Offshore wind farm cluster ultra-short-term power prediction method considering atmospheric stability effect

Abstract

This invention relates to a method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability. The method includes: acquiring the structure and historical operating parameters of the offshore wind power cluster; performing numerical simulations; calculating the Monin-Obukhov length to determine atmospheric stability; dynamically scaling the hyperedge search radius; combining real-time risk and inter-site azimuth angles to calculate the deflection angle difference and introduce a cosine attenuation operator; and constructing a dynamic wake hypergraph topology matrix based on the hyperedge search radius and the cosine attenuation operator. The structure, historical operating parameters, and dynamic wake hypergraph topology matrix of the offshore wind power cluster are input into an asymmetric split-integration spatiotemporal module to obtain the predicted spatiotemporal characteristics. The method then performs site-by-site predictions for the offshore wind power cluster and finally aggregates the results to obtain the total cluster power. Compared with existing technologies, this invention can accurately characterize high-order wake coupling between sites, significantly suppress non-stationary meteorological noise, and exhibits higher prediction accuracy, robustness, and physical interpretability under small sample sizes and complex marine conditions.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power cluster power prediction technology, and in particular to an ultra-short-term power prediction method for offshore wind power clusters that takes into account the influence of atmospheric stability. Background Technology

[0002] Currently, offshore wind power has become a core engine supporting the green transformation of modern power systems. The large-scale development of offshore wind farms has induced extremely complex wake effects between farm clusters. Affected by the low-turbulence environment at sea, the wake evolution distance can reach tens of kilometers, resulting in power losses of up to 10% to 20% for downstream power plants. This poses unprecedented challenges to the time-series coordinated scheduling of the power grid. Accurate power prediction of offshore wind power clusters is not only key to reducing scheduling costs and improving absorption levels, but also the cornerstone for ensuring the safe and stable operation of a power grid with a high proportion of renewable energy.

[0003] With the rapid development of offshore wind power towards large capacity and clustering, significant wake interactions between wind farms and long propagation distances, coupled with real-time modulation by atmospheric stability, lead to drastic power fluctuations. Traditional prediction methods often employ static graph structures or purely data-driven models, which struggle to characterize high-order wake couplings and the dynamic effects of atmospheric stratification. These methods suffer from insufficient spatial representation, poor robustness to non-stationary conditions, and weak physical interpretability.

[0004] Existing dynamic graph models only achieve binary correlation adaptive updates, failing to characterize the simultaneous influence of upstream wakes on multiple stations. Most methods do not explicitly embed atmospheric stability into the topology and convolution process, making it difficult for models to learn physical properties such as long wakes in stable strata and short wakes in unstable strata. Therefore, a high-precision prediction method that integrates WRF physical fields, dynamic wake hypergraphs, and atmospheric stability gating is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a method for ultra-short-term power prediction of offshore wind power clusters that integrates WRF physical field, dynamic wake hypergraph and atmospheric stability, taking into account the influence of atmospheric stability.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability includes the following steps: The structure and historical operating parameters of the offshore wind power cluster are obtained to conduct numerical simulations of the offshore wind power cluster area, calculate the Monin-Obukhov length, and thus determine the atmospheric stability. Based on the atmospheric stability, the hyperedge search radius is dynamically scaled, and the deviation angle difference is calculated by combining real-time risk and azimuth between stations to introduce a cosine attenuation operator. Based on the hyperedge search radius and the cosine attenuation operator, a dynamic wake hypergraph topology matrix driven by wind direction, wake intensity and atmospheric stability is constructed. The structure, historical operating parameters, and dynamic wake hypergraph topology matrix of the offshore wind power cluster are input into a pre-trained asymmetric split-integration spatiotemporal module to obtain the predicted spatiotemporal features. Based on the predicted spatiotemporal characteristics, the offshore wind power cluster is predicted field by field, and then the total predicted power of the cluster is obtained by summing them up.

