and the overall wind speed corresponding to the window's forward look-ahead step is used as the label. Derived features include wind vector components and wind shear index.

[0037] The training data is fed into a one-dimensional temporal convolutional neural network (CNN), which uses multi-scale convolution kernels to extract local features. The feature sequence output by the CNN is fed into a long-short-term memory (LSTM) network to model long-term dependencies. An attention mechanism is applied to the LSTM output sequence, calculating the attention weights for each time step and performing weighted aggregation to obtain a global feature representation.

[0038] Two multi-layer perceptron branches are set up for normal weather and extreme weather. Each branch contains multiple fully connected layers and nonlinear activation functions. The global feature representation is input into the two branches to obtain wind speed prediction values ​​for normal weather and extreme weather, respectively.

[0039] The wind direction circle is divided into multiple equal-angle sectors, and a correction curve is fitted within each sector based on historical data. During the forecast, the corresponding sector and adjacent sectors are determined based on the wind direction during the target period. The forecast wind speed is input into the correction curves for these sectors, and the predicted values ​​of the correction curves are obtained through inverse distance weighted fusion.

[0040] Finally, the normal weather branch predictions, extreme weather branch predictions, and correction curve predictions are weighted and fused to produce the final wind speed correction value. The weighting coefficients are dynamically adjusted based on the historical performance indicators of each prediction branch.

[0041] Through the above 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 ability to model nonlinear fluctuations and long-term dependencies in wind speed series, reducing prediction errors when wind speed changes rapidly or there are periodic disturbances. The introduction of an attention mechanism enables the model to dynamically focus on features at key time steps, improving its responsiveness to sudden changes.

[0042] A physics-inspired circular sectorization method, combined with a correction curve fitted from historical data, effectively characterizes the spatial heterogeneity of wind direction within a wind farm. This method can identify differences in local wind speed distribution corresponding to different wind directions, improving the model's response to sudden changes in wind direction or local eddies.

[0043] The dual-branch prediction architecture establishes independent prediction paths for both normal and extreme weather conditions, improving the model's adaptability to diverse meteorological conditions. The model's generalization capabilities are particularly enhanced under extreme weather conditions, such as severe convection or boundary layer instability.

[0044] In summary, this application improves the accuracy and reliability of wind speed correction and provides more accurate wind speed input for wind farm power prediction.

[0045] Some of the aforementioned solutions in this application propose obtaining a full-field power curve from the measured power curve to support subsequent wind speed corrections. However, in actual operation, differences in wind turbine models lead to inconsistent wind speed axis resolution in the measured power curve, resulting in missing, duplicated, or outlier data. Furthermore, differences in rated wind speeds between wind turbine models lead to discontinuities in the curve at wind speed boundaries. These issues directly affect the accuracy of the full-field power curve, thereby reducing the reliability of subsequent wind speed corrections.

[0046] The present application further proposes a method for preprocessing the measured power curve to obtain the power curve of the entire field, including performing interval normalization on the measured power curve of each wind turbine generator set on the wind speed axis to unify the wind speed grid, performing linear interpolation on the missing power data at the corresponding wind speed points and eliminating duplicate or abnormal points, setting boundary constraints at the wind speed boundary according to the maximum rated wind speed model and performing boundary value interpolation to ensure curve continuity, thereby obtaining the optimized power curve of each unit; adding the optimized power values ​​at each unified wind speed point, and traversing the wind speed grid to obtain the power curve of the entire field.

[0047] Among them, interval normalization maps the discrete wind speed points of different models to standardized grid nodes by setting a uniform wind speed grid spacing, eliminating the difference in original data resolution. Linear interpolation uses the power values ​​of adjacent wind speed points for linear fitting to fill in the missing power data of the wind speed grid nodes after normalization to ensure data integrity. Outlier removal identifies and removes data points that exceed the threshold by setting a threshold for the reasonable range of power variation relative to wind speed. Boundary constraints are based on the model with the largest rated wind speed, and the rated power value of the model is forcibly set at the upper and lower boundaries of the wind speed grid to avoid conflicts in boundary data of different models. Boundary value interpolation uses quadratic spline interpolation to generate a continuous curve within the boundary constraint range to ensure a smooth transition of the power curve of the entire field.

[0048] Specifically, during the interval normalization process, the wind speed axis of each turbine's measured power curve is uniformly divided into grid nodes spaced 0.5 m / s apart, covering the effective wind speed range for all turbine types. For each grid node, if no measured data is available for a particular turbine at that node, the power values ​​at two adjacent wind speed points are linearly interpolated to fill the gap. When multiple power measurements exist at the same wind speed point, the point with the smallest deviation from the average power is retained, and the remaining points are discarded as duplicates. At the upper and lower boundaries of the wind speed grid—for example, when the boundary corresponding to the maximum rated wind speed model is 25 m / s—the power curves of all turbine types are truncated above 25 m / s, and a continuous curve is generated using cubic polynomial interpolation in the range of 24.5-25 m / s. After optimizing each turbine, the power values ​​of all turbines at the same wind speed grid node are superimposed to form the overall power curve. For example, at a wind speed node of 15 m / s, if the optimized powers of three turbines are 3.2 MW, 2.8 MW, and 3.0 MW, respectively, the overall power curve is 9.0 MW. This process provides high-precision input for subsequent full-field wind speed calculations by eliminating data noise, unifying data formats, and enforcing boundary continuity.

[0049] As a preferred embodiment, the solution of this application is specifically implemented as follows: When preprocessing the measured power curve to obtain the full-field power curve, the measured power curves of each wind turbine are first normalized along the wind speed axis to create a unified wind speed grid. Specifically, a uniform wind speed interval of 0.5 m / s is selected, and the wind speed points of the original measured power curve are interpolated onto this unified grid.

[0050] Secondly, missing power data is linearly interpolated at the corresponding wind speed points. For example, if power data for a certain wind speed point is missing, it is filled by linear interpolation using the power values ​​of two adjacent valid wind speed points. At the same time, duplicate or outlier points are removed, such as data points with power values ​​significantly deviating from the normal range.

[0051] Furthermore, a boundary constraint is set at the wind speed boundary based on the turbine with the highest rated wind speed. Assuming the turbine with the highest rated wind speed in the wind farm is 15 m / s, the right boundary of the entire farm's power curve is set to 15 m / s. Based on this, boundary value interpolation is performed to ensure curve continuity, using cubic spline interpolation to achieve smooth transitions. Through these steps, the optimized power curve for each turbine unit is obtained.

