A wave field long-time prediction method and device, a terminal and a storage medium

Abstract

This invention discloses a method, apparatus, terminal, and storage medium for long-term prediction of wave fields, relating to the field of spatiotemporal signal processing. The method constructs a model input tensor using historical sample data of wave fields in a target region; it then uses a trained long-term prediction model to predict the wave field prediction data for the current prediction step based on the model input tensor; employing a multi-step autoregressive rolling mechanism, it updates the model input tensor based on the wave field prediction data, and determines the wave field prediction data for each prediction step using the updated model input tensor through the trained long-term prediction model; finally, it determines the short-term, medium-term, and long-term prediction sequences of the wave field based on the wave field prediction data for each prediction step through the output heads of each prediction span, and fuses them to determine the long-term wave field prediction result. Therefore, it effectively solves the problem that existing technologies cannot improve long-term rolling stability while reducing modeling complexity and taking into account the propagation characteristics at different time scales.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal signal processing, and more particularly to a method, apparatus, terminal, and storage medium for long-term prediction of wave fields. Background Technology

[0002] In applications such as seismic wave propagation analysis, earthquake early warning, and engineering safety assessment, it is often necessary to predict the evolution of the wave field in future periods based on historical or early observation data of the target area.

[0003] In existing technologies, wave field extrapolation mainly falls into two categories. One category relies on propagation equations and numerical discrete calculations for extrapolation, which has clear physical logic. However, in large-scale, high-resolution, and long-term prediction scenarios, it suffers from high computational load and low extrapolation efficiency, failing to meet the needs of rapid prediction. The other category uses deep learning models to fit the spatiotemporal evolution of wave fields, which can reduce online computation and improve prediction speed, but it has significant drawbacks: rolling predictions are prone to error accumulation over time, resulting in poor stability of long-term prediction results; high-dimensional joint modeling under large grids has high complexity and storage costs; and it is difficult to simultaneously consider the propagation details and overall evolution trends at different time scales, easily leading to missing details or trend deviations.

[0004] Therefore, how to improve long-term rolling stability while reducing modeling complexity and taking into account the propagation characteristics at different time scales has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a long-term prediction method, device, terminal and storage medium for wave fields, which addresses the above-mentioned deficiencies of the prior art. The aim is to solve the problem that the prior art cannot improve long-term rolling stability while reducing modeling complexity and taking into account the propagation characteristics of different time scales.

[0006] The technical solution adopted by this invention to solve the problem is as follows: In a first aspect, embodiments of the present invention provide a long-term prediction method for wave fields, wherein the method includes: The model input tensor is constructed based on historical sample data of the wave field in the target region. Based on the input tensor of the trained long-term prediction model, the fluctuation field prediction data corresponding to the current prediction step is determined. A multi-step autoregressive rolling mechanism is adopted to update the model input tensor based on the fluctuation field prediction data, and the fluctuation field prediction data for each prediction step is determined by the trained long-term prediction model based on the updated model input tensor. Based on the wave field prediction data of each prediction step, the short-term prediction sequence, medium-term prediction sequence, and long-term prediction sequence of the wave field are determined by the output heads of each prediction span of the trained long-term prediction model. The short-term prediction sequence, the medium-term prediction sequence, and the long-term prediction sequence are fused using the fusion module of the trained long-term prediction model to determine the long-term fluctuation field prediction result.

[0007] In one implementation method, a model input tensor is constructed based on historical sample data of the fluctuation field in the target region, including: The historical sample data of the wave field is divided based on the input window and the output window to determine each input time series and the corresponding output time series; The input time series and the corresponding output time series are downsampled according to the prediction time interval to determine each supervised learning sample pair; The supervised learning sample pairs are subjected to spatial grid organization, channel rearrangement and normalization to determine the model input tensor.

[0008] In one implementation, the long-term prediction model further includes: an input mapping layer, a position encoding layer, a backbone network, and an output decoding layer. Based on the input tensor of the trained long-term prediction model, the fluctuation field prediction data corresponding to the current prediction step is determined, including: The input mapping layer maps the model input tensor into high-dimensional implicit features to obtain the initial latent space features. The initial latent space features are encoded using Gaussian Fourier features through the location encoding layer to determine the latent space encoded features; The backbone network is used to extract features based on the latent space coding features to determine the coupled evolution features. The coupled evolution features are decoded by the output decoding layer to determine the wave field prediction data corresponding to the current prediction step.