[0007] Furthermore, the expression for calculating the matrix weights of the dynamic wake hypergraph topology matrix is ​​as follows: In the formula, Let the weight of the hyperedge between station i and station j at time t be denoted as . Let be the actual straight-line geographical distance between station i and station j. The maximum deviation angle, Let be the wake intensity factor of station i at time t. For the angle difference, For cosine attenuation operator, Let be the hyperedge search radius at time t.

[0008] Furthermore, the expression for calculating the hyperedge search radius is as follows: In the formula, Let be the hyperedge search radius at time t. The baseline search radius is set based on the average geographical spacing of offshore wind farm clusters. For the Morning-Obukhov length based on time t Determined atmospheric stability.

[0009] Furthermore, the process of determining the atmospheric stability includes: The corresponding atmospheric state is determined based on the range of the Monin-Obukhov length, which is either stable, neutral, or unstable. The formula for calculating atmospheric stability is as follows: In the formula, For the Morning-Obukhov length based on time t Determined atmospheric stability, This is the gain control coefficient, combined with the maximum upper limit of scaling the actual control radius of the wind farm during operation; The size is ,in, For wheel hub height, The sign of is: when the atmospheric state is stable. It is a positive value; when the atmospheric state is unstable, It is a negative value.

[0010] Furthermore, the calculation expression for the wake intensity factor is as follows: In the formula, The wake intensity factor, The actual wind speed at the wheel hub height after the wake effect. The free-flow wind speed is unaffected by the wake.

[0011] Furthermore, the Moning-Obukhoff length is calculated based on hub height wind speed, wind direction, sensible heat flux, and friction speed.

[0012] Furthermore, the asymmetric split-flow integrated spatiotemporal module includes a time-inertial split-flow branch, a space-physical split-flow branch, and a spatiotemporal integrated output layer; The time-inertial shunt branch mines historical power sequences to extract the temporal inertial features of the power evolution of the power station itself; The spatial physical diversion branch generates a diagonal gated matrix based on atmospheric stability through multilayer perceptron mapping; generates a spatial feature matrix by data projection based on the station's physical characteristics; generates physically constrained spatial features by feature extraction based on the dynamic wake hypergraph topology matrix; and fuses the diagonal gated matrix, the spatial feature matrix, and the physically constrained spatial features to form physical interactive features. The spatiotemporal integration output layer adaptively aligns and splices the temporal inertial features and physical interaction features, and then feeds them into the BiLSTM network to further capture bidirectional temporal dependencies and output predicted spatiotemporal features.

[0013] Furthermore, the time-inertial shunt branch utilizes multi-scale dilated convolution to mine historical power sequences, capture multi-scale temporal dependencies, and extract the temporal inertial features of the power evolution of the power station itself.

[0014] Furthermore, the physical gating matrix is ​​used to adjust the transmission threshold of wake information between stations in real time, enhance the transmission of long-distance wake characteristics in stable strata, and suppress turbulence noise in unstable strata.

[0015] Furthermore, the method also includes: during the training of the asymmetric split-integration spatiotemporal module, constructing a joint error metric for model accuracy verification, wherein the joint error metric includes MAE, RMSE, and R. 2 index.

[0016] Compared with the prior art, the present invention has the following advantages: (1) The present invention first constructs a dynamic wake hypergraph topology matrix based on the data simulation results of the target offshore wind power cluster area, and then predicts the spatiotemporal characteristics based on the structure, historical operating parameters and dynamic wake hypergraph topology matrix of the target offshore wind power cluster. Based on the predicted spatiotemporal characteristics, the offshore wind power cluster is predicted field by field, and then the predicted total power of the cluster is obtained by summarizing. In the process of constructing the dynamic wake hypergraph topology matrix, unlike the traditional static adjacency matrix, this invention dynamically scales the search radius of the hyperedge with atmospheric stability in order to simulate the physical nature of wake propagation distance being longer under stable strata and dissipation being faster under unstable strata. After determining the search radius, the deviation angle difference is calculated using real-time wind direction and azimuth angle between stations to construct a cosine attenuation operator to solve the numerical instability problem caused by wind direction jumps, thereby achieving a significant improvement in physical interpretability.

