A power flow speed prediction method based on data enhancement and space-time neural network
By using data augmentation and spatiotemporal neural network methods, a multidimensional spatiotemporal tensor was constructed and physical constraints were introduced to solve the problems of time dependence, spatial correlation and physical consistency in tidal current velocity prediction in tidal flat areas, thus achieving higher accuracy and more stable tidal current velocity prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI NORMAL UNIVERSITY
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-21
AI Technical Summary
Existing tidal current velocity prediction methods struggle to simultaneously consider time dependence, spatial neighborhood correlation, and physical consistency in tidal flat areas, resulting in insufficient prediction accuracy and stability, especially in terms of poor generalization ability under complex environments.
We employ data augmentation and spatiotemporal neural network methods to construct a multidimensional spatiotemporal tensor, combine spatial feature encoding, feature dimension attention, and spatiotemporal coupling loop, and introduce a physical operator constraint loss term to predict power flow velocity.
It improves the accuracy, stability, and generalization ability of tidal current velocity prediction, and can better characterize the tidal current evolution process in tidal flat areas, meeting the requirements of physical consistency.
Smart Images
Figure CN122432550A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of nearshore hydrodynamic parameter monitoring and inversion technology, specifically to a tidal current velocity prediction method based on data augmentation and spatiotemporal neural networks. Background Technology
[0002] Tidal current velocity is a crucial parameter characterizing the hydrodynamic processes of tidal flats. Its variations directly influence the initiation, transport, sedimentation, and scour evolution of sediment on the seabed, further affecting the stability of tidal flat landforms, the effectiveness of vegetation establishment, and the rationality of the layout of ecological restoration engineering components. In tidal flat ecological restoration scenarios, failure to accurately grasp the tidal current velocity variation patterns of the target area at different time scales can easily lead to inaccurate selection of vegetation transplantation areas, biased assessment of local scour and sedimentation risks, and insufficient design basis for restoration facilities such as wave-dissipating dams and diversion dikes, thereby affecting the ecological restoration effect and subsequent operational stability.
[0003] Currently, common methods for predicting tidal current velocity mainly include hydrodynamic numerical modeling, harmonic analysis prediction, and basic statistical or machine learning prediction. Hydrodynamic numerical modeling typically involves establishing a two-dimensional or three-dimensional hydrodynamic model and solving the governing equations under discrete grid conditions to obtain the tidal current evolution results for the target area. While this type of method has strong mechanistic interpretation capabilities, its prediction accuracy is highly dependent on fundamental data such as water depth, topography, seabed parameters, and boundary conditions. The model construction process is complex and computationally resource-intensive. In tidal flat areas where there are frequent changes in local micro-topography, significant differences in seabed sediment, and highly unsteady flow processes, it often suffers from long modeling cycles, difficulties in parameter adjustment, and local response lags.
[0004] Harmonic analysis forecasting methods are primarily based on the periodic characteristics of tidal motion. They decompose historical tidal current sequences into tidal constituents and reconstruct future current velocities using the amplitude and lag angle of each tidal constituent. This type of method is applicable to sea areas with strong periodicity and relatively stable boundary conditions. However, it is essentially a linear extrapolation method based on periodic laws and lacks sufficient responsiveness to non-periodic factors such as wind disturbances, local topographic changes, differences in seabed friction, and non-astronomical tides. Particularly in shallow tidal flats, river-sea confluence zones, or areas significantly affected by meteorological disturbances, it is prone to significant instantaneous prediction biases.
[0005] Basic statistical and machine learning prediction methods typically employ autoregressive models, support vector machines, or other sequential regression models to fit historical tidal current velocity data, establishing a mapping relationship between input and output. While these methods are relatively simple in implementation, most focus on single-point time series, emphasizing the discovery of correlations over time, and failing to adequately consider the spatial heterogeneity prevalent in tidal flat areas, the influence of surrounding topography, and the coupling relationships between multiple environmental factors. Furthermore, most existing methods lack the incorporation of physical constraints on fluid motion, making it difficult to integrate the intrinsic correlation between wind speed, elevation, seabed conditions, and tidal current velocity into the modeling process. Therefore, in applications with significant data fluctuations and complex environmental changes, they are prone to insufficient generalization ability.