[0052] Finally, at each uniform wind speed point, the optimized power values ​​are summed up, traversing the wind speed grid. For example, at each wind speed point with an interval of 0.5 m / s, the optimized power values ​​of all units at that wind speed are accumulated to obtain the power curve for the entire field.

[0053] Through the above technical solution, the present application realizes the standardized processing and integration of the measured power curves of each wind turbine generator set. As a result, the data differences between different units are eliminated, missing data are filled, outliers are eliminated, and the continuity and smoothness of the curve are guaranteed. Furthermore, by accumulating the optimized power curves of each unit to obtain the power curve of the entire field, a reliable data basis is provided for subsequent wind speed correction. This processing method improves the accuracy and representativeness of the power curve of the entire field, and helps to improve the accuracy and reliability of wind speed correction.

[0054] In some of the above-mentioned schemes of this application, when obtaining the entire field power curve by preprocessing the measured power curve and constructing a multimodal training data set, the complex relationship between the actual operation data of the wind farm and the meteorological characteristics is not fully integrated, resulting in the model training data lacking an effective representation of the dynamic change law of wind speed, affecting the subsequent prediction accuracy.

[0055] The present application further proposes to obtain the entire field power within a historical period and establish a monotonic correspondence between power and wind speed based on the entire field power curve, and obtain the entire field wind speed at each moment on the wind speed grid based on reverse search and piecewise linear interpolation; obtain the entire field wind direction representing the wind farm, de-ring the angle data and digitize it with sine and cosine coding, and time-align it with the entire field wind speed according to a uniform sampling step; collect meteorological features on the same time axis as the historical period and normalize the meteorological features, the meteorological features include wind speed, wind direction, temperature, humidity, air pressure and boundary layer height; construct samples based on a sliding time window, take the meteorological features within the window and the entire field wind direction and derived features as input, take the entire field wind speed corresponding to the window forward step as a label, constitute a multimodal training data set, and annotate the weather labels of normal weather and extreme weather according to the wind force level threshold; derived features include wind vector component, wind shear index, wind direction turning rate, boundary layer stability index and boundary layer height change rate.

[0056] Among them, reverse lookup and piecewise linear interpolation ensure the physical consistency of wind speed throughout the field by establishing a monotonic relationship between power and wind speed; sine and cosine encoding converts wind direction angles into continuous values ​​to avoid circular discontinuities in angle data; normalization processing eliminates dimensional differences in meteorological characteristics and improves model convergence efficiency; sliding time windows convert time series data into supervised learning samples to capture the temporal dependence of wind speed changes; derived features enhance data characterization capabilities through physical deduction, for example, the wind vector component is decomposed into east-west and north-south components, the wind shear index reflects the vertical wind speed gradient, and the boundary layer stability index is calculated based on temperature and boundary layer height.

[0057] Specifically, the entire field power curve provides a mapping between the overall power of the wind farm and wind speed. A reverse lookup is performed to determine the corresponding entire field wind speed on the wind speed grid. Piecewise linear interpolation fills in missing values ​​between discrete wind speed points to ensure a continuous and smooth wind speed series. The entire field wind direction is de-circularized to eliminate sudden changes between 0° and 360°. Sin-cosine encoding converts angles into sine and cosine values ​​to avoid model deviations due to angular periodicity. Meteorological feature normalization uses maximum and minimum value scaling to ensure that features of different dimensions are within the same numerical range. A sliding time window divides historical time series data into multiple samples. The window length and look-ahead step are set based on the wind speed variation period, for example, a window length of 6 hours and a look-ahead step of 1 hour. Among the 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 altitudes. The boundary layer stability index represents the atmospheric stability state by combining the temperature gradient and the rate of change of the boundary layer altitude. Weather labels are categorized based on wind speed thresholds—for example, wind speeds below 17.2 m / s are considered normal weather, while wind speeds above this threshold are considered extreme weather. These labels are then used for subsequent branch training. This multimodal training dataset, by integrating wind farm operating data, meteorological observations, and physically derived features, provides the model with multidimensional dynamic information input, improving the accuracy of wind speed correction.

[0058] As a preferred embodiment, the solution of this application is specifically implemented as follows: The entire field power is obtained over a historical period and a monotonic relationship between power and wind speed is established based on the entire field power curve. The entire field wind speed at each moment is calculated on the wind speed grid using reverse lookup and piecewise linear interpolation. The entire field wind direction representing the wind farm is obtained, the angle data is de-ringed and digitized using sine and cosine encoding, and time-aligned with the entire field wind speed using a uniform sampling step.

[0059] Meteorological characteristics on the same time axis as the historical period are collected and normalized. The meteorological characteristics include wind speed, wind direction, temperature, humidity, air pressure and boundary layer height.

[0060] The method constructs samples based on a sliding time window, using the meteorological features within the window, the overall wind direction, and derived features as input. The overall wind speed corresponding to the window's forward-looking step is used as the label to form a multimodal training dataset. Weather labels for normal and extreme weather are annotated based on wind force thresholds. Derived features include wind vector components, wind shear index, wind direction turning rate, boundary layer stability index, and boundary layer height change rate.

[0061] Specifically, a 6-hour sliding window can be used to construct samples with a 10-minute sampling step. For meteorological characteristics, wind speed and direction can be predicted at a height of 10 meters, temperature and humidity at a height of 2 meters, air pressure at sea level, and boundary layer height using the output of the forecast model. Normalization can be performed using the maximum and minimum value normalization method.

[0062] The calculation method of the derived characteristics is as follows: the wind vector component is obtained by decomposing the trigonometric functions of wind speed and wind direction; the wind shear index is calculated by the ratio of wind speed at heights of 10 meters and 80 meters; the wind direction turning rate is used to calculate the change in wind direction at two adjacent moments; the boundary layer stability index can be calculated using the Richardson number; and the boundary layer height change rate is used to calculate the change in boundary layer height at two adjacent moments.

[0063] Weather labels can be categorized using the Beaufort wind scale, with winds of force 7 and above defined as extreme weather. Samples can be constructed using a 36-hour forecast window, meaning the sliding window contains data from 24 historical moments, with the labels representing the wind speed for the next 12 moments.