[0017] (2) In the process of constructing the dynamic wake hypergraph topology matrix of this invention, the matrix weight of its own node is set to a fixed value, which strengthens the preservation of the central station's own features in matrix operations. During convolution aggregation, this ensures that the station's own historical power features are not completely submerged by the features of neighboring nodes.

[0018] (3) The present invention effectively suppresses non-stationary meteorological noise by dynamically adjusting the characteristic transmission through stability gating, and has higher robustness.

[0019] (4) The asymmetric split-integrated spatiotemporal module constructed in this invention is divided into a time inertial split-in branch and a spatial physical split-in branch, realizing the decoupling and complementary fusion of spatiotemporal features, and has obvious accuracy advantages in cluster total power prediction. Attached Figure Description

[0020] Figure 1 This is a wind turbine thrust coefficient versus power curve provided in an embodiment of the present invention; Figure 2 This is a 24-hour average wake loss heat map of each wind field provided in an embodiment of the present invention; Figure 3 This is a wake hypergraph topology diagram under different atmospheric stability conditions provided in an embodiment of the present invention; Figure 4 This is a diagram of a STHGDN spatiotemporal hypergraph dynamic network architecture provided in an embodiment of the present invention; Figure 5 This is a spatial distribution map of the wake of an offshore wind power cluster provided in an embodiment of the present invention; Figure 6 This is a comparison chart of wind speed differences under different atmospheric stability conditions provided in an embodiment of the present invention; Figure 7This is a comparison chart of prediction accuracy indicators for various wind farm stations provided in an embodiment of the present invention; Figure 8 This is a comparison chart of single-field prediction errors using multiple models provided in an embodiment of the present invention; Figure 9 This is a comparison chart of the power prediction effects of various models for a typical power station provided in an embodiment of the present invention; Figure 10 This is a comparison chart of the prediction effects of various models for the total power of a power field provided in an embodiment of the present invention; Figure 11 This is a flowchart illustrating an ultra-short-term power prediction method for offshore wind power clusters that takes into account the influence of atmospheric stability, as provided in the embodiments. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0024] Example 1 like Figure 11 As shown in the figure, this embodiment provides a method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability, including the following steps: S1: Obtain the structure and historical operating parameters of the offshore wind power cluster to conduct numerical simulation of the offshore wind power cluster area, calculate the Moning-Obukhov length, and thus determine the atmospheric stability. S2: Based on atmospheric stability, the hyperedge search radius is dynamically scaled. The deviation angle difference is calculated by combining real-time risk and azimuth between stations to introduce a cosine attenuation operator. Based on the hyperedge search radius and cosine attenuation operator, a dynamic wake hypergraph topology matrix driven by wind direction, wake intensity and atmospheric stability is constructed. S3: Input the structure, historical operating parameters and dynamic wake hypergraph topology matrix of the offshore wind power cluster into the pre-trained asymmetric split-integration spatiotemporal module to obtain the predicted spatiotemporal features; S4: Based on the predicted spatiotemporal characteristics, the offshore wind power cluster is predicted field by field, and then the total predicted power of the cluster is obtained by summing them up.

[0025] Specifically, in step S1, a refined simulation of the marine boundary layer is performed based on WRF-WFP, extracting physical quantities such as hub height wind speed, wind direction, sensible heat flux, and friction velocity, calculating the Moning-Obukhov length to achieve atmospheric stability classification, and quantifying the interfield wake intensity factor through Wake-On and Wake-Off simulations. In this embodiment, a three-layer bidirectional nested grid is specifically used to perform numerical simulation on the target offshore wind power cluster area, with horizontal resolutions of 27km, 9km, and 3km respectively. The initial field and boundary conditions are constructed using ERA5 global reanalysis data.

