Machine learning based offshore wind pile scour rate quantitative prediction method
A physics-driven attention-guided dual-channel diffusion network constructed through machine learning achieves coordinated protection against local and overall scour of offshore wind turbine pile foundations, improving protection effectiveness and economic benefits, and solving the problem of lack of coordinated protection in existing technologies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUANENG RUDONG BAXIANJIAO OFFSHORE WIND POWER GENERATION CO LTD
- Filing Date
- 2025-08-13
- Publication Date
- 2026-07-03
AI Technical Summary
In existing technologies, there is a lack of coordinated protection schemes for local and overall scour of offshore wind turbine foundations, resulting in limited protection effectiveness and high costs.
By employing a machine learning-based approach, a physical-driven attention-guided dual-channel diffusion network is constructed by collecting multi-source topographic observation data and tidal current and sediment dynamic field data. This network enables the coordinated prediction of local scour rates and global state evolution rates, and generates a coordinated protection scheme.
It significantly improves the protection effect and economic benefits of offshore wind power pile foundations, enhances the spatiotemporal resolution and physical consistency of scour rate and depth prediction, and realizes automated collaborative protection design.
Smart Images
Figure CN121009822B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of offshore wind turbine pile technology, and in particular to a method for quantitatively predicting the scour rate of offshore wind turbine pile foundations based on machine learning. Background Technology
[0002] Offshore wind turbine foundations are subjected to the dynamic forces of waves and tides in the marine environment for extended periods, making them susceptible to localized and overall scour, which severely impacts the stability and safety of the foundations. Current technologies primarily employ protective measures designed individually for either localized or overall scour, lacking a comprehensive solution for coordinated protection. This results in limited protective effectiveness and high costs.
[0003] Therefore, there is an urgent need for a collaborative protection optimization method that can simultaneously solve both local and overall scour problems. Summary of the Invention
[0004] One objective of this invention is to propose a quantitative prediction method for the scour rate of offshore wind turbine pile foundations based on machine learning. This invention achieves coordinated protection optimization design for local and overall scour of offshore wind turbine pile foundations, significantly improving protection effectiveness and economic benefits.
[0005] A method for quantitatively predicting the scour rate of offshore wind turbine foundations based on machine learning, according to an embodiment of the present invention, includes:
[0006] Historical multi-source topographic observation data were collected, and the morphological features of sand ridges and sand troughs in the pre-set analysis area around the offshore wind power pile foundation were extracted. The data were then vectorized in a unified coordinate system to generate a topographic spatiotemporal evolution map.
[0007] Data on tidal current and sediment dynamic fields were collected and combined with observations from an acoustic Doppler current profiler to construct a coupled mathematical model of tidal current and sediment, thereby obtaining calibrated tidal current and sediment dynamic field data.
[0008] Construct a learnable terrain perturbation mapper to output a digital seabed elevation field at the target prediction time;
[0009] A physical-driven attention-guided dual-channel diffusion network, including a first channel and a second channel, is constructed using the seabed digital elevation field at the target prediction time and calibrated tidal and sediment dynamic field data.
[0010] In each diffusion-anti-diffusion iteration of the first and second channels, a physical-driven attention mechanism is introduced to assign scale weights to the digital elevation field of the seabed at the target prediction time, thereby obtaining the local scour rate prediction field and the global state evolution rate prediction field, respectively.
[0011] Construct a coupling and fusion layer to output the predicted scour rate and scour depth data of the area surrounding the pile foundation.
[0012] The predicted scour rate data is compared with the field observation data of scour pit depth, and the data are jointly corrected to obtain an updated model, generating updated scour rate prediction data and updated scour depth prediction data.
[0013] Based on the updated scour rate prediction data and the updated scour depth prediction data, a collaborative protection scheme for pile foundations is automatically generated.
[0014] Optionally, the generation of the terrain spatiotemporal evolution map includes:
[0015] Define the analysis area around the offshore wind turbine pile foundation. The analysis area around the offshore wind turbine pile foundation is a circular area with the center of the pile foundation as the center and the radius as the analysis radius.
[0016] Historical multi-source topographic observation data were collected within the analysis area surrounding the offshore wind turbine pile foundations.
[0017] Historical multi-source topographic observation data are mapped to a unified grid coordinate system to form a standardized seabed elevation raster dataset.
[0018] A time series matrix of seabed evolution was constructed based on a standardized seabed elevation raster dataset.
[0019] At each observation time, the spatial trajectories of the center lines of sand ridges and sand troughs are extracted from the standardized seabed elevation raster dataset.
[0020] Vectorization encoding was performed on the center lines of the sand ridges and sand troughs at each observation time to form spatiotemporal migration maps of the sand ridges and sand troughs.
[0021] By combining the seabed evolution time series matrix with the spatiotemporal migration maps of sand ridges and sand troughs, a topographic spatiotemporal evolution map is constructed.
[0022] Optionally, the process of acquiring the calibration tidal current and sediment dynamic field data includes:
[0023] A tidal and sediment dynamic field analysis region consistent with the analysis region surrounding the offshore wind turbine pile foundation is defined and divided into a spatial grid composed of two-dimensional spatial coordinate points;
[0024] At each two-dimensional spatial coordinate point of the spatial grid and at each observation time, the main direction of the tidal current, the instantaneous velocity field, and the sediment concentration field are collected.
[0025] The measured data set of profile flow velocity was collected at several profile observation positions and at each observation time in the spatial grid using an acoustic Doppler velocity profiler.
[0026] A coupled mathematical model of tidal current and sediment is constructed using instantaneous velocity field, sediment concentration field and tidal current direction as input parameters, and outputs the tidal current and sediment dynamic field at the corresponding spatial location and observation time.
[0027] Define a calibration error function for the tidal current-sediment coupling mathematical model to quantify the difference between simulated and measured values;
[0028] By adjusting the parameter set of the tidal current-sediment coupling mathematical model, the calibration error function of the tidal current-sediment coupling mathematical model is minimized, and the optimal model parameter configuration is obtained. The simulation results corresponding to the optimal model parameter configuration are called the scour driving variable prediction results after model calibration, including the calibration velocity field, the calibration sediment concentration field, and the calibration tidal current mainstream direction.
[0029] Based on the predicted results of scour driving variables, calibrated tidal and sediment dynamic field data are generated.