[0006] With the development of deep learning methods, using neural networks to predict tidal current velocity has become a feasible approach. However, existing solutions still have two shortcomings in tidal flat scenarios: First, measured tidal current velocity data usually have limited sampling points and are easily affected by sensor obstruction, communication failures, and environmental noise, making it difficult to fully characterize the tidal current evolution process by relying solely on raw sequence data; second, tidal current velocity is affected by multiple factors such as tides, topography, bottom sediment type, and wind field, exhibiting significant spatiotemporal coupling characteristics. If the model only focuses on a single time series or static environmental factor, it is difficult to simultaneously take into account temporal dependence, spatial neighborhood correlation, and physical consistency.
[0007] In view of this, the present invention proposes a power flow velocity prediction method based on data augmentation and spatiotemporal neural networks. Summary of the Invention
[0008] The purpose of this invention is to provide a tidal current velocity prediction method based on data augmentation and spatiotemporal neural networks. This method aims to solve the technical problems of existing bed shear stress monitoring methods, which are unable to achieve the coordinated extraction of water flow velocity, wave trajectory velocity and wave height based on a single vertical acceleration signal under the combined action of tidal flats and waves and the alternating open and submerged conditions. This results in high cost, poor continuity and insufficient reliability in the estimation of bed shear stress.
[0009] In a first aspect, the present invention provides a method for predicting power flow velocity based on data augmentation and spatiotemporal neural networks, comprising the following steps:
[0010] Acquire tidal current velocity sequence data, elevation data, bed sediment type data, forecast wind speed data, and tidal timetable data for the target tidal flat area; perform outlier detection and missing value processing on the tidal current velocity sequence data;
[0011] Vertical acceleration and turbulence intensity features are constructed based on the processed tidal velocity sequence data; slope, aspect, and surface roughness features are constructed based on elevation data; and wind stress, friction velocity, and wind energy density features are constructed based on forecast wind speed data.
[0012] A local spatial analysis window is established with the monitoring point as the center. Spatial registration is performed on static and dynamic spatial data, and time sliding window reconstruction is performed on tidal velocity sequence data. The data are then spliced together to form a multidimensional spatiotemporal tensor.
[0013] The multidimensional spatiotemporal tensor is input into a spatiotemporal neural network that includes spatial feature encoding, feature dimension attention, and spatiotemporal coupling loop. The network is then trained and used to predict the power flow velocity at future time steps. The training process incorporates a physical operator constraint loss term constructed based on the continuity relationship of the power flow.
[0014] As a preferred embodiment of the present invention, outlier detection and missing value processing are performed on tidal current velocity sequence data, including:
[0015] Median filtering was used to detect and remove impulse noise and logical outliers in the tidal current velocity sequence, and Kriging interpolation was used to reconstruct missing values to obtain continuous tidal current velocity sequence data in the time dimension.
[0016] As a preferred embodiment of the present invention, the vertical acceleration characteristic is calculated based on the tidal current velocity sequence data through a preset empirical relationship, and the turbulence intensity characteristic is calculated based on the rolling standard deviation and rolling median of the tidal current velocity sequence data within a preset time window.
[0017] As a preferred technical solution of the present invention, the slope characteristics, aspect characteristics and surface roughness characteristics are obtained by performing local window analysis on the elevation data. The local window analysis determines the direction of terrain undulation and the degree of surface obstruction based on the rate of change of elevation values in different directions within the neighborhood of the monitoring point.
[0018] As a preferred technical solution of the present invention, the wind stress characteristics, friction velocity characteristics and wind energy density characteristics are calculated based on the predicted wind speed data, air density and drag coefficient, and are used to characterize the influence of the wind field on the momentum transfer of surface water.
[0019] As a preferred technical solution of the present invention, the step of establishing a local spatial analysis window centered on the monitoring point and performing spatial registration on static and dynamic spatial data includes: taking elevation data, bed sediment type data, and slope characteristics, aspect characteristics, and surface roughness characteristics derived from elevation data as static spatial data, and taking forecast wind speed data, tidal timetable data, and wind stress characteristics, friction velocity characteristics, and wind energy density characteristics derived from forecast wind speed data as dynamic spatial data, and uniformly mapping them into the local spatial analysis window.