[0026] The WRF physical parameterization scheme is configured as follows: WSM6-class for microphysics, RRTM for longwave radiation, Dudhia for shortwave radiation, Noah for land surface, MYNN2 for boundary layer, and Kain-Fritsch for cumulus convection. The thrust coefficient and power curves of the wind turbine used in the experiment are shown below. Figure 1 As shown.

[0027] Two parallel simulations were conducted using a parameterized WFP scheme with the wind farms turned on and off. The wake intensity factor was calculated using the difference between hub height and wind speed to quantify the wake shading effect between wind farms. The 24-hour average wake loss heat distribution for each wind farm is shown below. Figure 2 As shown.

[0028] This study defines the wake intensity factor. To quantify this physical shading effect, during data extraction, parallel simulation results with the wind farm parameterization scheme enabled (Wake-On) and disabled (Wake-Off) are compared to calculate the energy percentage of each cluster at hub height. The calculation formula is as follows: In the formula: This indicates the actual wind speed at wheel hub height after being affected by the wake (Wake-On scenario). This indicates the free-flow wind speed unaffected by the wake (Wake-Off scenario).

[0029] Extract the surface sensible heat flux HFX and friction velocity UST, and calculate the Monin-Obukhov length. L The atmospheric state is divided into three categories: stable, neutral, and unstable, thus completing the aggregation of atmospheric stability at the field group level.

[0030] Mourning-Obuhoff length The physical expression is: In the formula: air density, For isobaric specific heat, This is a virtual temperature. For friction speed, Let be the von Kármán constant (taken as 0.4). It is the acceleration due to gravity. This refers to the sensible heat flux at the Earth's surface.

[0031] Based on the calculation Based on relevant meteorological standards and the actual operation of wind farms, this study divides atmospheric conditions into three core categories (stable, neutral, and unstable).

[0032] when When the stratification is defined as unstable, thermal convection is intense and the wake dissipates rapidly; when When the vertical momentum exchange is hindered, the wake undergoes long-range, persistent evolution; when At this point, it is defined as a neutral stratification, dominated by pure mechanical shear.

[0033] The determining logic is based on the mechanism of atmospheric layer binding inhibiting and promoting turbulent mixing. A smooth adjustment factor is constructed using the hyperbolic tangent function, making... The output fits the physical boundary. In the formula: , This refers to the wheel hub height. It is an odd function, when the atmosphere is stable hour, When unstable hour, . It is the gain control coefficient, which is the maximum upper limit of scaling the actual control radius of the wind farm.

[0034] when In a stable stratification, the sea surface cold inversion layer inhibits vertical momentum exchange, resulting in a narrow and long wake that exhibits a long-range, persistent propagation characteristic. By expanding the hyperedge search radius, the model's ability to capture higher-order nonlinear interactions of distant receptor nodes is explicitly enhanced; conversely, when... In the case of (unstable stratification), strong thermal convection accelerates the recovery of wake energy, and the propagation distance is significantly reduced. At this point, the setting... By shrinking the hyperboundary, noise interference from distant irrelevant nodes is adaptively cut off.

[0035] In step S2, to capture the complex nonlinear wake interactions among offshore wind farm clusters, a dynamic wake hypergraph topology matrix based on physical constraints is constructed. Unlike traditional static adjacency matrices, this topology not only rotates with the wind direction but also achieves fine-grained modulation of spatial association strength by introducing a continuous attenuation operator and a wake intensity factor. The elements of the association matrix define the physical membership degree of a node to a hyperedge, and the assignment logic combines the geographic search boundary and the wake loss intensity. This invention provides spatial topology visualization, as shown below. Figure 3 As shown, each column of the matrix maps to a hyperedge patch in the diagram, and the overall wind direction of the field group is indicated by arrows. When the atmosphere is in a stable stratification, the matrix has a large spatial distribution of non-zero elements, corresponding to an expanding patch with wide coverage and many member nodes in the diagram; when the atmosphere is in an unstable stratification, it is subject to physical gating, and the matrix becomes sparse, corresponding to a degenerate topology with shrinking area and broken connections in the diagram.