[0030] Optionally, the process of acquiring the digital elevation field of the seabed at the target prediction time includes:
[0031] The time series matrix of seabed evolution is normalized to obtain a normalized elevation tensor. The physical dimensions of the calibration velocity field, calibration sediment concentration field and calibration tidal current direction are made consistent to obtain a dimensionless dynamic tensor.
[0032] A learnable terrain disturbance mapper is constructed, which takes the normalized elevation tensor and the dimensionless dynamic tensor as joint inputs, and outputs the digital elevation field of the seabed at the target prediction time through a combination of nonlinear feature mapping and residual mapping.
[0033] Define an elevation prediction loss function, which measures the error between the target-time digital elevation field of the seabed output by the learnable terrain perturbation mapper and the actual observed elevation field.
[0034] The parameter set of the learnable terrain perturbation mapper is iteratively optimized by minimizing the elevation prediction loss function;
[0035] A learnable terrain disturbance mapper based on the optimal parameter set outputs a digital elevation field of the seabed at the target prediction time.
[0036] Optionally, the construction process of the physically driven attention-guided dual-channel diffusion network includes:
[0037] Using the seabed digital elevation field and calibrated tidal and sediment dynamic field data at the target prediction time as joint inputs, a physical-driven attention-guided dual-channel diffusion network including the first and second channels is constructed.
[0038] The first channel learns the formation trend of scour pits dominated by sediment diffusion and outputs the first channel prediction field;
[0039] The second channel learns the seabed evolution trend driven by sand ridge migration and outputs the second channel prediction field.
[0040] Optionally, the diffusion-anti-diffusion iterative process includes:
[0041] During the diffusion-anti-diffusion iteration process of the first and second channels, the calibration tidal and sediment dynamic field data is embedded into the network as a physical condition. At each spatial location and each diffusion step, the seabed digital elevation field at the target prediction time is jointly feature-processed with the physical condition.
[0042] A physical-driven attention mechanism is introduced into the physical-driven attention-guided dual-channel diffusion network. The physical-driven attention mechanism assigns attention weights based on the calibrated velocity field, calibrated sediment content field, and calibrated tidal current mainstream direction.
[0043] The first channel learns the influence of sediment diffusion on the formation of local scour pits through a physics-driven attention mechanism, and obtains a local scour rate prediction field.
[0044] The second channel learns the impact of sand ridge migration on the overall evolution of the seabed through a physical-driven attention mechanism, and obtains a prediction field of the overall evolution rate.
[0045] Optionally, the construction process of the coupling fusion layer includes:
[0046] The local scour rate prediction field and the global state evolution rate prediction field are fused with soft constraints based on local energy conservation constraints and mass conservation constraints.
[0047] Define a local energy conservation constraint function, which is used to quantify the deviation in local energy between the fused scour rate field output by the coupled fusion layer and the input local scour rate prediction field and global state evolution rate prediction field.
[0048] Define a mass conservation constraint function, which is used to quantify the deviation of the fused scour rate field output by the coupled fusion layer from satisfying the mass conservation of sediment in the spatial range;
[0049] A fusion constraint optimization objective function is constructed by weighted summation of the local energy conservation constraint function and the mass conservation constraint function;
[0050] By optimizing the objective function of the fusion constraint, the fusion scour rate field is iteratively adjusted to obtain the fusion scour rate prediction results data.
[0051] Based on the fusion of scour rate prediction data, the scour depth prediction data is calculated based on the seabed digital elevation field at the target prediction time.
[0052] Optionally, the output process of the pile foundation collaborative protection scheme includes:
[0053] Based on the updated scour rate prediction results and the updated scour depth prediction results, and combined with the spatial distribution and characteristic values of high scour rate areas, deep scour pit areas and areas with significant sand ridge migration, output rules are set for local toe protection parameters and compliant flow guide structure protection parameters, respectively.
[0054] A collaborative protection scheme for pile foundations was developed to achieve the design goal of collaborative protection for areas with high scour rates, deep scour pits, and areas with significant sand ridge migration.
[0055] Optionally, the output rules for the local foot protection parameters are as follows:
[0056] When there is a spatial location within the analysis area surrounding the pile foundation where the scour rate is greater than the scour rate threshold or the scour depth is greater than the scour depth threshold, the local toe protection structure type is output as reinforced toe protection, the toe protection radius is output as the maximum value of the distance from the spatial location to the center of the pile foundation, and the toe protection depth is output as the sum of the maximum scour depth of the spatial location and the safety redundancy depth.
[0057] When the scour rate of all spatial locations within the analysis area around the pile foundation is less than or equal to the scour rate threshold and the scour depth is less than or equal to the scour depth threshold, the local toe protection structure type is output as conventional toe protection, the toe protection radius is output as a specified multiple of the pile foundation diameter, and the toe protection depth is output as a conventional value.
[0058] The output rules for the protective parameters of the compliant flow guiding structure are as follows:
[0059] When the width of the significant migration zone of the sand ridge is greater than the migration width threshold or the angle between the main migration direction and the main tidal current direction is less than the angle threshold, the output of the guiding structure type is a reinforced guiding wall, the output of the guiding wall length is the maximum spatial span of the significant migration zone of the sand ridge, and the output of the guiding wall orientation is the angle between the main migration direction and the main tidal current direction.
[0060] When the width of the significant migration zone of the sand ridge is less than or equal to the migration width threshold and the angle between the main migration direction and the main tidal current direction is greater than or equal to the angle threshold, the output of the diversion structure type is an compliant diversion wall, the output of the diversion wall length is a specified multiple of the pile foundation diameter, and the output of the diversion wall orientation is the main tidal current direction.
[0061] The beneficial effects of this invention are:
[0062] (1) This invention improves the spatiotemporal resolution and physical interpretability of scour prediction by jointly modeling historical topographic disturbance sequences and physical driving conditions. It constructs a topographic spatiotemporal evolution map using multi-source topographic observation data under a unified coordinate system. It achieves high-resolution reproduction of historical sand ridge and sand trough spatiotemporal evolution sequences through a learnable topographic disturbance mapper. It also introduces numerically calibrated tidal sediment dynamic field data as model input conditions to achieve causal modeling of disturbance-response, which significantly improves the analytical ability for the complex interaction mechanism of sand ridge migration and local scour. This enables the model to distinguish and model scour trends in different regions and different historical stages, thereby improving the spatiotemporal resolution and physical consistency of scour rate and scour depth prediction.