[0064] Through the above technical solution, this application constructs a training data set containing multimodal meteorological features, derived features and label information, providing rich input information for subsequent dynamic spatiotemporal modeling. The time series characteristics of wind speed are captured by constructing a sliding time window. At the same time, derived features are introduced to enhance the characterization of the physical characteristics of the wind field. In addition, through the division of weather labels, the model is provided with prior information to distinguish between normal weather and extreme weather. This multi-dimensional and multi-scale data construction method helps to improve the model's ability to learn the laws of wind speed changes, thereby improving the accuracy and robustness of wind speed correction.

[0065] In some of the above-mentioned schemes in this application, the multimodal training data set contains high-dimensional spatiotemporal features and complex meteorological elements. It is difficult to effectively extract local spatiotemporal correlation patterns by directly inputting them into traditional neural networks, resulting in the model being unable to accurately capture the local mutations and periodic fluctuation characteristics in the dynamic changes of wind speed.

[0066] This application further proposes stacking the multimodal training data set in the time dimension as a sliding time window and performing channel splicing on the meteorological features and the entire wind direction in the feature dimension to form a one-dimensional time series tensor; using a convolutional neural network with one-dimensional time series convolution, and using causal filling in the first layer, and using multi-scale convolution kernels and optional void convolution in the middle layer, and the kernel length of the multi-scale convolution kernel is one or more of 3, 5, and 7; setting nonlinear activation in turn after each convolution layer and suppressing deep layer degradation through residual connection, and finally outputting local features.

[0067] The one-dimensional time series tensor converts time series data into a fixed-length input structure through a sliding window, and multimodal features are integrated through splicing to achieve cross-modal information fusion. Causal padding ensures that convolution operations rely solely on historical data, preventing future information leakage. Multi-scale convolution kernels use different kernel lengths to capture local features of short, medium, and long periods, respectively. Dilated convolution expands the receptive field by adjusting the dilation rate, enhancing the detection of sparse events. Residual connections propagate gradients through skip paths, alleviating the vanishing gradient problem in deep network training.

[0068] Specifically, after the multimodal training dataset is segmented using a sliding window, meteorological features and wind direction data are concatenated along the feature dimension to form a one-dimensional time series tensor, where each time step contains a multidimensional feature vector. Causal padding is applied at the beginning of the convolutional layer to ensure that the output length matches the input and depends only on the left-hand input. Multi-scale convolutional layers employ kernels of lengths 3, 5, and 7 in parallel to extract local patterns across different time spans. For example, a kernel length of 3 captures sudden changes between adjacent time points, while a kernel length of 7 identifies fluctuations between periodic intervals. Dilated convolutions expand temporal coverage without increasing the number of parameters by sampling the input data at intervals. For example, with a dilation rate of 2, a convolution with a kernel length of 5 actually covers 11 time steps. The feature maps output by each convolutional layer are activated with ReLU and then added to the input via residual connections to suppress degradation in deep networks. The final output local features contain multi-scale spatiotemporal patterns, providing a foundation for modeling long-term dependencies in the subsequent long-short-term memory network.

[0069] As a preferred embodiment, the solution of this application is specifically implemented as follows: The multimodal training dataset is stacked in sliding time windows along the time dimension, and the meteorological features and overall wind direction are channel-joined along the feature dimension to form a one-dimensional time series tensor. A convolutional neural network with one-dimensional time series convolution is employed, with causal padding applied to the first layer and multi-scale convolution kernels with optional dilated convolutions in the intermediate layers. The kernel lengths of the multi-scale convolution kernels are set to 3, 5, and 7, respectively. Nonlinear activations are applied after each convolution layer, and residual connections are used to suppress deep layer degradation, ultimately outputting local features.

[0070] Specifically, the sliding time window length is set to 24 hours, with a step size of 1 hour. The shape of the one-dimensional time series tensor is (batch_size, 24, num_features), where num_features is the total dimension of the meteorological features and the wind direction of the entire field. The convolutional neural network contains three convolutional layers. The first layer uses causal padding with a convolution kernel size of 3. The second and third layers use convolution kernels of lengths 5 and 7, respectively. Each convolutional layer is followed by a ReLU activation function, and the input is directly added to the output through a residual connection. After the final convolution layer, local features are obtained through global average pooling.

[0071] Through the above technical solution, this application can effectively extract local features of wind speed spatiotemporal data. By using multi-scale convolution kernels, the model can simultaneously capture wind speed variation patterns at different time scales. Causal filling ensures that the model uses only current and past information for prediction, avoiding information leakage. Residual connections help alleviate the vanishing gradient problem, enabling better training of deep networks. This design enables the model to more accurately characterize the spatiotemporal correlation of wind speed, thereby improving the accuracy of wind speed correction.

[0072] In some of the above-mentioned schemes in this application, when long-term dynamic dependencies are modeled through 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, resulting in the averaging of key time step features, which cannot fully capture the key fluctuation patterns in the dynamic changes of wind speed, affecting the global feature representation's ability to characterize complex spatiotemporal correlations, and thereby reducing the accuracy of wind speed correction.

[0073] This application further proposes to input local features into a stacked long short-term memory network to model long-term dynamic dependencies and obtain a hidden state sequence; calculate the correlation score based on the attention mechanism, and obtain the attention weight through Softmax normalization; perform weighted aggregation of time step features according to the attention weight to obtain a context vector, and obtain a global feature representation by nonlinearly projecting the context vector.

[0074] Among them, the stacked long short-term memory network adopts a multi-layer structure to enhance the sequence modeling capability, and each time step in the hidden state sequence corresponds to a feature vector; the attention mechanism measures the correlation by calculating the dot product score between the query vector and the key vector of each time step, and the score is converted into the attention weight of the probability distribution after exponential operation and normalization; the weighted aggregation process linearly combines the feature vectors of each time step according to the weight to generate a context vector, and the nonlinear projection uses a fully connected layer and activation function to reduce or increase the dimension of the context vector.

[0075] Specifically, each time-step feature vector in the hidden state sequence is mapped into a query vector, a key vector, and a value vector. After calculating the dot product score between the query vector and all key vectors, the score is converted into attention weights using the Softmax function. The value vectors of each time step are weighted and summed to obtain a context vector, which integrates the contributions of different time steps in the sequence. The nonlinear projection layer maps the context vector to the target dimensional space, forming a global feature representation that contains information about 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, nonlinear projection enhances the nonlinear expression ability of 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.

[0076] As a preferred embodiment, the solution of this application is specifically implemented as follows: Local features are input into a stacked LSTM network to model long-term dynamic dependencies and obtain a hidden state sequence. Specifically, a two-layer LSTM network is used, with each layer containing 128 hidden units. The input of the first LSTM layer is the local feature sequence, and the output serves as the input of the second layer. The output of the second LSTM layer is the hidden state sequence.