[0009] In one implementation, the backbone network decomposes the joint spatial attention computation into directional low-rank kernel attention along different directions, and recursively updates the latent space encoding features using a shared parameter multiple iterative update method to determine the coupled evolution features.

[0010] In one implementation method, a multi-step autoregressive rolling mechanism is employed to update the model input tensor based on the fluctuation field prediction data, including: Remove the first frame of the model input tensor in the current prediction step and add the fluctuation field prediction data as the last frame of the model input tensor to determine the updated model input tensor.

[0011] In one implementation method, the training method of the long-term prediction model includes: Perform single-step prediction training on the long-term prediction model; The long-term prediction model trained based on single-step prediction is trained by performing multi-step autoregression training for a first preset number of time steps, and the model parameters are updated based on the multi-step prediction loss. The long-term prediction model trained by multi-step autoregression is fine-tuned for a second preset number of time steps to determine the trained long-term prediction model.

[0012] In one implementation, the method further includes: When performing multi-step autoregressive expansion training on a long-term prediction model trained on a single-step prediction basis, random noise is added to the input window.

[0013] Secondly, embodiments of the present invention also provide a long-term wave field prediction device, wherein the long-term wave field prediction device includes: The data acquisition module is used to construct the model input tensor based on historical sample data of the fluctuation field in the target area; The single-step prediction module is used to determine the fluctuation field prediction data corresponding to the current prediction step based on the input tensor of the trained long-term prediction model. The autoregressive rolling prediction module is used to update the model input tensor based on the fluctuation field prediction data using a multi-step autoregressive rolling mechanism, and to determine the fluctuation field prediction data for each prediction step based on the updated model input tensor using a trained long-term prediction model. The sequence output module is used to determine the short-term, medium-term, and long-term prediction sequences of the fluctuation field based on the fluctuation field prediction data of each prediction step, using the prediction span output heads of the trained long-term prediction model. The multi-scale fusion module is used to fuse the short-term prediction sequence, the medium-term prediction sequence, and the long-term prediction sequence through the fusion module of the trained long-term prediction model to determine the long-term fluctuation field prediction result.

[0014] Thirdly, embodiments of the present invention also provide a terminal, the terminal including a memory and one or more processors; the memory stores one or more programs; the programs include instructions for executing the long-term prediction method of the fluctuation field as described above; the processor is used to execute the programs.

[0015] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a plurality of instructions, wherein the instructions are adapted to be loaded and executed by a processor to implement any of the above-described long-term prediction methods for wave fields.

[0016] The beneficial effects of this invention are as follows: In this embodiment, a model input tensor is constructed using historical sample data of the fluctuation field in the target region; a trained long-term prediction model predicts the fluctuation field prediction data for the current prediction step based on the model input tensor; a multi-step autoregressive rolling mechanism is employed to update the model input tensor based on the fluctuation field prediction data, and the trained long-term prediction model determines the fluctuation field prediction data for each prediction step based on the updated model input tensor; the output heads of each prediction span of the trained long-term prediction model determine the short-term, medium-term, and long-term prediction sequences of the fluctuation field based on the fluctuation field prediction data for each prediction step, and these sequences are then fused to determine the long-term fluctuation field prediction result. Therefore, this invention effectively solves the problem that existing technologies cannot improve long-term rolling stability while reducing modeling complexity and taking into account the propagation characteristics at different time scales. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the long-term prediction method for wave fields provided in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram illustrating a specific implementation of the long-term prediction method for wave fields provided in this embodiment of the invention.

[0020] Figure 3 This is a schematic diagram illustrating the model training, combined loss target, and parameter update process of the long-term prediction method for wave fields provided in this embodiment of the invention.