[0063] (2) This invention achieves the separation and collaborative prediction of local scour and overall evolution mechanisms through a physical-driven attention-guided dual-channel diffusion network. The diffusion network sets up a channel for sediment diffusion to dominate the formation of scour pits and a channel for sand ridge migration to drive seabed state evolution. The physical-driven attention mechanism effectively achieves responsive identification of high scour risk areas and areas with significant sand ridge migration by weighting the calibration velocity field, sediment content field and tidal current mainstream direction. The dual-channel structure can learn for local scour rate and overall state evolution rate respectively. In the coupling layer, physical consistency fusion is achieved through soft constraints of energy conservation and mass conservation, which improves the model's prediction accuracy of scour rate and scour depth under different hydrodynamic fields and geomorphological evolution mechanisms. The prediction error of local extreme scour depth is reduced to less than 10 cm.
[0064] (3) This invention outputs an automated pile foundation collaborative protection scheme, which significantly improves the intelligence and adaptability of engineering protection design. Based on the updated scour rate prediction results and updated scour depth prediction results, the protection parameters of local toe protection and compliant flow guiding structures are automatically generated through spatial partition identification and parameter judgment. The collaborative protection scheme is automatically recommended and the parameters are quantified. It can automatically output the protection structure type, layout and key size parameters according to the spatiotemporal distribution of high scour rate area, deep scour pit area and sand ridge migration area, which improves the design efficiency and the pertinence and scientific nature of the protection scheme. It realizes the collaborative protection optimization design of local scour and overall scour of offshore wind power pile foundation, which significantly improves the protection effect and economic benefits. Attached Figure Description
[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0066] Figure 1 This is a flowchart of a machine learning-based quantitative prediction method for the scour rate of offshore wind power pile foundations proposed in this invention. Detailed Implementation
[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0068] refer to Figure 1 A method for quantitative prediction of scour rate of offshore wind turbine pile foundations based on machine learning, comprising:
[0069] Historical multi-source topographic observation data was collected, including historical nautical chart data, depth sounding data and lidar measurement data. Based on the historical multi-source topographic observation data, the morphological features of sand ridges and sand troughs in the preset analysis area around the offshore wind turbine pile foundation were extracted. The extracted topographic features were vectorized in a unified coordinate system to generate a topographic spatiotemporal evolution map.
[0070] Data on tidal current and sediment dynamics were collected, including the main tidal current direction, instantaneous velocity field, and sediment concentration field. The data were used as input and combined with the observation results of an acoustic Doppler current profiler to construct a coupled mathematical model of tidal current and sediment. The data on tidal current and sediment dynamics were then numerically calibrated to obtain calibrated tidal current and sediment dynamics data.
[0071] A learnable topographic disturbance mapper is constructed, and the spatiotemporal evolution map of topography and the calibrated tidal current and sediment dynamic field data are synchronously input into the learnable topographic disturbance mapper, which outputs the digital elevation field of the seabed at the target prediction time.
[0072] A physics-driven attention-guided dual-channel diffusion network is constructed using seabed digital elevation field and calibrated tidal and sediment dynamic field data at the target prediction time. The physics-driven attention-guided dual-channel diffusion network includes a first channel and a second channel. The first channel learns the trend of scour pit formation dominated by sediment diffusion; the second channel learns the trend of seabed morphological evolution driven by sand ridge migration.
[0073] In each diffusion-anti-diffusion iteration of the first and second channels, the tidal shear stress and turbulent dissipation intensity are calculated based on the calibrated tidal and sediment dynamic field data. A physical-driven attention mechanism is introduced to perform scale weight allocation on the seabed digital elevation field at the target prediction time, thereby obtaining the local scour rate prediction field and the global state evolution rate prediction field, respectively.
[0074] A coupling and fusion layer is constructed, and the local scour rate prediction field and the global state evolution rate prediction field are input. The two prediction fields are fused with soft constraints based on local energy conservation constraints and mass conservation constraints, and the scour rate prediction results and scour depth prediction results of the area surrounding the pile foundation are output.
[0075] The predicted scour rate data is compared with the field observation data of scour pit depth. Based on the comparison error, the parameters of the learnable terrain disturbance mapper and the parameters of the physical-driven attention-guided dual-channel diffusion network are jointly corrected to obtain the updated model and generate updated scour rate prediction data and updated scour depth prediction data.
[0076] The updated scour rate prediction data and the updated scour depth prediction data are used as inputs for the protection design to automatically generate a pile foundation collaborative protection scheme. The pile foundation collaborative protection scheme includes local toe protection parameters and compliant diversion structure protection parameters.
[0077] In this embodiment, the generation of the terrain spatiotemporal evolution map includes:
[0078] Define the analysis area around the offshore wind turbine pile foundation. The analysis area around the offshore wind turbine pile foundation is a circular area with the center of the pile foundation as the center and the radius as the analysis radius.
[0079] The analysis radius is the product of the pile diameter and the multiplication factor. The pile diameter is used to characterize the cross-sectional scale of a single pile, and the multiplication factor is used to reflect the expansion scale of the influence area of the pile on the analysis range.
[0080] Historical multi-source topographic observation data were collected within the analysis area surrounding offshore wind turbine pile foundations;
[0081] Historical multi-source topographic observation data includes historical nautical chart data, bathymetry data, and lidar measurement data. All of these data are elevation data indexed by two-dimensional spatial coordinates and multiple observation times. The two-dimensional spatial coordinates represent the spatial location of the topographic observation point on the horizontal plane, and the observation times represent the time points when the data was acquired.
[0082] Historical multi-source topographic observation data are mapped to a unified grid coordinate system to form a standardized seabed elevation raster dataset.
[0083] The standardized seabed elevation raster dataset is used to represent standardized elevation values from elevation data from different sources on the same spatial grid.
[0084] A time series matrix of seabed evolution was constructed based on a standardized seabed elevation raster dataset.
[0085] The seabed evolution time series matrix is used to express the dynamic changes in seabed elevation at multiple observation times and multiple spatial locations. Each item in the seabed evolution time series matrix represents the standardized elevation value at a specific spatial location at a specific observation time.