[0077] The attention mechanism calculates relevance scores and obtains attention weights through softmax normalization. Furthermore, 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 by linearly transforming the hidden state at the last time step. The relevance scores are normalized through the softmax function to obtain attention weights.

[0078] The context vector is obtained by weighted aggregation of time-step features according to the attention weights, and the context vector is nonlinearly projected to obtain the global feature representation. The context vector is then weighted summed over the hidden state sequence using the attention weights. The context vector is then nonlinearly transformed through a single-layer feedforward neural network to obtain the final global feature representation.

[0079] 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. This global feature representation comprehensively considers the dynamic changes in historical wind speed data, improving the accuracy and robustness of wind speed corrections. Furthermore, the attention mechanism enhances the model's interpretability, facilitating analysis of key time points that influence wind speed predictions.

[0080] In some of the above-mentioned schemes in this application, the multi-layer perceptron structure may cause conflicts in the direction of model parameter updates during the training process due to the differences in the distribution of normal weather and extreme weather samples. The sparsity of extreme weather samples makes it difficult for the model to fully learn its unique pattern, thereby affecting the accuracy of wind speed prediction under extreme weather conditions.

[0081] The present application further proposes setting up two multi-layer perceptron branches, including a normal weather multi-layer perceptron branch and an extreme weather multi-layer perceptron branch; each multi-layer perceptron branch includes at least two fully connected layers and nonlinear activation, and random inactivation is set between the layers to suppress overfitting; in the training phase, only the parameters of the corresponding multi-layer perceptron branch are updated based on the weather label, and the other multi-layer perceptron branch is frozen. In the prediction phase, the global feature representation is forward-inputted into the two multi-layer perceptron branches respectively to simultaneously generate the normal weather branch prediction value and the extreme weather branch prediction value, and the quantile loss is used as the robust loss function to constrain the training error.

[0082] Among them, the fully connected layers of the normal 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 is dynamically adjusted during the training process; the quantile loss function selects 0.05, 0.5, and 0.95 quantiles to calculate the weighted error; the weather label is divided according to the wind level threshold, and when the measured wind speed exceeds 20 meters per second, it is marked as an extreme weather sample.

[0083] Specifically, during the training phase, when the weather label of the input sample is normal weather, only the gradient of the fully connected layer parameters of the normal weather branch is updated, and the parameters of the extreme weather branch remain frozen; conversely, for extreme weather samples, only the extreme weather branch parameters are updated. The two branches learn the wind speed error distribution patterns under normal and extreme weather modes respectively through independent optimization processes. During the prediction phase, the global feature representation is input into both branches at the same time, and the corresponding predicted values ​​are output respectively. The quantile loss function enhances the robustness of the model to abnormal wind speed values ​​by minimizing the deviation between the predicted value and the true value at multiple quantiles. The random dropout operation randomly blocks some neurons in the hidden layer, forcing the network to learn redundant feature expressions and avoid overfitting to specific weather modes. This technical solution effectively alleviates the learning interference between different weather modes through branch decoupling and parameter isolation mechanisms, thereby improving the prediction stability under extreme weather conditions.

[0084] As a preferred embodiment, the solution of this application is specifically implemented as follows: Two multi-layer perceptron branches are set up, including a normal weather multi-layer perceptron branch and an extreme weather multi-layer perceptron branch. Each multi-layer perceptron branch consists of three fully connected layers with nonlinear activation, and random dropout is set between layers to suppress overfitting. During the training phase, based on the weather label, only the corresponding multi-layer perceptron branch is updated with parameters, while the other multi-layer perceptron branch is frozen. During the prediction phase, the global feature representation is forward-feeded into the two multi-layer perceptron branches to simultaneously generate prediction values ​​for the normal weather branch and the extreme weather branch. The quantile loss is used as a robust loss function to constrain the training error.

[0085] Specifically, the three fully-connected layers of the multi-layer perceptron for normal weather have 128, 64, and 32 neurons, respectively, while the three fully-connected layers of the multi-layer perceptron for extreme weather have 64, 32, and 16 neurons, respectively. Both branches use ReLU as the nonlinear activation function, and a random dropout layer with a dropout rate of 0.3 is placed between the first and second layers. During training, the corresponding branch is selected for backpropagation based on the weather label of the sample, and the parameters of the unselected branches remain unchanged. During prediction, the global feature representation is simultaneously input into both branches, resulting in two prediction results. The quantile loss function is set to the 0.5 quantile, that is, median regression, to enhance the model's robustness to outliers.

[0086] Through the above technical solution, this application can construct specialized forecasting models for both normal and extreme weather conditions, improving the model's adaptability and forecasting accuracy under different weather conditions. Furthermore, by adopting a quantile loss function, the model's robustness to outliers is enhanced, improving the reliability of the forecast results. Furthermore, by selectively updating parameters during the training phase, the model's generalization ability is improved by avoiding interference from irrelevant samples.

[0087] In some of the above-mentioned schemes of this application, a method of constructing a correction curve by dividing the circular sectors is proposed. However, in scenarios with spatial heterogeneity and uneven wind speed distribution dominated by wind direction, the existing methods fail to effectively combine wind speed levels and wind direction angles for hierarchical division, resulting in insufficient adaptability of the correction curve in local areas and difficulty in accurately capturing the dynamic relationship between different wind speed ranges and wind direction sectors.

[0088] The present application further proposes to divide the wind direction circle into circular sectors of equal angular width using the entire wind direction as the angular variable, and to divide the sector extension into several ring belts according to the wind speed level to form a hierarchical structure; within each circular sector, a monotonic constrained regression fitting is performed using the historical meteorological forecast wind speed as the independent variable and the entire wind speed as the dependent variable to obtain a correction curve.

[0089] The wind direction circle is divided into multiple equal 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 sector extension is further divided into multiple ring belts according to wind speed level, for example, the wind speed is divided into levels of 0-5 m / s, 5-10 m / s, and 10-15 m / s. Within the sub-area of ​​each sector and ring belt combination, rank-preserving regression or monotone spline fitting methods are used to establish a mapping relationship between the weather 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 belt 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 angles and changes in wind speed intensity.