[0020] As a preferred technical solution of the present invention, the constructed multidimensional spatiotemporal tensor is a four-dimensional tensor, which includes at least a time dimension, a spatial height dimension, a spatial width dimension, and a feature channel dimension;
[0021] The feature channel dimensions include the tidal velocity feature channel and the auxiliary feature channel consisting of elevation data, bed sediment type data, forecast wind speed data, tidal timetable data and enhanced feature set.
[0022] As a preferred technical solution of the present invention, the spatiotemporal neural network includes a spatial feature encoding submodule, a feature dimension attention mechanism submodule, a spatiotemporal coupling loop submodule, and an output submodule;
[0023] The spatial feature encoding submodule is used to perform convolutional encoding on the multidimensional spatiotemporal tensor to generate a joint spatial representation.
[0024] The feature dimension attention mechanism submodule is used to generate channel weights based on the contribution of each feature channel to the tidal current velocity prediction result and to perform feature recalibration on the joint space representation.
[0025] The spatiotemporal coupling loop submodule is used to perform spatiotemporal correlation modeling on the time series features after feature recalibration.
[0026] The output submodule is used to output the power flow velocity prediction results for future time steps.
[0027] As a preferred technical solution of the present invention, the spatial feature encoding module includes a multi-layer convolutional structure, each convolutional layer uses a convolutional kernel to extract local spatial patterns, wherein a normalization processing unit is set after at least two convolutional layers.
[0028] The feature dimension attention mechanism module includes a global average pooling unit, a feature compression unit, a weight activation unit, and a channel weighted operation unit;
[0029] The spatiotemporal coupling loop module includes multiple ConvLSTM2D units, which are used to jointly model the temporal dependencies and spatial correlations in multidimensional spatiotemporal tensors.
[0030] As a preferred technical solution of the present invention, the physical operator constraint loss term is constructed based on the power flow continuity relationship. The physical operator constraint loss term and the data fidelity loss term together form the objective function for network training, so as to simultaneously constrain the degree of fit between the prediction result and the historical samples and the degree of satisfaction of the prediction result with the physical conservation relationship.
[0031] The technical effects and advantages provided by the present invention in the above technical solution are as follows:
[0032] This invention improves the continuity and effectiveness of input data in the temporal dimension by detecting outliers and reconstructing missing values in tidal velocity sequence data. Based on tidal velocity, elevation, and predicted wind speed, it constructs enhanced features such as vertical acceleration, turbulence intensity, slope, aspect, surface roughness, wind stress, friction velocity, and wind energy density, expanding the original single flow velocity information into a multi-dimensional set of physical features including flow response, topographic constraints, and wind field driving forces. Simultaneously, by establishing a local spatial analysis window centered on monitoring points, spatial registration is performed on static and dynamic spatial data, and a multi-dimensional spatiotemporal tensor is reconstructed using a time sliding window, ensuring that the network input possesses temporal continuity, spatial neighborhood, and multi-feature correlation. Furthermore, it combines spatial feature encoding, feature dimension attention, and spatiotemporal coupling cyclic modeling, and introduces a physical operator constraint loss term constructed based on tidal continuity during training. This allows the prediction results to not only more fully characterize the tidal evolution process in tidal flat areas but also maintain physical consistency, thereby improving the accuracy, stability, and generalization ability of tidal velocity prediction in complex tidal flat environments. Attached Figure Description
[0033] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0034] Figure 1 This is a general technical flow chart of the present invention;
[0035] Figure 2 This is a schematic diagram of the spatiotemporal neural network of the present invention. Detailed Implementation
[0036] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.
[0037] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The described embodiments are only a part of the embodiments of this application, not all of them. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.
[0038] Example 1
[0039] like Figure 1 As shown, this embodiment provides a tidal current velocity prediction method based on data augmentation and spatiotemporal neural networks, which is applicable to continuous monitoring scenarios in tidal flat areas under the combined action of waves and currents and the alternating conditions of open and submerged areas. By deploying a single vertical acceleration acquisition unit in the target area, the vertical acceleration response signal caused by the movement of the medium near the bottom is obtained, and the water flow velocity, wave trajectory velocity and wave height are extracted based on the signal, thereby calculating the continuous time series of bottom shear stress.