[0086] At each observation time, the spatial trajectories of the center lines of sand ridges and sand troughs are extracted from the standardized seabed elevation raster dataset.
[0087] The centerline of a sand ridge is used to express the spatial distribution characteristics of the sand ridge region at the same time, and the centerline of a sand trough is used to express the spatial distribution characteristics of the sand trough region at the same time. Both the centerline of a sand ridge and the centerline of a sand trough are represented by a series of spatial coordinate points. The number of points in the spatial coordinate point sequence corresponds to the number of sand ridge points and the number of sand trough points, respectively.
[0088] Vectorization encoding was performed on the center lines of the sand ridges and sand troughs at each observation time to form spatiotemporal migration maps of the sand ridges and sand troughs.
[0089] Vectorization encoding refers to mapping the spatial coordinates of the centerlines of sand ridges and sand troughs at various times to a unified coordinate system. The spatiotemporal migration maps of sand ridges and sand troughs are used to describe the geometric migration process of sand ridges and sand troughs within the analysis area.
[0090] By combining the seabed evolution time series matrix with the spatiotemporal migration maps of sand ridges and sand troughs, a topographic spatiotemporal evolution map is constructed. This map is used to comprehensively express and analyze the changes in seabed elevation over time and the geometric migration processes of sand ridges and sand troughs within the region.
[0091] This implementation improves the spatiotemporal resolution and physical interpretability of scour prediction by jointly modeling historical topographic disturbance sequences and physical driving conditions. It constructs a topographic spatiotemporal evolution map using multi-source topographic observation data in a unified coordinate system, and achieves high-resolution reproduction of historical sand ridge and sand trough spatiotemporal evolution sequences through a learnable topographic disturbance mapper. Furthermore, it introduces numerically calibrated tidal and sediment dynamic field data as model input conditions to achieve causal modeling of disturbance-response, significantly improving the analytical ability for the complex interaction mechanism of sand ridge migration and local scour. This enables the model to differentiate and model scour trends in different regions and at different historical stages, thereby improving the spatiotemporal resolution and physical consistency of scour rate and scour depth predictions.
[0092] In this embodiment, the process of acquiring calibration tidal and sediment dynamic field data includes:
[0093] A tidal and sediment dynamic field analysis region consistent with the analysis region surrounding the offshore wind turbine pile foundation is defined and divided into a spatial grid composed of two-dimensional spatial coordinate points;
[0094] Two-dimensional spatial coordinates are used to uniquely identify each spatial location within the analysis area of the tidal and sediment dynamic field.
[0095] At each two-dimensional spatial coordinate point of the spatial grid and at each observation time, the main direction of the tidal current, the instantaneous velocity field, and the sediment concentration field are collected.
[0096] The current direction is used to characterize the direction of hydrodynamics, the instantaneous velocity field is used to characterize the speed of water movement per unit time, and the sediment concentration field is used to characterize the concentration of suspended sediment in a unit volume of water.
[0097] The measured velocity data of the profile was collected at several profile observation locations and at each observation time using an acoustic Doppler velocity profiler on a spatial grid. Where m and n represent the coordinate indices of the profile location within the spatial grid;
[0098] A coupled mathematical model of tidal current and sediment is constructed using instantaneous velocity field, sediment concentration field and tidal current direction as input parameters, and outputs the tidal current and sediment dynamic field at the corresponding spatial location and observation time.
[0099] The instantaneous velocity field, sediment load field, and tidal current direction collected at each two-dimensional spatial coordinate point and at each observation time in the spatial grid are input as parameters into the tidal current-sediment coupling mathematical model. The tidal current-sediment coupling mathematical model performs coupled numerical calculations of the dynamic equation and the sediment transport equation on the input parameters. The dynamic equation is used to simulate the movement process of water flow, and the sediment transport equation is used to simulate the migration and deposition process of sediment with the movement of water. Based on the coupling results of the dynamic equation and the sediment transport equation, the tidal current-sediment coupling mathematical model outputs a simulated scour rate field at each spatial grid point and at each observation time. The simulated scour rate field is used to describe the rate of change of seabed surface elevation per unit time at each spatial location and at each observation time. Positive values of the rate of change of seabed surface elevation indicate sedimentation, and negative values indicate scour.
[0100] Define the calibration error function ε of the tidal current-sediment coupling mathematical model. cal Used to quantify the difference between simulated and measured values:
[0101]
[0102] Where M represents the number of discrete points along the x-direction in the spatial grid, i.e., the total number of nodes in the spatial grid along the x-direction; N represents the number of discrete points along the y-direction in the spatial grid, i.e., the total number of nodes in the spatial grid along the y-direction; T represents the total number of observation times within the analysis period of the tidal and sediment dynamic field; M represents the index along the x-direction in the spatial grid; and t... k This represents the k-th observation time. This indicates the spatial location (x) of the tidal current-sediment coupling mathematical model. m ,y n ) and observation time t k The flow velocity results obtained from the simulation;
[0103] By adjusting the parameter set of the tidal current-sediment coupling mathematical model, the calibration error function of the tidal current-sediment coupling mathematical model is minimized, and the optimal model parameter configuration is obtained. The simulation results corresponding to the optimal model parameter configuration are called the scour driving variable prediction results after model calibration, including the calibration velocity field, the calibration sediment concentration field, and the calibration tidal current mainstream direction.
[0104] Based on the calibrated velocity field, calibrated sediment concentration field, and calibrated tidal current mainstream direction after model calibration, calibrated tidal current sediment dynamic field data are generated.
[0105] In this embodiment, the process of obtaining the digital elevation field of the seabed at the target prediction time includes:
[0106] The time series matrix of seabed evolution is normalized to obtain a normalized elevation tensor. The physical dimensions of the calibration velocity field, calibration sediment concentration field and calibration tidal current direction are made consistent to obtain a dimensionless dynamic tensor.
[0107] The normalized elevation tensor is used to characterize the standardized elevation values at different spatial locations at various observation times, while the dimensionless dynamic tensor is used to characterize the normalization characteristics of hydrodynamic parameters and sediment parameters in space and time.
[0108] A learnable terrain disturbance mapper is constructed, which takes the normalized elevation tensor and the dimensionless dynamic tensor as joint inputs, and outputs the digital elevation field of the seabed at the target prediction time through a combination of nonlinear feature mapping and residual mapping.