[0036] Specifically, the construction process of the dynamic wake hypergraph topology matrix is ​​as follows: To simulate the physical nature of wakes propagating further under stable strata and dissipating faster under unstable strata, the search radius of the hyperedge is dynamically scaled with atmospheric stability: In the formula, Let be the hyperedge search radius at time t. The baseline search radius is set based on the average geographical spacing of offshore wind farm clusters. For the Morning-Obukhov length based on time t Determined atmospheric stability.

[0037] After determining the search radius, use real-time wind direction. Azimuth between stations Calculate the difference in deflection angle A cosine attenuation operator is introduced to address numerical instability caused by abrupt wind direction changes. The weights of the topology matrix are no longer simple binary values, but are instead determined through... Synthesis with continuous decay operator: In the formula: This represents the actual straight-line geographical distance between station i (as the center or upstream source of the superedge) and station j (as a potentially affected node); The maximum deviation angle; It is the wake intensity factor of station i at time t; This is to enhance the preservation of the central power station's own features during matrix operations. During convolutional aggregation, this ensures that the station's own historical power features are not completely overwhelmed by the features of its neighboring nodes.

[0038] Ultimately, this mechanism transforms the endogenous physical quantities output by WRF-WFP into tensors with spatiotemporal evolution properties. Through Adjust the rotation direction of the super-edge in real time, through Change the connection range, utilize As a basic value, it reflects the real-time energy deficit between power stations.

[0039] In step S3, the asymmetric split-flow integrated spatiotemporal dynamic hypergraph dynamic network STHGDN built based on the ASI-STM architecture mainly includes three parts: 1) Physically guided dynamic evolution hypergraph convolutional branch PE-DHGCN; 2) Temporal inertial split-flow branch based on multi-scale dilated convolution TCN; 3) Spatiotemporal integrated output layer based on BiLSTM to long short-term memory network.

[0040] The time-inertia shunt branch mines historical power sequences to extract the temporal inertial characteristics of the power evolution of the power station itself; The spatial physics diversion branch generates a diagonal gated matrix based on atmospheric stability through multilayer perceptron mapping; it generates a spatial feature matrix by projecting data based on the station's physical characteristics; it extracts features based on the dynamic wake hypergraph topology matrix to generate physically constrained spatial features; and it fuses the diagonal gated matrix, the spatial feature matrix, and the physically constrained spatial features to form physical interactive features. The spatiotemporal integration output layer adaptively aligns and concatenates temporal inertial features and physical interaction features, and then feeds them into the BiLSTM network to further capture bidirectional temporal dependencies and output predicted spatiotemporal features.

[0041] The time-inertial shunt branch uses multi-scale dilated convolution to mine historical power sequences, capture multi-scale temporal dependencies, and extract the temporal inertial features of the power evolution of the power station itself.

[0042] The physical gating matrix is ​​used to adjust the transmission threshold of wake information between stations in real time, enhance the transmission of long-distance wake characteristics in stable strata, and suppress turbulence noise in unstable strata.

[0043] Essentially, the STHGDN model employs an asymmetric splitting strategy, feeding the extracted features into the ASI-STM module for asymmetric splitting processing: the dynamic hypergraph topology matrix and the station's physical features are input into the spatial branch. A multi-layer perceptron (MLP) is used to map atmospheric stability parameters to generate a physical gating matrix, adjusting the transmission threshold of wake information between stations in real time. This enhances long-distance feature transmission in stable strata and suppresses turbulence noise in unstable strata. Simultaneously, multi-scale dilated convolution (TCN) is used for deep mining of historical power sequences to extract the temporal inertial features of the station's own power evolution. The STHGDN spatiotemporal hypergraph dynamic network architecture is as follows: Figure 4 As shown.

[0044] STHGDN (Spatio-Temporal Hypergraph Dynamic Network) is the overall model architecture. It is responsible for handling the complete end-to-end process from the initial input (WRF meteorological data, historical power, wake characteristics) to the final output (predicted total power of the cluster).