[0040] In this embodiment, the monitoring area is selected as an observation section on the nearshore tidal flat profile, with an elevation range of -0.8m to +1.6m, and the bottom sediment is mainly silty sand. A monitoring device is installed along the middle of the section, comprising a protective shell, a vertical acceleration sensor, a data acquisition module, a clock synchronization module, a storage module, and a power supply module. The vertical acceleration sensor is fixedly installed inside the protective shell, with its sensitive axis set vertically. The height of the sensor installation point from the bottom surface is controlled within the range of 0.03m to 0.08m. In this embodiment, the installation height is 0.05m. The sampling frequency of the data acquisition module is set to 16Hz for continuously recording the raw vertical acceleration sequence.
[0041] After monitoring begins, a buoy-type tidal current velocity measurement device is deployed in the target tidal flat area. The vertical acceleration sensor continuously outputs the raw sequence of vertical acceleration, which is denoted as:
[0042] A = {a1, a2, …, aL};
[0043] Where: L is the past time step, and aL represents the vertical acceleration value at the Lth sampling time. In this embodiment, data is acquired over 24 hours with a sampling interval of 5 minutes, so the total time step L = 1440.
[0044] Simultaneously acquire auxiliary environmental data B for the area, including elevation data b1, bottom sediment type data b2, forecast wind speed data b3, and tidal timetable data b4.
[0045] To address abnormal fluctuations and data loss caused by sensor obstruction and signal transmission failures during monitoring, a data cleaning process is performed on the original sequence data A, specifically including the following steps: 1) Outlier detection: Median filtering algorithm is used to remove impulse noise and logical outliers in A; 2) Missing value reconstruction: Kriging interpolation method is used to fill in data gaps to ensure the continuity of A in the time dimension.
[0046] Data augmentation is performed on the cleaned sequence data A and the auxiliary environmental data to construct more features, including the following steps:
[0047] 1) The vertical acceleration b5 is calculated from the tidal velocity a using Formula 1 to capture transient changes in the vertical direction; the turbulence intensity b6 is calculated using Formula 2, which helps the neural network identify complex flow regime transitions.
[0048] Formula 1: ;
[0049] Formula 2: ;
[0050] Where k1 and k2 are empirical constants, typically taken as 0.944 and -1.107 respectively; σ and μ are the 15-minute rolling standard deviation and rolling median of the sequence data A, respectively;
[0051] 2) Based on the 3×3 window analysis algorithm, the elevation data b1 is used to calculate the slope b7, aspect b8, and surface roughness b9 using equations 3, 4, and 5 respectively:
[0052] Formula 3: ;
[0053] Formula 4: ;
[0054] Formula 5: ;
[0055] in, This represents the rate of change of the elevation value z in the x-direction. It represents the rate of change of the elevation value z in the y direction; atan2 is the four-quadrant arctangent function, which returns the azimuth of a point on a plane;
[0056] 3) Based on the predicted wind speed data b3, wind stress b10, friction velocity b11, and wind energy density b12 are calculated using equations 6, 7, and 8 respectively, to improve the spatiotemporal neural network model's ability to perceive the effects of wind-induced flow and momentum transfer in surface water.
[0057] Formula 6: ;
[0058] Formula 7: ;
[0059] Formula 8: ;
[0060] Where ρa is the air density, and in this embodiment, the constant Cd is taken as the dimensionless drag force coefficient, which is usually set to the empirical constant 0.0012.
[0061] 4) Construct a local spatial analysis window centered on the buoy deployment point, and spatially register the static b1-2, b4-9 with the dynamic b3, b10-12; then reconstruct A along the time axis L using a time sliding window, and concatenate it with the corresponding auxiliary environmental data B to finally construct a multidimensional spatiotemporal tensor used as input, denoted as... L represents the past time step, H×W is the local grid window centered on the buoy's position, and C is the feature dimension, including 13 dimensions in total, a and b1-12.
[0062] This can be understood as deploying a buoy-type tidal current velocity measurement device in the target tidal flat area to collect tidal current velocity sequence data a in real time; simultaneously acquiring elevation data b1, bottom sediment type data b2, forecast wind speed data b3, and tidal timetable data b4 corresponding to the tidal flat area. Among these, the tidal current velocity sequence data a represents one-dimensional raw observation data that varies over time; the elevation data b1 and bottom sediment type data b2 represent regional spatial attribute data; and the forecast wind speed data b3 and tidal timetable data b4 represent time-varying auxiliary environmental data.