[0090] Specifically, when constructing the correction curve, the sector to which it belongs is first determined based on the wind direction angle in the historical data, and then the corresponding ring belt level is determined based on the meteorological forecast wind speed value; within the same sector and ring belt, the historical meteorological forecast wind speed is used as the input variable and the entire wind speed is used as the target variable. The constrained regression algorithm is used to fit the monotonic relationship between the two to generate the correction curve for each sub-area. For example, in the northeast wind direction 30-degree sector and the ring belt with a wind speed of 5-10 meters per second, piecewise linear regression is used to force the slope to be non-negative, ensuring that the entire wind speed forecast 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, and improve the local adaptability and physical consistency of the correction curve.

[0091] As a preferred embodiment, the solution of this application is specifically implemented as follows: The wind direction circle is divided into circular sectors of equal angular width using the overall wind direction as the angular variable. The sectors are then divided into several rings based on wind speed levels to form a hierarchical structure. Specifically, the wind direction circle is divided into eight sectors of 45-degree angular width, and each sector is divided into three rings, corresponding to low, medium, and high wind speed ranges, respectively.

[0092] Within each circular sector, a correction curve is obtained by performing a monotonic constrained regression fit using historical forecast wind speed as the independent variable and the overall wind speed as the dependent variable. Furthermore, an rank-preserving regression algorithm is used to fit the historical data within each sector to ensure the monotonicity of the correction curve. Thus, a unique correction curve is generated for each sector and wind speed range combination.

[0093] For example, for the medium wind speed range in the 45-degree north sector, corresponding data points are collected for the historical forecast wind speed and the actual wind speed throughout the entire field within that sector. These data points are fitted using an rank-preserving regression algorithm to produce a monotonically increasing correction curve. This curve can be used to map the forecast wind speed within that sector to a more accurate estimate of the actual wind speed.

[0094] Through the above technical solution, the present application realizes refined correction based on wind direction and wind speed. Since the spatial heterogeneity of wind direction is taken into account, the correction curve can better capture the local terrain effects and wind field characteristics under different wind directions. At the same time, the hierarchical correction structure improves the adaptability of the model to different wind speed ranges. This physical-inspired explicit correction mechanism enhances the interpretability and stability of the model, making the correction results more reliable and easy to understand. Furthermore, the method constructs the correction curve through historical statistical relationships, reduces the dependence on a large amount of training data, and improves the robustness of the model in the case of sparse data.

[0095] In some of the above-mentioned schemes of the present application, when dividing the multimodal training data set into several levels of circular sectors according to the entire wind direction and constructing the correction curve of each sector, only the correction curve of the sector where the target wind direction is located is considered for wind speed prediction, and the influence of adjacent sectors on the target wind direction is not considered. As a result, when there are natural fluctuations or measurement errors in the wind direction, the correction result of a single sector is easy to deviate from the actual wind speed distribution, and the prediction stability is insufficient.

[0096] The present application further proposes to determine the target circular sector and its adjacent circular sectors corresponding to the entire wind direction during the target period, and substitute the meteorological forecast wind speed of 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 entire wind direction, and perform weighted summation of the sector correction results to obtain the correction curve prediction value.

[0097] Among them, when determining the target circular sector and its adjacent sectors, matching is performed in pre-divided circular sectors according to the angular value of the entire wind direction. The adjacent sectors include two sectors that are angularly adjacent to the target sector. When substituting into the correction curve, the correction curve of each sector is constructed based on the monotonic constrained regression relationship between the historical meteorological forecast wind speed and the entire wind speed, ensuring that the corresponding correction result can be directly output after the meteorological forecast wind speed is input. When calculating the inverse distance weight, the absolute value of the angle between the central axis of each sector and the entire wind direction is used as the distance measurement, and the inverse distance weighted formula is used to calculate the weight coefficient, and the weight is inversely proportional to the angle. When weighted summation is performed, the correction result of each sector is multiplied by its corresponding inverse distance weight and then added to obtain the comprehensive correction curve prediction value.

[0098] Specifically, after obtaining the entire wind direction during the target period, the target sector is determined in the pre-divided circular sectors according to the angle value, and the two adjacent sectors on its left and right are selected as adjacent sectors. The meteorological forecast wind speed of the target period is input into the correction curves corresponding to the target sector and its left and right adjacent sectors respectively to obtain the correction results of the three sectors. The absolute value of the angle between the central axis of each sector and the entire wind direction is calculated. For example, if the central axis of the target sector is 30 degrees and the entire wind direction is 28 degrees, the angle is 2 degrees; the central axis of the adjacent left sector is 15 degrees, the angle is 13 degrees; the central axis of the adjacent right sector is 45 degrees, the angle is 17 degrees. The weight is calculated using the inverse distance weighted 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 three sector correction results are multiplied by the corresponding weights and added together to obtain the final correction curve prediction value. By introducing the correction results of adjacent sectors and performing weighted fusion based on spatial proximity, the prediction error caused by the angle deviation of a single sector can be effectively reduced, and the robustness and accuracy of the correction results can be improved.

[0099] As a preferred embodiment, the solution of the present application is specifically implemented as follows: when the entire wind direction of the target period is 90 degrees, first determine the target circular sector corresponding to the wind direction. Assume that the wind direction circle is divided into 16 sectors of equal angular width, and 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, that is, the areas corresponding to 78.75 degrees to 86.25 degrees and 93.75 degrees to 101.25 degrees. The meteorological forecast wind speed of the target period is input into the correction curve corresponding to the target sector respectively, and the correction curves of the two adjacent sectors are input at the same time to obtain three sector correction results. Calculate the angle between the central axis of each sector and the 90-degree wind direction. For example, the central axis of the target sector is 90 degrees, and the central axes of the adjacent sectors are 82.5 degrees and 97.5 degrees respectively. Calculate the inverse distance weight based on the absolute value of the angle, where the weight is inversely proportional to the absolute value of the angle. The weighted sum of the correction results of the three sectors is performed to obtain the predicted value of the correction curve.

[0100] Through the above technical solution, this application effectively reduces the impact of localized wind direction deviations on wind speed corrections. By integrating the correction results of adjacent sectors, it suppresses the boundary mutation errors that may be caused by single-sector divisions. This solution can adaptively balance the historical statistical patterns of adjacent wind direction regions, enhance the model's robustness to wind direction measurement errors and short-term fluctuations, and improve the spatial continuity of the correction results.

[0101] In some of the above-mentioned schemes of the present application, a method of dividing the circular sectors according to the entire wind direction and constructing a correction curve is proposed. However, when there are fluctuations in wind direction or measurement errors, the entire wind direction during the target period may deviate from the central axis of the sector, resulting in a single sector correction curve that cannot accurately reflect the actual wind speed distribution, and the sudden changes in wind speed between adjacent sectors are not effectively smoothed.