[0045] The ASI-STM (Asymmetric Shunting Integrated Spatio-Temporal Module) is the most critical core computing unit within STHGDN. Most of STHGDN's innovations (such as temporal inertia extraction, physical gating mechanisms, and dynamic hypergraph convolution) are implemented within the ASI-STM module.

[0046] Therefore, the STHGDN model employs an asymmetric splitting strategy to feed extracted features into the ASI-STM module. This process involves the following steps: During the feature processing phase of the model architecture, after the STHGDN model has completed its perception of multidimensional heterogeneous data at the input end, it uses an asymmetric splitting strategy to precisely decouple the features. Historical power and wind speed sequences are separately extracted and fed into the temporal splitting branch of the ASI-STM module; while physical field features such as power, wake factor, and atmospheric stability are fed into the spatial physics branch of the ASI-STM module in parallel. This asymmetric cascading mechanism, which decouples data flow and control flow, explicitly injects highly interpretable physical priors into the early stages of feature entry into the core module.

[0047] This process adaptively aligns and splices the physical interaction characteristics output by the spatial branch with the temporal inertia characteristics output by the temporal branch, and feeds them into the BiLSTM network to further capture bidirectional temporal dependencies and filter temporal noise, making the prediction curve smoother and more in line with the physical laws of continuous fluctuations in offshore wind power.

[0048] In step S4, the fused spatiotemporal feature stream is further propagated within the STHGDN model and input into its global temporal refinement layer (BiLSTM) for in-depth global temporal dependency mining. First, predictions are made field-by-field, then the total cluster power is obtained by summarizing the data. MAE, RMSE, and RMS are then used as the metrics. 2 A joint error index was constructed using various indicators to complete the accuracy verification.

[0049] like Figure 5 The image shows the wake space distribution of an offshore wind farm cluster. In large-scale wind farm clusters, upstream wind farms extract kinetic energy, leading to a decrease in downwind speed and an increase in turbulence intensity. This energy loss directly affects the power output of downstream wind farms. Figure 5 As shown, a significant wind speed attenuation zone appears in the wind power cluster area due to the superposition of wake currents. Figure 5 It can be observed that, under the influence of wake interaction, a continuous wind speed attenuation zone of up to 137 km is generated in the wind power cluster area. Figure 5 b shows that the width of the wake formed in the dense local wind field under the southeasterly wind direction is about 58 km.

[0050] like Figure 6 The image shows a comparison of wind speed differences under different atmospheric stability conditions. The wind speed difference between stable and unstable atmospheres is mostly less than zero, clearly indicating that wind speeds in stable atmospheres are lower than those in unstable atmospheres. This is a manifestation of the long wake effect: the wake in unstable atmospheres has recovered, while the wake in stable atmospheres continues, causing the wind speed deficit to be amplified in this region.

[0051] like Figure 7 The figure shown is a comparison of the prediction accuracy indicators of each wind farm station. After introducing atmospheric stability gating, the prediction accuracy of each station is significantly improved.

[0052] like Figure 8 As shown in the figure, the prediction error of the STHGDN model in this invention is significantly lower than that of the LSTM, GRU, TCN, AGCN-TCN and other comparative models.

[0053] like Figure 9 As shown, this is a comparison chart of the power prediction effects of various models for typical power stations. The curve of this invention is closest to the measured value.

[0054] like Figure 10 The figure shows a comparison of the prediction results of various models for the total power of the power field cluster. The present invention shows the best fit in predicting the total power of the power field cluster, with R0. 2 It can reach over 99%.

[0055] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability, characterized in that, Includes the following steps: The structure and historical operating parameters of the offshore wind power cluster are obtained to conduct numerical simulations of the offshore wind power cluster area, calculate the Monin-Obukhov length, and thus determine the atmospheric stability. Based on the atmospheric stability, the hyperedge search radius is dynamically scaled, and the deviation angle difference is calculated by combining real-time risk and azimuth between stations to introduce a cosine attenuation operator. Based on the hyperedge search radius and the cosine attenuation operator, a dynamic wake hypergraph topology matrix driven by wind direction, wake intensity and atmospheric stability is constructed. The structure, historical operating parameters, and dynamic wake hypergraph topology matrix of the offshore wind power cluster are input into a pre-trained asymmetric split-integration spatiotemporal module to obtain the predicted spatiotemporal features. Based on the predicted spatiotemporal characteristics, the offshore wind power cluster is predicted field by field, and then the total predicted power of the cluster is obtained by summing them up.