[0063] Preprocessing of the raw data, including outlier detection and missing value detection, ensures the continuity of the original tidal velocity sequence over time. Vertical acceleration (b5) and turbulence intensity (b6) are derived from the tidal velocity sequence data (a); slope aspect (b7), slope aspect (b8), and surface roughness (b9) are derived from the elevation data (b1); and wind stress (b10), friction velocity (b11), and wind energy density (b12) are derived from the predicted wind speed data (b3). Thus, the original single tidal velocity observation is expanded into a multi-dimensional environmental representation encompassing flow response, topographic constraints, and wind field effects. Subsequently, a spatiotemporal tensor construction step is performed, with two processing modules: "time sliding window reconstruction" and "dynamic / static spatial registration." "Time sliding window reconstruction" slices the historical tidal velocity sequence along the time axis, reconstructing continuous observation data into past time windows suitable for sequence prediction; "dynamic / static spatial registration" aligns static geographic attribute data with dynamic environmental driving data under a unified spatial reference. The static data includes at least elevation data b1, bottom sediment type data b2, and slope, aspect, and surface roughness derived from elevation; the dynamic data includes at least forecast wind speed data b3 and wind stress, friction velocity, and wind energy density derived from it, as well as tidal timetable data b4 corresponding to time. The multidimensional spatiotemporal tensor is input into the enhanced neural network.
[0064] To further explain, such as Figure 2 The spatiotemporal neural network shown has an input-output interface, and the spatiotemporal neural network includes:
[0065] 1) Input layer: used to receive a multidimensional spatiotemporal tensor X of shape L×H×W×C;
[0066] 2) Spatial feature encoding module: Convolutional layer 1 containing 32 feature maps, convolutional layer 2 containing 64 feature maps, and convolutional layer 3 containing 128 feature maps. Batch normalization is performed on convolutional layer 1 and convolutional layer 2. All convolutional layers use 3×3 convolutional kernels and ReLU activation function.
[0067] 3) Feature Dimension Attention Mechanism Module: Includes global average pooling layer, feature compression layer, weight activation layer, and channel weighted operation layer;
[0068] 4) Spatiotemporally Coupled Recurrent Layer: 128 ConVLSTM2D units with 3×3 convolutional kernels to capture temporal dynamics and spatial constraints;
[0069] 5) Output layer: contains tensor flattening and fully connected layers, and finally outputs the predicted value of power flow velocity aL+1 at time step T+1 and its quantile.
[0070] Furthermore, the spatiotemporal neural network is trained using the loss functions of Equations 9 and 10:
[0071] Formula 9: ;
[0072] Formula 10: ;
[0073] Where L is the hyperparameter weight coefficient of λ1, λ2, λ3, and λ4, Ldfl is the data fidelity loss, Lpde is the physical operator constraint loss, Lspc is the frequency domain consistency loss, and Lshp is the spatiotemporal gradient smoothing loss; η is the instantaneous height of the water surface relative to the mean sea level of the tidal timetable data b4, h is the total lost water depth, u and v are velocity vector components, x and y are the horizontal and vertical coordinates respectively, t represents the continuous time dimension, and ||·||2 represents the Euclidean norm, used to quantify the global energy of the physical residual.
[0074] The spatiotemporal neural network is trained with a large amount of past data. A spatiotemporal tensor X of L=1440 is input in real time to the model. The future time step to be predicted is T=120 (i.e., the next 2 hours). The final output is A={a1, a2, …, a1440,a1441, …, a1580}.
[0075] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the scope of protection of the appended claims.