[0102] The present application further proposes to determine the target circular sector and its adjacent circular sectors corresponding to the entire wind direction during the target period, and substitute the meteorological forecast wind speed of 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 entire wind direction, and perform weighted summation of the sector correction results to obtain the correction curve prediction value.

[0103] Among them, the determination of adjacent sectors is based on the equal angular width of the wind direction circle. The number of adjacent sectors is set to two and is 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, and the weight coefficient is proportional to the square of the cosine value of the included angle. During the fusion process of the sector correction results, when the target wind direction is located at the junction of two adjacent sectors, the weight allocation threshold is set to an included angle of 15 degrees. When the included angle between the target wind direction and the central axis of any sector exceeds 45 degrees, only the two sectors with the largest weights are retained for calculation.

[0104] Specifically, the angle between the target wind direction and the central axis of each sector is obtained by vector dot product calculation, and the square of the cosine value of the calculated angle is used as the initial weight. The initial weight is normalized to ensure that the sum of the weights of 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 weight of the main sector is 0.85, and the adjacent left and right sectors receive weights of 0.10 and 0.05, respectively. This weighting method enables the correction results of adjacent sectors to produce a smooth transition effect on the final prediction value, effectively suppressing the prediction jump caused by wind direction measurement deviation or short-term fluctuations. Through inverse distance weighted fusion, the dominant role of the main sector correction curve is retained, and the adjacent sector data is used to correct the prediction error in the boundary area, thereby improving the correction stability under complex wind conditions.

[0105] As a preferred embodiment, the solution of this application is specifically implemented as follows: During the wind farm wind speed correction process, for the fusion of the normal weather branch prediction value, the extreme weather branch prediction value, and the correction curve prediction value, the mean absolute error and root mean square error of each prediction value on the validation set are first calculated as performance indicators. Furthermore, the analytic hierarchy process is used to determine subjective weights, with the subjective weight of the normal weather branch prediction value set to 0.4, the subjective weight of the extreme weather branch prediction value set to 0.3, and the subjective weight of the correction curve prediction value set to 0.3. Simultaneously, the objective weight of each prediction value is calculated based on the entropy weight method, with the objective weight of the normal weather branch prediction value set to 0.35, the objective weight of the extreme weather branch prediction value set to 0.25, and the objective weight of the correction curve prediction value set to 0.4. The subjective weight and the objective weight are linearly combined in a 1:1 ratio. After normalization, the resulting combined weights are 0.375, 0.275, and 0.35, respectively. Finally, the combined weight is linearly weighted summed with the three prediction values ​​to output the fused wind speed correction value.

[0106] Through the above technical solution, this application effectively balances the complementarity between the prediction results based on data-driven deep learning and the prediction results based on physical statistics correction curves, and solves the problem of large prediction deviations of a single model under complex weather conditions. By introducing a subjective and objective weight combination mechanism, it not only retains the qualitative judgment of the model performance based on expert experience, but also makes full use of the statistical characteristics of the data to quantitatively adjust the weight distribution, thereby enhancing the model's adaptability in scenarios where normal weather and extreme weather alternate. This fusion method reduces the fluctuations in prediction errors caused by abnormal meteorological events, and improves the continuity of wind speed correction results in the time dimension and the stability in the spatial dimension.

[0107] In the above embodiment, the measured power curves of each wind turbine in the wind farm are obtained and preprocessed to generate the overall power curve. The overall wind speed is then inverted and time-series paired with the overall wind direction. This constructs a multimodal training dataset encompassing meteorological characteristics, enabling 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. Its time-dependent properties are then captured using a long-short-term memory network. An attention mechanism is introduced into the output sequence to weight the importance of each time step, enhancing the model's ability to model wind speed trends. Two multi-layer perceptron branches, one for normal weather and one for extreme weather, enable learning prediction strategies under different meteorological conditions, improving responsiveness to extreme weather events. Spatially, the training dataset is partitioned into circular sectors based on the overall wind direction. Correction curves are constructed for each sector based on the relationship between historically predicted and measured wind speeds. Correction results from adjacent sectors are then integrated using an inverse distance weighting strategy to generate physically meaningful correction curve predictions. By integrating the normal weather forecast values, extreme weather forecast values ​​and correction curve forecast values, 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 adaptation to extreme situations.

[0108] In another preferred embodiment based on the above embodiment, refer to Figure 2 As shown, this embodiment provides a wind farm wind speed correction system based on dynamic spatiotemporal modeling, which is used to apply the above-mentioned wind farm wind speed correction method based on dynamic spatiotemporal modeling, including: The acquisition unit is configured to obtain the measured power curve of each wind turbine generator set in the wind farm, and pre-process the measured power curve to obtain the power curve of the entire farm; The correlation unit is configured to obtain the entire wind speed of the historical period based on the entire power curve, and perform time series correlation with the entire wind direction, while collecting meteorological characteristics of the corresponding period to construct a multimodal training data set; The first processing unit is configured to input the multimodal training data set into the convolutional neural network to extract local features of the wind speed spatiotemporal data, 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 the features of each time step and obtain the attention weight, and perform weighted aggregation on the time step features according to the attention weight to obtain a global feature representation; The second processing unit is configured to set two multi-layer perceptron branches, input the global feature representation to the corresponding multi-layer perceptron branch to perform wind speed prediction correction, and obtain a normal weather branch prediction value and an extreme weather branch prediction value; a third processing unit configured to divide the multimodal training dataset into a plurality of circular sectors according to wind direction, construct a correction curve for each sector based on the relationship between the historical forecast wind speed and the actual wind speed, input the meteorological forecast wind speed for the target period into the correction curve corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain a sector correction result, and fuse the sector correction results of the adjacent sectors through inverse distance weighting to obtain a predicted value of the correction curve; The correction unit is configured to obtain a final wind speed correction value based on a weighted fusion of the normal weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value.