[0021] Figure 4 This is a schematic diagram of the internal modules of the long-term wave field prediction device provided in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of the terminal provided in the embodiment of the present invention. Detailed Implementation

[0023] This invention discloses a method, apparatus, terminal, and storage medium for long-term prediction of wave fields. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0024] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0025] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0026] In applications such as seismic wave propagation analysis, earthquake early warning, and engineering safety assessment, it is often necessary to predict the evolution of the wave field in future periods based on historical or early observation data of the target area. This type of technology can be used for tasks such as emergency decision support, structural response analysis, propagation mechanism research, and supplementary calculations for numerical simulations.

[0027] In existing technologies, one type of method mainly relies on solving the propagation equation and numerical discretization to extrapolate wave fields. This type of method has the advantage of clear physical meaning, but in scenarios with large spatial ranges, high grid resolutions, and long prediction time domains, it usually suffers from problems such as large computational load, slow extrapolation speed, and difficulty in meeting the needs of rapid prediction.

[0028] Another approach employs deep learning models to learn the evolution of wave fields in both time and space dimensions, thereby reducing the computational cost of online inference. While this approach has advantages in prediction speed, it still has shortcomings in long-term prediction tasks for high-dimensional wave fields: First, as the prediction duration increases, rolling prediction errors tend to accumulate, leading to decreased stability of results over longer periods; second, wave fields exhibit significant spatiotemporal coupling characteristics, and direct high-dimensional joint modeling in large-grid scenarios often results in high computational complexity and large storage overhead; third, it is difficult to simultaneously capture propagation details and overall trends at different time scales, easily leading to the loss of local details or long-term trend drift.

[0029] To address the aforementioned deficiencies in existing technologies, this invention provides a long-term fluctuation field prediction method. The method constructs a model input tensor using historical fluctuation field sample data of the target region; it then uses a trained long-term prediction model to predict the fluctuation field prediction data for the current prediction step based on the model input tensor; employing a multi-step autoregressive rolling mechanism, it updates the model input tensor based on the fluctuation field prediction data, and uses the trained long-term prediction model to determine the fluctuation field prediction data for each prediction step based on the updated model input tensor; finally, it uses the prediction span outputs of the trained long-term prediction model to determine the short-term, medium-term, and long-term fluctuation field prediction sequences based on the fluctuation field prediction data for each prediction step, and fuses these sequences to determine the long-term fluctuation field prediction result.

[0030] This method, to consider the short-term local details, mid-term propagation structure, and long-term overall evolution trend of the wave field, sets up at least two output heads with different prediction spans. Each output head generates corresponding wave field prediction results for different time scales. The shorter prediction span output head preserves the fine local changes in near-term predictions, while the longer prediction span output head enhances the modeling ability for long-term propagation trends and overall evolution patterns. Furthermore, the prediction results generated by the different prediction span output heads are fused to obtain the final long-term wave field prediction result, thereby improving the stability and accuracy of the model in long-term rolling prediction tasks. Therefore, it effectively solves the problem that existing techniques cannot improve long-term rolling stability while reducing modeling complexity and considering the propagation characteristics at different time scales.

[0031] Exemplary method: like Figure 1 The long-term wave field prediction method shown is applied to the scenario of predicting future periods of regional two-dimensional grid wave fields. The wave field can be a physical wave field discretely represented on a two-dimensional or three-dimensional spatial grid; preferably, the wave field is a seismic wave field, specifically a three-component velocity field, wherein the three-component velocity field includes a velocity component in the X direction, a velocity component in the Y direction, and a velocity component in the Z direction. By unifying wave information from multiple directions into the same prediction framework, continuous evolution prediction of the wave field in the target area can be achieved while maintaining the correlation of multi-directional propagation information. The method includes: Step S100: Construct the model input tensor based on historical sample data of the fluctuation field in the target region; In simple terms, based on the historical sample data of the wave field already collected in the target area, and according to the preset temporal arrangement and spatial dimension division rules, the historical sample data of the wave field at different times and spatial locations are organized and combined, the data dimensions and formats are unified, and the data is systematically spliced ​​to form a multidimensional input tensor that conforms to the model input specifications, thus completing the structured construction of the model's pre-input data.