[0109] The digital elevation field of the seabed at the target prediction time is used to characterize the seabed elevation values at each spatial location at the target prediction time, and the spatial location is uniquely determined by the spatial grid coordinates.
[0110] Define an elevation prediction loss function, which measures the error between the target-time digital elevation field of the seabed output by the learnable terrain perturbation mapper and the actual observed elevation field.
[0111] By calculating the sum of squares of the differences between the predicted and observed elevation values at all spatial grid coordinate points, and then taking the average of the error values for all spatial grid points, the error between the digital elevation field of the seabed at the target prediction time and the actual observed elevation field is obtained.
[0112] The parameter set of the learnable terrain perturbation mapper is iteratively optimized by minimizing the elevation prediction loss function;
[0113] Iterative optimization is used to continuously update the parameter set so that the elevation prediction loss function reaches its minimum value, thereby obtaining the optimal parameter set, which is the final parameter configuration of the learnable terrain disturbance mapper.
[0114] A learnable terrain disturbance mapper based on the optimal parameter set outputs a digital elevation field of the seabed at the target prediction time.
[0115] In this embodiment, the physical-driven attention-guided construction process of the dual-channel diffusion network includes:
[0116] Using the seabed digital elevation field and calibrated tidal and sediment dynamic field data at the target prediction time as joint inputs, a physical-driven attention-guided dual-channel diffusion network including the first and second channels is constructed.
[0117] The first channel takes the seabed digital elevation field and calibrated tidal and sediment dynamic field data at the target prediction time as input, learns the trend of scour pit formation dominated by sediment diffusion, and outputs the first channel prediction field.
[0118] The second channel takes the seabed digital elevation field and calibrated tidal and sediment dynamic field data at the target prediction time as input, learns the seabed state evolution trend driven by sand ridge migration, and outputs the second channel prediction field.
[0119] In this embodiment, the diffusion-anti-diffusion iterative process includes:
[0120] During the diffusion-reverse diffusion iteration process of the first and second channels, the calibration velocity field, calibration sediment concentration field, and calibration tidal current mainstream direction in the calibration tidal current and sediment dynamic field data are embedded into the network as physical conditions. At each spatial location and each diffusion step, the seabed digital elevation field at the target prediction time is jointly feature-processed with the physical conditions.
[0121] A physical-driven attention mechanism is introduced into the physical-driven attention-guided dual-channel diffusion network. The physical-driven attention mechanism assigns attention weights based on the calibrated velocity field, calibrated sediment content field, and calibrated tidal current mainstream direction.
[0122] Attention weights are used to control the model’s sensitivity to scour or sedimentation responses at various spatial locations.
[0123] The first channel learns the influence of sediment diffusion on the formation of local scour pits through a physics-driven attention mechanism, and obtains a local scour rate prediction field.
[0124] The local scour rate prediction field is used to characterize the distribution of local scour rates at each spatial location at the target prediction time.
[0125] The second channel learns the impact of sand ridge migration on the overall evolution of the seabed through a physical-driven attention mechanism, and obtains a prediction field of the overall evolution rate.
[0126] The global state evolution rate prediction field is used to characterize the distribution of the global state evolution rate of the seabed at each spatial location at the target prediction time.
[0127] This implementation achieves the separation, characterization, and collaborative prediction of local scour and overall evolution mechanisms through a physically driven attention-guided dual-channel diffusion network. The diffusion network incorporates a sediment diffusion-driven scour pit formation channel and a ridge migration-driven seabed evolution channel. By using a physically driven attention mechanism to weight the calibrated velocity field, sediment content field, and mainstream tidal current direction, it effectively identifies high-scour-risk areas and areas with significant ridge migration. The dual-channel structure can learn for both local scour rates and overall evolution rates separately. Within the coupling layer, energy and mass conservation soft constraints achieve physical consistency fusion, improving the model's prediction accuracy for scour rates and depths under different hydrodynamic fields and geomorphological evolution mechanisms. The prediction error for local extreme scour depths is reduced to within 10 cm.
[0128] In this embodiment, the construction process of the coupling fusion layer includes:
[0129] The local scour rate prediction field and the global state evolution rate prediction field are fused with soft constraints based on local energy conservation constraints and mass conservation constraints.
[0130] Local energy conservation constraints are used to ensure that the predicted scour rate output by the coupled fusion layer satisfies the energy conservation characteristics of the scour and sedimentation process within the local area, while mass conservation constraints are used to ensure that the predicted scour rate output by the coupled fusion layer satisfies the mass conservation characteristics of the sediment transport process within the analysis area.
[0131] Define a local energy conservation constraint function, which is used to quantify the deviation in local energy between the fused scour rate field output by the coupled fusion layer and the input local scour rate prediction field and global state evolution rate prediction field.
[0132] The local energy conservation constraint function is calculated by averaging the squared difference between the value at the corresponding position of the fused scour rate field and the values of the input local scour rate prediction field and the global state evolution rate prediction field at each spatial grid coordinate point.
[0133] Define a mass conservation constraint function, which is used to quantify the deviation of the fused scour rate field output by the coupled fusion layer from satisfying the mass conservation of sediment in the spatial range;
[0134] The mass conservation constraint function is calculated by calculating the net sediment transport mass value corresponding to the fused scour rate field at each spatial grid coordinate point, and summing the net sediment transport mass values at all spatial locations within the analysis area to obtain the overall net sediment transport mass value.
[0135] A fusion constraint optimization objective function is constructed by weighted summation of the local energy conservation constraint function and the mass conservation constraint function;
[0136] By optimizing the objective function of the fusion constraint, the fusion scour rate field is iteratively adjusted to obtain the fusion scour rate prediction results data.
[0137] The fusion of scour rate prediction results data is used to characterize the scour rate distribution at various spatial locations around the pile foundation at the target prediction time.
[0138] Based on the fused scour rate prediction data, the scour depth prediction data is calculated on the basis of the seabed digital elevation field at the target prediction time.
[0139] The scour depth prediction results are used to characterize the predicted depth distribution of seabed scour pits at various spatial locations around the pile foundation at the target prediction time.