[0109] Specifically, the acquisition unit normalizes the measured power curves of multiple units through a unified wind speed grid, performs linear interpolation and boundary constraint interpolation on the wind speed axis, and generates the power curve of the entire field by traversing and summing along the grid; the association unit aligns the meteorological characteristics with the entire wind direction based on a sliding time window, constructs a multimodal training data set containing derived features, and annotates the weather labels; the first processing unit uses a one-dimensional temporal convolutional neural network to extract local features, models long-term dependencies through a long short-term memory network, and introduces an attention mechanism to generate a global feature representation; the second processing unit sets up two independent multi-layer perceptron branches, activates the corresponding branches based on the weather labels for prediction, and configures a fully connected layer and a random inactivation layer for the normal weather branch and the extreme weather branch respectively; the third processing unit divides the hierarchical sectors according to the wind direction to construct a correction curve, and generates the correction curve prediction value by fusing the results of adjacent sectors through inverse distance weighting; the correction unit calculates the combination weight based on the performance index and performs weighted fusion on the prediction results of multiple branches.

[0110] 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 turbine models, fills in missing data points through interpolation, and ultimately outputs a full-field power curve. The correlation unit simultaneously receives meteorological monitoring data and full-field wind direction data, performs sine and cosine encoding on the wind direction to eliminate angular circularization errors, constructs a multimodal input matrix combining 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 time series tensor, extracts local wind speed fluctuation patterns using multi-scale convolution kernels, employs residual connections to prevent gradient vanishing, and uses an attention mechanism to filter key time-step features to form a global representation. The second processing unit dynamically freezes non-relevant branch parameters based on weather labels during the training phase. During the prediction phase, it performs normal and extreme weather forecasts in parallel, suppressing outlier interference using a quantile loss function. The third processing unit establishes a corrected wind speed mapping relationship between sectors based on historical data. During real-time prediction, it matches adjacent sectors based on wind direction angles, incorporating spatial proximity effects using inverse distance weighting. The correction unit dynamically adjusts weight distribution by calculating the prediction error rate of each branch, ultimately outputting a fused wind speed correction value. This system, through a modular design, enables parallel processing of data acquisition, feature extraction, multi-model prediction, and physical correction. This system improves response speed to extreme weather events while leveraging the advantages of statistical learning and physical models through a weighted fusion mechanism.

[0111] As a preferred embodiment, the solution of the present application is specifically implemented as follows: the wind speed correction system for a wind farm 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 curve from each wind turbine generator set in the wind farm, and generates the power curve of the entire field after preprocessing. The association unit infers the wind speed of the entire field in the historical period based on the power curve of the entire field, aligns it with the wind direction of the entire field in time, and integrates meteorological characteristics to construct a multimodal training data set. The first processing unit inputs the data set 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 in the feature sequence to calculate the weight of each time step, and forms a global feature representation after weighted aggregation. The second processing unit sets up two multi-layer perceptron branches to perform wind speed prediction correction for normal weather and extreme weather respectively. Non-related branch parameters are frozen during training, and predicted values ​​under two types of weather conditions are output in parallel during prediction. The third processing unit divides the data into multi-level circular sectors based on the overall wind direction. Correction curves are fitted to each sector based on the relationship between historical forecast wind speeds and actual wind speeds. The target period wind speed forecast is input into the correction curves for the corresponding sector and adjacent sectors. The correction curve predictions are then fused using inverse distance weighted fusion. The correction unit then combines the predictions from the two multi-layer perceptron branches with the correction curve predictions using a weighted combination to generate the final wind speed correction.

[0112] Through the above technical solutions, this application effectively solves the problems of insufficient spatiotemporal dynamic modeling, lack of processing of wind direction spatial heterogeneity, and weak extreme weather response capabilities in the existing technology. Through multimodal data fusion and attention mechanism weighting, the ability to capture the spatiotemporal evolution of wind speed is enhanced; the use of sector correction curves and inverse distance weighting strategies improves the adaptability to spatial heterogeneity under different wind directions; through the dual-branch multi-layer perceptron structure, differential processing of normal and extreme weather is achieved, enhancing the stability of predictions under extreme conditions; the combined weighting mechanism of multiple prediction results further improves the robustness and generalization of the correction results.

[0113] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0115] 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 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational 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 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0117] 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, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A wind farm wind speed correction method based on dynamic spatiotemporal modeling, characterized in that: include: Obtaining a measured power curve of each wind turbine generator set in the wind farm, and preprocessing the measured power curve to obtain a power curve for the entire farm; The entire wind speed of the historical period is obtained according to the entire power curve, and a time series correlation is performed with the entire wind direction. At the same time, the meteorological characteristics of the corresponding period are collected to construct a multimodal training data set; Inputting the multimodal training data set into a convolutional neural network to extract local features of the wind speed spatiotemporal data, and inputting the local features into a long short-term memory network, introducing an attention mechanism into the feature sequence output by the long short-term memory network, calculating the correlation of the features of each time step and obtaining attention weights, and performing weighted aggregation on the time step features according to the attention weights to obtain a global feature representation; Two multi-layer perceptron branches are set, and the multi-layer perceptron branches corresponding to the global feature representation input are used to perform wind speed prediction correction to obtain a normal weather branch prediction value and an extreme weather branch prediction value; Dividing the multimodal training data set into a plurality of circular sectors according to the entire wind direction, constructing a correction curve for each sector based on the relationship between the historical weather forecast wind speed and the entire wind speed, inputting the weather forecast wind speed for the target period into the correction curve corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain a sector correction result, and obtaining a correction curve prediction value 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 normal weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value.

2. The wind speed correction method for a wind farm based on dynamic spatiotemporal modeling according to claim 1 is characterized in that: Preprocessing the measured power curve to obtain the entire field power curve includes: The measured power curves of each wind turbine generator set are normalized on the wind speed axis to unify the wind speed grid. Missing power data are linearly interpolated at the corresponding wind speed points and duplicate or abnormal points are eliminated. Boundary constraints are set at the wind speed boundary according to the maximum rated wind speed model and boundary value interpolation is performed to ensure curve continuity, thereby obtaining the optimized power curve of each unit. The optimized power values ​​are added at each uniform wind speed point, and the sum is traversed along the wind speed grid to obtain the entire field power curve.