[0032] The historical sample data can be derived from numerical simulation data, measured observation data, inversion and reconstruction data, or any combination of the above. Preferably, the historical sample data is organized into a continuous time frame sequence in chronological order, and each moment corresponds to a multi-component fluctuation field value on the same spatial grid.

[0033] To address the inconsistencies in the number of channels, duration, and grid organization of historical sample data from different sources of wave fields, the original samples were first organized into a unified five-dimensional tensor. , is represented as: , In the formula, For the set of real numbers, Indicates the number of sample trajectories. Indicates the number of channels. and These represent the number of discrete points in the spatial grid in the two directions, respectively. Indicates the time frame number; when the wave field is a three-component velocity field, .

[0034] After unifying the representation of the raw data, historical fluctuation field samples from different sources and of different lengths can be incorporated into the same processing framework, thereby reducing the data adaptation complexity during subsequent model training and deployment, and providing a unified interface for window segmentation, normalization and batch inference.

[0035] In one implementation, the model input tensor is constructed based on historical sample data of the fluctuation field in the target region, including: Step S101: Divide the historical sample data of the fluctuation field based on the input window and the output window to determine each input time series and the corresponding output time series; Step S102: Downsample the input time series and the corresponding output time series according to the prediction time interval to determine each supervised learning sample pair; Step S103: Perform spatial grid organization, channel rearrangement, and normalization on each supervised learning sample pair to determine the model input tensor.

[0036] Specifically, to enable the model to make predictions over longer time spans using historical wave field sample data of a finite length, continuous time series need to be constructed as input and output windows. Based on these windows, corresponding input and output time series are divided from the historical wave field sample data, allowing the model to learn the evolution patterns of the wave field at future moments based on historical propagation information. Simultaneously, the divided input and output time series are sampled at preset time intervals, and the resulting discrete input and output time samples are combined to obtain supervised learning sample pairs. This allows for the coverage of longer physical time spans with the same number of rolling prediction steps, thereby improving long-term prediction efficiency and reducing the number of rolling steps.

[0037] Let the first If the window starts at a given time, then the input window can be represented as: , The corresponding output window can be represented as: , In the formula, Indicates time The wave field, Indicates the start time of the input window. Indicates the length of the input window. Indicates the length of the output window. Indicates the time downsampling interval. , These represent length variables.

[0038] For example, frames of the fluctuating field can be extracted at fixed time intervals, such as one frame every four original time steps, thereby expanding the physical time span corresponding to a single-step prediction without significantly increasing the number of model rolls. This approach can reduce temporal redundancy while preserving the main propagation information and improve the model's ability to model long-scale propagation processes.

[0039] In one implementation, the input window length It can be set to 30, the time downsampling interval. It can be set to 4.

[0040] After obtaining supervised learning sample pairs, it is also necessary to unify the data scale of different channels to avoid the difference in scale affecting the model optimization process.

[0041] To mitigate training instability caused by differences in the dimensions of different physical components, the supervised learning sample pairs are standardized by channel, as follows: , In the formula, This represents the c-th channel input. and Let these represent the mean and standard deviation of the c-th channel on the training set, respectively. To prevent stable terms with a denominator of zero.

[0042] The standardization process described above adjusts data from different channels to similar numerical scales, thereby improving convergence speed during model training and numerical stability during inference. After standardization, the data is then rearranged into a tensor form suitable for model input, which can then be used by the subsequent prediction model.

[0043] Step S200: Determine the fluctuation field prediction data corresponding to the current prediction step based on the input tensor of the trained long-term prediction model; In simple terms, the trained long-term prediction model is invoked, the constructed model is input into the tensor, and based on the spatiotemporal evolution characteristics and time series inference rules of the wave field learned by the model, the wave field prediction data corresponding to the current prediction step is output through model calculation and inference.

[0044] In one implementation, in addition to the output heads for each prediction span and the fusion module, the long-term prediction model further includes: an input mapping layer, a position encoding layer, a backbone network, and an output decoding layer. Based on the input tensor of the trained long-term prediction model, the fluctuation field prediction data corresponding to the current prediction step is determined, including: The input mapping layer maps the model input tensor into high-dimensional implicit features to obtain the initial latent space features. The initial latent space features are encoded using Gaussian Fourier features through the location encoding layer to determine the latent space encoded features; The backbone network is used to extract features based on the latent space coding features to determine the coupled evolution features. The coupled evolution features are decoded by the output decoding layer to determine the wave field prediction data corresponding to the current prediction step.