[0140] In this embodiment, the output process of the pile foundation collaborative protection scheme includes:
[0141] Based on the updated scour rate prediction results and the updated scour depth prediction results, and combined with the spatial distribution and characteristic values of high scour rate areas, deep scour pit areas and areas with significant sand ridge migration, output rules are set for local toe protection parameters and compliant flow guide structure protection parameters, respectively.
[0142] By integrating the local toe protection parameters and the compliant flow guiding structure protection parameters, a collaborative protection scheme for pile foundations is formed, achieving the design goal of collaborative protection for high scour rate areas, deep scour pit areas, and areas with significant sand ridge migration.
[0143] In this embodiment, the output rules for local foot protection parameters are as follows:
[0144] When there is a spatial location within the analysis area surrounding the pile foundation where the scour rate is greater than the scour rate threshold or the scour depth is greater than the scour depth threshold, the local toe protection structure type is output as reinforced toe protection, the toe protection radius is output as the maximum value of the distance from the spatial location to the center of the pile foundation, and the toe protection depth is output as the sum of the maximum scour depth of the spatial location and the safety redundancy depth.
[0145] When the scour rate of all spatial locations within the analysis area around the pile foundation is less than or equal to the scour rate threshold and the scour depth is less than or equal to the scour depth threshold, the local toe protection structure type is output as conventional toe protection, the toe protection radius is output as a specified multiple of the pile foundation diameter, and the toe protection depth is output as a conventional value.
[0146] Output rules for protection parameters of compliant flow guiding structure:
[0147] When the width of the significant migration zone of the sand ridge is greater than the migration width threshold or the angle between the main migration direction and the main tidal current direction is less than the angle threshold, the output of the guiding structure type is a reinforced guiding wall, the output of the guiding wall length is the maximum spatial span of the significant migration zone of the sand ridge, and the output of the guiding wall orientation is the angle between the main migration direction and the main tidal current direction.
[0148] When the width of the significant migration zone of the sand ridge is less than or equal to the migration width threshold and the angle between the main migration direction and the main tidal current direction is greater than or equal to the angle threshold, the output of the diversion structure type is an compliant diversion wall, the output of the diversion wall length is a specified multiple of the pile foundation diameter, and the output of the diversion wall orientation is the main tidal current direction.
[0149] This embodiment outputs an automated pile foundation collaborative protection scheme, which significantly improves the intelligence and adaptability of engineering protection design. Based on the updated scour rate prediction results and updated scour depth prediction results, it automatically generates protection parameters for local toe protection and compliant diversion structures through spatial partition identification and parameter criteria, realizing the automated recommendation and parameter quantification of collaborative protection schemes. It can automatically output the type, layout and key dimension parameters of protection structures according to the spatiotemporal distribution of high scour rate areas, deep scour pit areas and areas with significant sand ridge migration, thereby improving design efficiency and the pertinence and scientific nature of protection schemes.
[0150] In this embodiment, based on the updated scour rate prediction results and the updated scour depth prediction results, numerical simulation calculations are performed to optimize the protection parameters in the pile foundation collaborative protection scheme. The optimized pile foundation collaborative protection scheme is then verified by physical model tests to form a final protection configuration recommendation.
[0151] By combining numerical simulation and physical experiments, a coordinated protection optimization design for local and overall scour of offshore wind turbine pile foundations was achieved, which significantly improved the protection effect and economic benefits.
[0152] Example 1: During the pile foundation construction and operation and maintenance of an offshore wind farm, the project team used this invention to carry out quantitative prediction of scour rate and intelligent protection design.
[0153] In this embodiment, after the pile foundation installation was completed, the on-site technicians deployed multiple sets of high-precision water depth sensors and lidar equipment to intensively sample the seabed elevation within different radii around the pile foundation. The collected data included multi-source seabed elevation information. In Example 1, the elevation at grid point A was 4.37 meters, at grid point B it was 4.41 meters, and at grid point C it was 4.52 meters. Over the following two-month period, continuous observation data showed that the elevation at grid point D gradually decreased from 4.35 meters to 4.01 meters, while the elevation at grid point E remained relatively stable, consistently between 4.60 and 4.62 meters.
[0154] Through data preprocessing in a unified coordinate system, the system automatically constructed a seabed elevation grid containing 48,000 sets of spatial-temporal data. Further, using spatial analysis algorithms, the system extracted the ridge centerline trajectory within the R1 to R3 grid interval, finding that the ridge migrated from point A to point F, with a horizontal distance change of 8.2 meters. At the same time, the trough centerline also showed a southward trend in the P1-P4 region, with a cumulative displacement of 5.6 meters.
[0155] To accurately assess the impact of sediment transport, the project team deployed three hydrodynamic profilers at different depths to continuously collect data on local flow velocity, mainstream direction, and sediment concentration. The sampling data showed that the maximum flow velocity at point G was 1.44 m / s, corresponding to a suspended sediment concentration of 1.12 kg / m³. 3 The flow direction is 12° southeast of the main axis, with a peak flow velocity of 1.27 m / s at point H and a sediment concentration of 1.05 kg / m³. 3 The flow direction is 19° south.
[0156] All historical elevation data and dynamic field data were synchronously input into the terrain disturbance mapper. After normalization and convolutional residual calculation, the system predicted that the elevation of the N-point grid (13 meters from the pile center) would drop from 4.13 meters to 3.85 meters within the next month. Compared with the real-time elevation sensor data on site, the actual elevation change of the point was 4.14 meters to 3.81 meters, with an absolute error of 0.04 meters. Compared with the prediction value of the traditional empirical formula (dropped to 3.97 meters, with an error of 0.16 meters) and the prediction value of the conventional neural network (dropped to 3.72 meters, with an error of 0.09 meters), the present invention shows better accuracy in high-risk areas.
[0157] In the dual-channel diffusion network stage, the system automatically identifies a local scour rate extremum in the M region, with a local scour rate of -0.033 m / d and an overall evolution rate of -0.008 m / d. Comparing this with actual observation data, the measured scour rate at point M is -0.031 m / d, a difference of 0.002 m / d, which is better than the traditional model (-0.024 m / d, error 0.007 m / d). Statistical analysis of 20 sets of data for all scour rate extremum points shows that the average absolute error of this invention is 0.0028 m / d, while that of the traditional method is 0.0096 m / d.