3. The wind speed correction method for a wind farm based on dynamic spatiotemporal modeling according to claim 2 is characterized in that: The entire wind speed of the historical period is obtained according to the entire power curve, and a time series correlation is performed with the entire wind direction. At the same time, the meteorological characteristics of the corresponding period are collected to construct a multimodal training dataset, including: Obtaining the entire field power within a historical period and establishing a monotonic correspondence between power and wind speed based on the entire field power curve, calculating the entire field wind speed at each moment on the wind speed grid based on reverse lookup and piecewise linear interpolation; obtaining the entire field wind direction representing the wind farm, de-ringing the angle data and digitizing it with sine and cosine coding, and time-aligning it with the entire field wind speed at a uniform sampling step; Collecting meteorological characteristics on the same time axis as the historical period and normalizing the meteorological characteristics, the meteorological characteristics including wind speed, wind direction, temperature, humidity, air pressure and boundary layer height; Samples are constructed based on a sliding time window. The meteorological characteristics within the window, the entire wind direction, and the derived characteristics are used as input. The entire wind speed corresponding to the window forward step is used as a label to form the multimodal training dataset. Weather labels for normal weather and extreme weather are annotated according to the wind force level threshold. The derived characteristics include wind vector component, wind shear index, wind direction turning rate, boundary layer stability index, and boundary layer height change rate.

4. The wind speed correction method for a wind farm based on dynamic spatiotemporal modeling according to claim 3 is characterized in that: When the multimodal training data set is input into a convolutional neural network to extract local features of wind speed spatiotemporal data, the method includes: The multimodal training data set is stacked according to the sliding time window in the time dimension and the meteorological features and the entire wind direction are channel-joined in the feature dimension to form a one-dimensional time series tensor; A convolutional neural network using one-dimensional temporal convolutions, causal padding in the first layer, and multi-scale convolution kernels with optional dilated convolutions in the middle layers. The kernel length of the multi-scale convolution kernel is one or more of 3, 5, and 7. Nonlinear activation is set in sequence after each convolutional layer and deep layer degradation is suppressed through residual connection, and finally the local features are output.

5. The wind speed correction method for a wind farm based on dynamic spatiotemporal modeling according to claim 4 is characterized in that: Calculating the correlation of each time step feature and obtaining the attention weight, performing weighted aggregation on the time step features according to the attention weight, and obtaining a global feature representation, including: Inputting the local features into a stacked long short-term memory network to model long-term dynamic dependencies and obtain a hidden state sequence; Calculating the relevance score based on the attention mechanism and obtaining the attention weight through Softmax normalization; The time step features are weightedly aggregated according to the attention weight to obtain a context vector, and the context vector is nonlinearly projected to obtain the global feature representation.

6. The wind speed correction method for a wind farm based on dynamic spatiotemporal modeling according to claim 5 is characterized in that: Two multi-layer perceptron branches are set, and the multi-layer perceptron branch corresponding to the global feature representation input is used to perform wind speed prediction correction, and the prediction value of the normal weather branch and the prediction value of the extreme weather branch are obtained, including: When setting two multi-layer perceptron branches, including a normal weather multi-layer perceptron branch and an extreme weather multi-layer perceptron branch; Each of the multi-layer perceptron branches includes at least two fully connected layers and nonlinear activation, and random inactivation is set between the layers to suppress overfitting; in the training phase, only the parameters of the corresponding multi-layer perceptron branch are updated based on the weather label, and the other multi-layer perceptron branch is frozen; in the prediction phase, the global feature representation is forward-inputted into the two multi-layer perceptron branches respectively to simultaneously generate the prediction value of the normal weather branch and the prediction value of the extreme weather branch, and the quantile loss is used as the robust loss function to constrain the training error.

7. The wind speed correction method for a wind farm based on dynamic spatiotemporal modeling according to claim 6, characterized in that: Dividing the multimodal training data set into a plurality of circular sectors according to the entire wind field, and constructing a correction curve for each sector based on the relationship between the historical meteorological forecast wind speed and the entire wind field speed, including: The wind direction circle is divided into the circular sectors of equal angular width using the entire wind direction as an angle variable, and the sector extension is divided into a plurality of ring belts according to the wind speed level to form a hierarchical structure; In each of the circular sectors, monotonic constrained regression fitting is performed with the historical weather forecast wind speed as the independent variable and the entire wind speed as the dependent variable to obtain the correction curve.

8. The wind speed correction method for a wind farm based on dynamic spatiotemporal modeling according to claim 7 is characterized in that: Inputting the meteorological forecast wind speed for the target period into the correction curve corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain a sector correction result, and fusing the sector correction results of the adjacent sectors by inverse distance weighted fusion to obtain a correction curve prediction value, including: Determine the target circular sector and its adjacent circular sectors corresponding to the entire wind direction during the target period, substitute the meteorological forecast wind speed during 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 entire wind direction, and perform weighted summation of the sector correction results to obtain the predicted value of the correction curve.

9. The wind speed correction method for a wind farm based on dynamic spatiotemporal modeling according to claim 8, characterized in that: When obtaining the final wind speed correction value according to the weighted fusion of the normal weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value, the method includes: Performance indicators are calculated for the normal weather branch prediction value, the extreme weather branch prediction value, and the correction curve prediction value, respectively, and subjective weights and objective weights are determined based on the performance indicators. Weight normalization and non-negative constraints are used to obtain a combined weight, and the sum of all weights in the combined weight is 1. The final wind speed correction value is calculated based on the combined weight and the normal weather branch prediction value, the extreme weather branch prediction value, and the correction curve prediction value.

10. 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 according to any one of claims 1 to 9, characterized in that: include: A collection unit is configured to obtain a measured power curve of each wind turbine generator set in the wind farm, and pre-process the measured power curve to obtain a power curve of the entire farm; a correlation unit configured to obtain the entire wind speed of a historical period according to the entire power curve, perform time series correlation with the entire wind direction, and simultaneously collect meteorological characteristics of the corresponding period to construct a multimodal training data set; A first processing unit is configured to input the multimodal training data set into a convolutional neural network to extract local features of the 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 the features of each time step and obtain attention weights, and perform weighted aggregation on the time step features according to the attention weights to obtain a global feature representation; The second processing unit is configured to set two multi-layer perceptron branches, input the global feature representation into the corresponding multi-layer perceptron branch to perform wind speed prediction correction, and obtain a normal weather branch prediction value and an extreme weather branch prediction value; a third processing unit configured to divide the multimodal training data set into a plurality of circular sectors according to wind direction, construct a correction curve for each sector based on a relationship between historical wind speed forecasts and actual wind speeds, input the meteorological forecast wind speed for a target period into the correction curve corresponding to the sector where the target wind direction is located and its adjacent sectors to obtain a sector correction result, and fuse the sector correction results of the adjacent sectors by inverse distance weighted fusion to obtain a predicted value of the correction curve; The correction unit is configured to obtain a final wind speed correction value based on weighted fusion of the normal weather branch prediction value, the extreme weather branch prediction value and the correction curve prediction value.

Images