[0045] Specifically, such as Figure 2 As shown, to enhance the model's ability to characterize spatial relationships and propagation path relationships, the model input tensor is first mapped to the latent space through an output mapping layer, and positional encoding is superimposed to obtain the initial latent space features, represented as: , In the formula, This represents the normalized input window. Indicates the input mapping operator, Indicates the position encoding operator, Represents a set of spatial coordinates. This represents the initial latent space features.

[0046] The above formula shows that this embodiment does not directly perform calculations on the original grid values, but first converts the input wave field into a high-dimensional latent feature representation with position-aware capabilities. This is beneficial for expressing the wavefront propagation direction, propagation path, and local spatial coupling relationship, and also provides a more suitable feature foundation for subsequent spatiotemporal joint modeling.

[0047] In the positional encoding layer, Gaussian Fourier transform is used to encode the two-dimensional or three-dimensional spatial coordinates of the initial latent space features, resulting in latent space encoded features. This enhances the model's ability to perceive coordinate continuity, wavefront position changes, and regional structural differences. Compared to using only the original grid index, this positional encoding method can better characterize the spatial phase relationships and propagation path dependencies during wave propagation.

[0048] Furthermore, to address the issues of unstable training, excessively large parameter size, and high cost of spatial dependency modeling in long-term prediction of high-dimensional fluctuating fields, the backbone network employs an implicit iterative update mechanism with shared parameters to extract features from the latent space encoded features and determine the coupled evolutionary features, represented as: , In the formula, Indicates the first Latent space encoding features of the next iteration. Indicates the total number of iterations. This indicates a shared parameter update operator. and This represents the attention kernel in two spatial directions.

[0049] This mechanism can increase the effective depth of the network without significantly increasing the number of model parameters, enabling the model to gradually extract higher-level propagation and evolution features while maintaining the stability of the training process and the efficiency of parameter utilization. Since each iteration is performed within a shared parameter framework, this embodiment achieves a good balance between model depth and parameter size.

[0050] In one implementation, to reduce the computational complexity of traditional two-dimensional joint attention on large-scale grids, the backbone network decomposes the joint spatial attention computation into directional low-rank kernel attention along different directions, and uses a multi-iterative update method with shared parameters to recursively update the latent space encoded features to determine the coupled evolution features.

[0051] Specifically, the low-rank kernel along the x-direction can be represented as: , The low-rank kernel along the y-direction can be represented as: , in, , Indicates directional query characteristics. , Indicating directional bond features, , , , They represent learnable mapping functions, This indicates transpose.

[0052] This method employs a backbone network based on an implicit iterative factor decomposition attention structure to decompose the originally highly complex two-dimensional joint attention into low-rank kernel attention modeling in two directions. This approach can reduce computational and storage overhead while maintaining the ability to express wave propagation relationships, making the model more suitable for wave field prediction tasks with large grids and long time periods.

[0053] For the coupled evolution characteristics output by the backbone network, the output decoding layer restores them to the two-dimensional three-component wave field corresponding to the current prediction step, that is, the wave field prediction data corresponding to the current prediction step.

[0054] Step S300: Using a multi-step autoregressive rolling mechanism, the model input tensor is updated based on the fluctuation field prediction data, and the fluctuation field prediction data for each prediction step is determined by the trained long-term prediction model based on the updated model input tensor. To extend the single-step prediction to obtain a continuous sequence of fluctuation field predictions for multiple future time periods, a multi-step autoregressive rolling iteration mechanism is introduced based on the long-term prediction model. The obtained fluctuation field prediction data updates the original input tensor, completing the sliding iteration of the input time series window. Then, the updated model input tensor is continuously fed into the fully trained long-term prediction model, and the fluctuation field prediction results corresponding to each prediction step are iteratively calculated in turn, realizing continuous long-term autonomous extrapolation.