[0158] The system inputs all local scour rate prediction fields and global state evolution rate prediction fields into a coupled fusion layer. After optimization with energy and mass conservation constraints, the final output is the scour rate prediction surface within the Q1-Q5 interval. Data analysis shows that the maximum scour depth in the Q3 interval is predicted to be 4.08 meters, while the actual observation is 4.11 meters; the predicted depth in the Q4 interval is 3.62 meters, while the actual depth is 3.67 meters, with the error controlled within 0.05 meters. The maximum error using the traditional model is 0.18 meters.
[0159] During the protection design output phase, the system automatically determined that the scour rate in section S1 (12.5 meters from the center of the pile foundation) exceeded -0.03 m / d and the scour depth was greater than 4.00 meters. Based on the set rules, a locally reinforced toe protection was recommended, with a toe protection radius of 11.2 meters and a burial depth of 4.8 meters. The area of significant sand ridge migration was identified as the main control direction in the southeast. The system automatically recommended a conforming guide wall with a length of 29 meters and an azimuth adjusted to an angle of 12° with the main flow direction. After the actual engineers adopted the solution, scour observation data for two consecutive months showed that the maximum scour rate in section S1 decreased by 16%, the scour depth remained stable below 4.15 meters, and the maintenance of the protection structure was reduced by 11% compared to the past.
[0160] In the sample training and comparison phase, the system used 36,000 sets of historical spatiotemporal elevation-dynamic data to train and validate the model. The test set contained 4,000 sets of spatial-temporal points, covering all typical working conditions. The traditional RANS model could only perform regression fitting based on 230 sets of historical scour pit data and simplified dynamic field data. The single-channel convolutional neural network used 10,200 sets of training samples, but failed to combine sand ridge migration and dynamic spatiotemporal coupling information.
[0161] The comparative results further show that the average absolute error between the maximum scour depth predicted by the present invention and the actual observation data is 0.05 meters and the maximum error is 0.12 meters within the entire pile foundation group. The errors of the traditional RANS model are 0.17 meters and 0.31 meters, respectively, and the errors of the single-channel network are 0.10 meters and 0.21 meters. In the high scour rate region, the error of the present invention is always less than 0.003 m / d, while the maximum error of the traditional model is as high as 0.015 m / d.
[0162] The method of this invention has significantly higher predictive reliability than traditional empirical methods under extreme tidal and strong wind and wave conditions, effectively improving the safety level and operation and maintenance efficiency of wind power pile foundations. Through the automatic output of intelligent protection schemes, the project cycle is shortened by 13%, and the material usage optimization rate reaches 8.5%.
[0163] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for quantitatively predicting the scour rate of offshore wind turbine foundation piles based on machine learning, characterized in that, include: Historical multi-source topographic observation data were collected, and the morphological features of sand ridges and sand troughs in the pre-set analysis area around the offshore wind power pile foundation were extracted. The data were then vectorized in a unified coordinate system to generate a topographic spatiotemporal evolution map. Data on tidal current and sediment dynamic fields were collected and combined with observations from an acoustic Doppler current profiler to construct a coupled mathematical model of tidal current and sediment, thereby obtaining calibrated tidal current and sediment dynamic field data. Construct a learnable terrain perturbation mapper to output a digital seabed elevation field at the target prediction time; A physical-driven attention-guided dual-channel diffusion network, including a first channel and a second channel, is constructed using the seabed digital elevation field at the target prediction time and calibrated tidal and sediment dynamic field data. In each diffusion-anti-diffusion iteration of the first and second channels, a physical-driven attention mechanism is introduced to assign scale weights to the digital elevation field of the seabed at the target prediction time, thereby obtaining the local scour rate prediction field and the global state evolution rate prediction field, respectively. Construct a coupling and fusion layer to output the predicted scour rate and scour depth data of the area surrounding the pile foundation. The predicted scour rate data is compared with the field observation data of scour pit depth, and the data are jointly corrected to obtain an updated model, generating updated scour rate prediction data and updated scour depth prediction data. Based on the updated scour rate prediction data and the updated scour depth prediction data, a collaborative protection scheme for pile foundations is automatically generated.
2. The method for quantitative prediction of scour rate of offshore wind turbine pile foundation based on machine learning according to claim 1, characterized in that, The generation of the spatiotemporal evolution map of the terrain includes: Define the analysis area around the offshore wind turbine pile foundation. The analysis area around the offshore wind turbine pile foundation is a circular area with the center of the pile foundation as the center and the radius as the analysis radius. Historical multi-source topographic observation data were collected within the analysis area surrounding the offshore wind turbine pile foundations. Historical multi-source topographic observation data are mapped to a unified grid coordinate system to form a standardized seabed elevation raster dataset. A time series matrix of seabed evolution was constructed based on a standardized seabed elevation raster dataset. At each observation time, the spatial trajectories of the center lines of sand ridges and sand troughs are extracted from the standardized seabed elevation raster dataset. Vectorization encoding was performed on the center lines of the sand ridges and sand troughs at each observation time to form spatiotemporal migration maps of the sand ridges and sand troughs. By combining the seabed evolution time series matrix with the spatiotemporal migration maps of sand ridges and sand troughs, a topographic spatiotemporal evolution map is constructed.
3. The method for quantitative prediction of scour rate of offshore wind turbine pile foundations based on machine learning according to claim 1, characterized in that, The process of acquiring the calibration tidal and sediment dynamic field data includes: A tidal and sediment dynamic field analysis region consistent with the analysis region surrounding the offshore wind turbine pile foundation is defined and divided into a spatial grid composed of two-dimensional spatial coordinate points; At each two-dimensional spatial coordinate point of the spatial grid and at each observation time, the main direction of the tidal current, the instantaneous velocity field, and the sediment concentration field are collected. The measured data set of profile flow velocity was collected at several profile observation positions and at each observation time in the spatial grid using an acoustic Doppler velocity profiler. A coupled mathematical model of tidal current and sediment is constructed using instantaneous velocity field, sediment concentration field and tidal current direction as input parameters, and outputs the tidal current and sediment dynamic field at the corresponding spatial location and observation time. Define a calibration error function for the tidal current-sediment coupling mathematical model to quantify the difference between simulated and measured values; By adjusting the parameter set of the tidal current-sediment coupling mathematical model, the calibration error function of the tidal current-sediment coupling mathematical model is minimized, and the optimal model parameter configuration is obtained. The simulation results corresponding to the optimal model parameter configuration are called the scour driving variable prediction results after model calibration, including the calibration velocity field, the calibration sediment concentration field, and the calibration tidal current mainstream direction. Based on the predicted results of scour driving variables, calibrated tidal and sediment dynamic field data are generated.