2. The method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability as described in claim 1, characterized in that, The expression for calculating the matrix weights of the dynamic wake hypergraph topology matrix is ​​as follows: In the formula, Let the weight of the hyperedge between station i and station j at time t be denoted as . Let be the actual straight-line geographical distance between station i and station j. The maximum deviation angle, Let be the wake intensity factor of station i at time t. For the angle difference, For cosine attenuation operator, Let be the hyperedge search radius at time t.

3. The method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability as described in claim 1, characterized in that, The expression for calculating the hyperedge search radius is: In the formula, Let be the hyperedge search radius at time t. The baseline search radius is set based on the average geographical spacing of offshore wind farm clusters. For the Morning-Obukhov length based on time t Determined atmospheric stability.

4. The method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability as described in claim 1, characterized in that, The process of determining atmospheric stability includes: The corresponding atmospheric state is determined based on the range of the Monin-Obukhov length, which is either stable, neutral, or unstable. The formula for calculating atmospheric stability is as follows: In the formula, For the Morning-Obukhov length based on time t Determined atmospheric stability, This is the gain control coefficient, combined with the maximum upper limit of scaling the actual control radius of the wind farm during operation; The size is ,in, For wheel hub height, The sign of is: when the atmospheric state is stable. It is a positive value; when the atmospheric state is unstable, It is a negative value.

5. The method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability as described in claim 1, characterized in that, The formula for calculating the wake intensity factor is as follows: In the formula, The wake intensity factor, The actual wind speed at the wheel hub height after the wake effect. The free-flow wind speed is unaffected by the wake.

6. The method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability as described in claim 1, characterized in that, The Moning-Obukhoff length is calculated based on hub height, wind speed, wind direction, sensible heat flux, and friction speed.

7. The method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability as described in claim 1, characterized in that, The asymmetric split-flow integrated spatiotemporal module includes a time-inertial split-flow branch, a space-physical split-flow branch, and a spatiotemporal integrated output layer; The time-inertial shunt branch mines historical power sequences to extract the temporal inertial features of the power evolution of the power station itself; The spatial physical diversion branch generates a diagonal gated matrix based on atmospheric stability through multilayer perceptron mapping; generates a spatial feature matrix by data projection based on the station's physical characteristics; generates physically constrained spatial features by feature extraction based on the dynamic wake hypergraph topology matrix; and fuses the diagonal gated matrix, the spatial feature matrix, and the physically constrained spatial features to form physical interactive features. The spatiotemporal integration output layer adaptively aligns and splices the temporal inertial features and physical interaction features, and then feeds them into the BiLSTM network to further capture bidirectional temporal dependencies and output predicted spatiotemporal features.

8. The method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability according to claim 7, characterized in that, The time-inertial shunt branch uses multi-scale dilated convolution to mine historical power sequences, capture multi-scale temporal dependencies, and extract the temporal inertial features of the power evolution of the power station itself.

9. A method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability, as described in claim 7, is characterized in that... The physical gating matrix is ​​used to adjust the transmission threshold of wake information between stations in real time, enhance the transmission of long-distance wake characteristics in stable strata, and suppress turbulence noise in unstable strata.

10. A method for ultra-short-term power prediction of offshore wind power clusters considering the influence of atmospheric stability, as described in claim 7, is characterized in that... The method further includes: during the training of the asymmetric split-integration spatiotemporal module, constructing a joint error metric for model accuracy verification, wherein the joint error metric includes MAE, RMSE, and R. 2 index.

Images