[0055] In one implementation, a multi-step autoregressive rolling mechanism is employed to update the model input tensor based on the fluctuation field prediction data, including: Remove the first frame of the model input tensor in the current prediction step and add the fluctuation field prediction data as the last frame of the model input tensor to determine the updated model input tensor.

[0056] Specifically, let the first The model input tensor corresponding to each step input window is: The predicted data of the fluctuation field in the current prediction step can be expressed as: , in, This represents a long-term prediction model. Indicates the first The output of the wave field prediction data.

[0057] Get the first After the prediction result is obtained, it is written to the end of the input window, and the oldest historical frame in the window is removed to form the next input window, which is the model input tensor for the next step. , is represented as: , in, This represents an input window update operation that deletes the earliest frame along the time dimension and shifts the entire window to the left. This indicates that the fluctuation field prediction data of the current prediction step is spliced ​​together at the end of the time dimension.

[0058] By continuously feeding the fluctuation field prediction data from the current prediction step back into the input window for the next step, a continuous fluctuation field prediction sequence for multiple future moments in the target region can be gradually generated. Compared with directly predicting the entire future sequence at once, this method is more convenient for continuously extrapolating the future field evolution process under limited model capacity.

[0059] Step S400: Determine the short-term prediction sequence, medium-term prediction sequence and long-term prediction sequence of the fluctuation field based on the fluctuation field prediction data of each prediction step through the prediction span output heads of the trained long-term prediction model. To account for short-term local details, medium-term propagation structure, and long-term overall trends, multiple prediction output heads can be set to output the fluctuation field prediction results corresponding to different prediction spans. For the s-th prediction head, its output can be expressed as: , in, Indicates the first The mapping function corresponding to each prediction head Indicates the first Step-by-step wave field prediction data, , , These represent the short-time prediction output header, the medium-time prediction output header, and the long-time prediction output header, respectively. Indicates the first Step on the scale The predicted data of the fluctuation field obtained below.

[0060] The above formulas show that different prediction heads focus on propagation information at different time scales. Specifically, short-time prediction heads emphasize near-field details and local wavefront morphology, medium-time prediction heads emphasize the transition process of the propagation structure, and long-time prediction heads emphasize the overall evolution trend and far-field stability. Multi-head parallel modeling can improve the model's ability to represent features at different time scales.

[0061] Step S500: The short-term prediction sequence, the medium-term prediction sequence, and the long-term prediction sequence are fused using the fusion module of the trained long-term prediction model to determine the long-term fluctuation field prediction result.

[0062] To improve the stability of the connection between prediction results at different scales, the fusion module uses a gated routing fusion method to generate fusion weights, represented as: , The final fused prediction result can be expressed as: , in, Indicates the first The fusion weight generation function corresponding to each scale Indicates the first Step 1 The fusion weights of the prediction heads, This represents the prediction result of the fused long-term fluctuation field.

[0063] The above formula can adaptively allocate the contribution of different prediction span results according to the current time series status, thereby improving the continuity and stability between cross-scale prediction results and avoiding the problem of loss of details or overall trend drift in long-term scenarios with a single prediction span scheme.

[0064] In one implementation, the prediction results corresponding to different prediction spans can also be segmented and connected according to a preset time period. When there are preset time boundary points between prediction spans, the long-term fluctuation field prediction results can be generated by segmenting and connecting them according to time periods.

[0065] In one implementation, the training method for the long-term prediction model includes: Perform single-step prediction training on the long-term prediction model; The long-term prediction model trained based on single-step prediction is trained by performing multi-step autoregression training for a first preset number of time steps, and the model parameters are updated based on the multi-step prediction loss. The long-term prediction model trained by multi-step autoregression is fine-tuned for a second preset number of time steps to determine the trained long-term prediction model.

[0066] Considering that errors in long-term prediction accumulate gradually with increasing rolling steps, this method employs a progressive multi-stage training strategy to enhance the model's multi-step prediction stability. In the first stage, single-step prediction training is performed, allowing the model to learn the mapping relationship between the historical window and the fluctuation field at the next prediction time. In the second stage, multi-step autoregressive training is performed based on the model parameters from the first stage, allowing the model to learn the error propagation law under the condition of prediction result backfeedback. In the third stage, multi-step fine-tuning training with a longer time span is performed based on the second stage to improve the model's stability and generalization ability in long-term prediction tasks. For example, the first stage can perform single-step prediction training, the second stage can perform 5-step multi-step autoregressive expansion training, and the third stage can further perform fine-tuning training with 10 or more steps.