[0076] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A power flow velocity prediction method based on data augmentation and spatiotemporal neural networks, characterized in that, Includes the following steps: Acquire tidal velocity sequence data, elevation data, bed sediment type data, forecast wind speed data, and tidal timetable data for the target tidal flat area; Outlier detection and missing value processing are performed on the tidal current velocity sequence data; Vertical acceleration and turbulence intensity features are constructed based on the processed tidal velocity sequence data; slope, aspect, and surface roughness features are constructed based on elevation data; and wind stress, friction velocity, and wind energy density features are constructed based on forecast wind speed data. A local spatial analysis window is established with the monitoring point as the center. Spatial registration is performed on static and dynamic spatial data, and time sliding window reconstruction is performed on tidal velocity sequence data. The data are then spliced together to form a multidimensional spatiotemporal tensor. The multidimensional spatiotemporal tensor is input into a spatiotemporal neural network that includes spatial feature encoding, feature dimension attention, and spatiotemporal coupling loop. The network is then trained and used to predict the power flow velocity at future time steps. The training process incorporates a physical operator constraint loss term constructed based on the continuity relationship of the power flow.
2. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 1, characterized in that, Outlier detection and missing value handling are performed on tidal current velocity sequence data, including: Median filtering was used to detect and remove impulse noise and logical outliers in the tidal current velocity sequence, and Kriging interpolation was used to reconstruct missing values to obtain continuous tidal current velocity sequence data in the time dimension.
3. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 2, characterized in that, The vertical acceleration characteristic is calculated based on the tidal current velocity sequence data through a preset empirical relationship, and the turbulence intensity characteristic is calculated based on the rolling standard deviation and rolling median of the tidal current velocity sequence data within a preset time window.
4. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 3, characterized in that, The slope characteristics, aspect characteristics, and surface roughness characteristics are obtained by performing local window analysis on the elevation data. The local window analysis determines the direction of terrain undulation and the degree of surface obstruction based on the rate of change of elevation values in different directions within the neighborhood of the monitoring point.
5. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 4, characterized in that, The wind stress characteristics, friction velocity characteristics, and wind energy density characteristics are calculated based on forecast wind speed data, air density, and drag coefficient, and are used to characterize the influence of the wind field on the momentum transfer of surface water.
6. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 5, characterized in that, The establishment of a local spatial analysis window centered on the monitoring point and the spatial registration of static and dynamic spatial data include: taking elevation data, bed sediment type data, and slope, aspect, and surface roughness characteristics derived from elevation data as static spatial data, and taking forecast wind speed data, tidal timetable data, and wind stress, friction velocity, and wind energy density characteristics derived from forecast wind speed data as dynamic spatial data, and mapping them uniformly into the local spatial analysis window.
7. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 1, characterized in that, The constructed multidimensional spatiotemporal tensor is a four-dimensional tensor, which includes at least the time dimension, spatial height dimension, spatial width dimension, and feature channel dimension; The feature channel dimensions include the tidal velocity feature channel and the auxiliary feature channel consisting of elevation data, bed sediment type data, forecast wind speed data, tidal timetable data and enhanced feature set.
8. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 7, characterized in that, The spatiotemporal neural network includes a spatial feature encoding submodule, a feature dimension attention mechanism submodule, a spatiotemporal coupling loop submodule, and an output submodule; The spatial feature encoding submodule is used to perform convolutional encoding on the multidimensional spatiotemporal tensor to generate a joint spatial representation. The feature dimension attention mechanism submodule is used to generate channel weights based on the contribution of each feature channel to the tidal current velocity prediction result and to perform feature recalibration on the joint space representation. The spatiotemporal coupling loop submodule is used to perform spatiotemporal correlation modeling on the time series features after feature recalibration. The output submodule is used to output the power flow velocity prediction results for future time steps.
9. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 8, characterized in that, The spatial feature encoding module includes a multi-layer convolutional structure, each convolutional layer uses a convolutional kernel to extract local spatial patterns, and a normalization processing unit is set after at least two convolutional layers. The feature dimension attention mechanism module includes a global average pooling unit, a feature compression unit, a weight activation unit, and a channel weighted operation unit; The spatiotemporal coupling loop module includes multiple ConvLSTM2D units, which are used to jointly model the temporal dependencies and spatial correlations in multidimensional spatiotemporal tensors.
10. The power flow velocity prediction method based on data augmentation and spatiotemporal neural networks according to claim 9, characterized in that, The physical operator constraint loss term is constructed based on the power flow continuity relationship. The physical operator constraint loss term and the data fidelity loss term together form the objective function for network training, so as to simultaneously constrain the degree of fit between the prediction result and the historical samples and the degree to which the prediction result satisfies the physical conservation relationship.