4. The method for quantitative prediction of scour rate of offshore wind turbine foundation based on machine learning according to claim 1, characterized in that, The process of acquiring the digital elevation field of the seabed at the target prediction time includes: The time series matrix of seabed evolution is normalized to obtain a normalized elevation tensor. The physical dimensions of the calibration velocity field, calibration sediment concentration field and calibration tidal current direction are made consistent to obtain a dimensionless dynamic tensor. A learnable terrain disturbance mapper is constructed, which takes the normalized elevation tensor and the dimensionless dynamic tensor as joint inputs, and outputs the digital elevation field of the seabed at the target prediction time through a combination of nonlinear feature mapping and residual mapping. Define an elevation prediction loss function, which measures the error between the target-time digital elevation field of the seabed output by the learnable terrain perturbation mapper and the actual observed elevation field. The parameter set of the learnable terrain perturbation mapper is iteratively optimized by minimizing the elevation prediction loss function; A learnable terrain disturbance mapper based on the optimal parameter set outputs a digital elevation field of the seabed at the target prediction time.
5. The method for quantitative prediction of scour rate of offshore wind turbine foundation based on machine learning according to claim 1, characterized in that, The construction process of the physical-driven attention-guided dual-channel diffusion network includes: Using the seabed digital elevation field and calibrated tidal and sediment dynamic field data at the target prediction time as joint inputs, a physical-driven attention-guided dual-channel diffusion network including the first and second channels is constructed. The first channel learns the formation trend of scour pits dominated by sediment diffusion and outputs the first channel prediction field; The second channel learns the seabed evolution trend driven by sand ridge migration and outputs the second channel prediction field.
6. The method for quantitative prediction of scour rate of offshore wind turbine foundation based on machine learning according to claim 1, characterized in that, The diffusion-anti-diffusion iterative process includes: During the diffusion-anti-diffusion iteration process of the first and second channels, the calibration tidal and sediment dynamic field data is embedded into the network as a physical condition. At each spatial location and each diffusion step, the seabed digital elevation field at the target prediction time is jointly feature-processed with the physical condition. A physical-driven attention mechanism is introduced into the physical-driven attention-guided dual-channel diffusion network. The physical-driven attention mechanism assigns attention weights based on the calibrated velocity field, calibrated sediment content field, and calibrated tidal current mainstream direction. The first channel learns the influence of sediment diffusion on the formation of local scour pits through a physics-driven attention mechanism, and obtains a local scour rate prediction field. The second channel learns the impact of sand ridge migration on the overall evolution of the seabed through a physical-driven attention mechanism, and obtains a prediction field of the overall evolution rate.
7. The method for quantitative prediction of scour rate of offshore wind turbine pile foundation based on machine learning according to claim 1, characterized in that, The construction process of the coupling fusion layer includes: The local scour rate prediction field and the global state evolution rate prediction field are fused with soft constraints based on local energy conservation constraints and mass conservation constraints. Define a local energy conservation constraint function, which is used to quantify the deviation in local energy between the fused scour rate field output by the coupled fusion layer and the input local scour rate prediction field and global state evolution rate prediction field. Define a mass conservation constraint function, which is used to quantify the deviation of the fused scour rate field output by the coupled fusion layer from satisfying the mass conservation of sediment in the spatial range; A fusion constraint optimization objective function is constructed by weighted summation of the local energy conservation constraint function and the mass conservation constraint function; By optimizing the objective function of the fusion constraint, the fusion scour rate field is iteratively adjusted to obtain the fusion scour rate prediction results data. Based on the fusion of scour rate prediction data, the scour depth prediction data is calculated based on the seabed digital elevation field at the target prediction time.
8. The method for quantitative prediction of scour rate of offshore wind turbine foundation based on machine learning according to claim 1, characterized in that, The output process of the pile foundation collaborative protection scheme includes: Based on the updated scour rate prediction results and the updated scour depth prediction results, and combined with the spatial distribution and characteristic values of high scour rate areas, deep scour pit areas and areas with significant sand ridge migration, output rules are set for local toe protection parameters and compliant flow guide structure protection parameters, respectively. A collaborative protection scheme for pile foundations was developed to achieve the design goal of collaborative protection for areas with high scour rates, deep scour pits, and areas with significant sand ridge migration.
9. A method for quantitatively predicting the scour rate of offshore wind turbine foundations based on machine learning, as described in claim 8, is characterized in that... The output rules for the local foot protection parameters are as follows: When there is a spatial location within the analysis area surrounding the pile foundation where the scour rate is greater than the scour rate threshold or the scour depth is greater than the scour depth threshold, the local toe protection structure type is output as reinforced toe protection, the toe protection radius is output as the maximum value of the distance from the spatial location to the center of the pile foundation, and the toe protection depth is output as the sum of the maximum scour depth of the spatial location and the safety redundancy depth. When the scour rate of all spatial locations within the analysis area around the pile foundation is less than or equal to the scour rate threshold and the scour depth is less than or equal to the scour depth threshold, the local toe protection structure type is output as conventional toe protection, the toe protection radius is output as a specified multiple of the pile foundation diameter, and the toe protection depth is output as a conventional value. The output rules for the protective parameters of the compliant flow guiding structure are as follows: When the width of the significant migration zone of the sand ridge is greater than the migration width threshold or the angle between the main migration direction and the main tidal current direction is less than the angle threshold, the output of the guiding structure type is a reinforced guiding wall, the output of the guiding wall length is the maximum spatial span of the significant migration zone of the sand ridge, and the output of the guiding wall orientation is the angle between the main migration direction and the main tidal current direction. When the width of the significant migration zone of the sand ridge is less than or equal to the migration width threshold and the angle between the main migration direction and the main tidal current direction is greater than or equal to the angle threshold, the output of the diversion structure type is an compliant diversion wall, the output of the diversion wall length is a specified multiple of the pile foundation diameter, and the output of the diversion wall orientation is the main tidal current direction.