[0067] During training, to balance robustness under large error samples with smooth optimization capability in small error regions, the Huber loss function is used during the training phase, expressed as: , in, Indicates the prediction error. δ represents the Huber threshold. A smoothed quadratic penalty is used when the error is small to facilitate the model's fine fitting of local fluctuation structures; a linear penalty is used when the error is large to avoid excessive interference from extreme error points in the overall training process. This loss function improves training stability and enhances the model's adaptability to regions of sudden fluctuations. For example, δ can be set to 0.2.

[0068] The model optimizer can be the Adam optimizer, and the learning rate can be adjusted by decaying according to the changes in the validation set loss to further improve the convergence efficiency of the training process.

[0069] In one implementation, the method further includes: When performing multi-step autoregressive expansion training on a long-term prediction model trained on a single-step prediction basis, random noise is added to the input window.

[0070] Specifically, to reduce the input distribution shift caused by the backfeeding of prediction results between the training and inference phases, random noise can be added to the input window during the multi-step training phase, as follows: , In the formula, This represents the training input after adding noise. This represents the normalized input. Indicates the noise intensity factor. This represents the random noise term.

[0071] The above steps can simulate the perturbations caused by error backfeeding during the inference phase of training, allowing the model to adapt to the input bias in autoregressive prediction earlier, thereby improving the robustness of long-term rolling prediction. For example, the noise intensity coefficient can be set to 0.01.

[0072] In one implementation, such as Figure 3 As shown, to further improve the cross-scale consistency and physical rationality of prediction results in long-term extrapolation, a physical consistency enhancement strategy can be introduced based on the multi-scale prediction and multi-step autoregressive framework. This physical consistency enhancement strategy can be used to constrain the continuity between prediction results at different time scales, the consistency of wavefront propagation direction, the smoothness of local energy distribution changes, or the rationality of prediction results near boundaries.

[0073] A physical consistency regularization term can be added to the training objective to form a joint loss function, which can be expressed as: , In the formula, Indicates the basic predicted loss. Indicates multi-scale monitoring loss. This represents a physical consistency regularization term. and These represent the corresponding weight coefficients, used to further constrain the consistency of results in the long-term prediction model across different prediction scales and rolling steps. The multi-scale supervision loss term... It can be used to enhance the consistency between short-term, medium-term, and long-term forecasts; physical consistency constraint. It can be used to enhance boundary continuity, propagation direction rationality, and overall evolutionary stability.

[0074] It should be noted that the physical consistency constraint term in this formula is not limited to a specific form. It can be set according to the specific type of wave field, application requirements, or prior knowledge. For example, it can be implemented based on local residual smoothing constraints, cross-scale difference constraints, boundary continuity constraints, or other equivalent physical law constraints.

[0075] This joint constraint approach can further reduce structural distortion and unreasonable oscillations in long-term rolling forecasts. Therefore, the joint loss function can enhance the coordination between results at different scales and the rationality of the overall propagation evolution while ensuring forecast accuracy.

[0076] Based on the above embodiments, this method has at least the following beneficial effects: This method introduces an attention structure based on implicit iterative factorization into wave field prediction, which reduces the computational complexity and storage overhead in long-term prediction of high-dimensional wave fields while maintaining the ability to represent spatiotemporal coupling propagation relationships.

[0077] This method, through the coordinated design of time downsampling, historical window construction, and autoregressive rolling mechanism, enables the prediction of fluctuation fields over a longer physical period with a limited number of rolling cycles, which is beneficial to improving the extrapolation efficiency in long-term prediction tasks.

[0078] This method enhances the model's adaptability to rolling prediction scenarios by employing a progressive multi-stage training strategy and injecting random noise during multiple training stages, thereby mitigating the cumulative impact of continuous error backfeeding during long-term inference.

[0079] This method, by setting multiple output heads with different prediction spans and performing fusion processing, can take into account the propagation information at different time scales, so that the long-term prediction results can be balanced in terms of detail, trend and stability.

[0080] Based on the above embodiments, the present invention also provides a long-term wave field prediction device, such as... Figure 4 As shown, the device includes: Data acquisition module 01 is used to construct the model input tensor based on historical sample data of the fluctuation field in the target area; The single-step prediction module 02 is used to determine the fluctuation field prediction data corresponding to the current prediction step based on the input tensor of the trained long-term prediction model. The autoregressive rolling prediction module 03 is used to update the model input tensor based on the fluctuation field prediction data using a multi-step autoregressive rolling mechanism, and to determine the fluctuation field prediction data for each prediction step based on the updated model input tensor using a trained long-term prediction model. The sequence output module 04 is used to determine the short-term prediction sequence, medium-term prediction sequence and long-term prediction sequence of the fluctuation field based on the fluctuation field prediction data of each prediction step through the prediction span output heads of the trained long-term prediction model. The multi-scale fusion module 05 is used to fuse the short-term prediction sequence, the medium-term prediction sequence, and the long-term prediction sequence through the fusion module of the trained long-term prediction model to determine the long-term fluctuation field prediction result.

[0081] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 5As shown, the terminal includes a processor, memory, network interface, and display screen connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a long-term prediction method for wave fields. The display screen can be an LCD screen or an e-ink screen.

[0082] Those skilled in the art will understand that Figure 5 The block diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0083] In one implementation, the terminal's memory stores one or more programs, and these programs are configured to be executed by one or more processors, and contain instructions for performing a long-term prediction method for wave fields.

[0084] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0085] In summary, this invention discloses a method, apparatus, terminal, and storage medium for long-term prediction of wave fields. The method constructs a model input tensor using historical sample data of wave fields in a target region; it then uses a trained long-term prediction model to predict the wave field prediction data for the current prediction step based on the model input tensor; employing a multi-step autoregressive rolling mechanism, it updates the model input tensor based on the wave field prediction data, and determines the wave field prediction data for each prediction step using the updated model input tensor through the trained long-term prediction model; finally, it determines the short-term, medium-term, and long-term prediction sequences of the wave field based on the wave field prediction data for each prediction step through the prediction span output heads of the trained long-term prediction model, and fuses them to determine the long-term wave field prediction result. Therefore, this method effectively solves the problem that existing technologies cannot improve long-term rolling stability while reducing modeling complexity and taking into account the propagation characteristics at different time scales.

[0086] It should be understood that the application of the present invention is not limited to the examples above. 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 protection scope of the appended claims.

Claims

1. A long-term prediction method for wave fields, characterized in that, The method includes: The model input tensor is constructed based on historical sample data of the wave field in the target region. Based on the input tensor of the trained long-term prediction model, the fluctuation field prediction data corresponding to the current prediction step is determined. A multi-step autoregressive rolling mechanism is adopted to update the model input tensor based on the fluctuation field prediction data, and the fluctuation field prediction data for each prediction step is determined by the trained long-term prediction model based on the updated model input tensor. Based on the wave field prediction data of each prediction step, the short-term prediction sequence, medium-term prediction sequence, and long-term prediction sequence of the wave field are determined by the prediction span output heads of each prediction span of the trained long-term prediction model. The short-term prediction sequence, the medium-term prediction sequence, and the long-term prediction sequence are fused using the fusion module of the trained long-term prediction model to determine the long-term fluctuation field prediction result.

2. The long-term prediction method for wave fields according to claim 1, characterized in that, The model input tensor is constructed based on historical sample data of the wave field in the target region, including: The historical sample data of the wave field is divided based on the input window and the output window to determine each input time series and the corresponding output time series; The input time series and the corresponding output time series are downsampled according to the prediction time interval to determine each supervised learning sample pair; The supervised learning sample pairs are subjected to spatial grid organization, channel rearrangement and normalization to determine the model input tensor.

3. The long-term prediction method for wave fields according to claim 1, characterized in that, The long-term prediction model further includes: an input mapping layer, a position encoding layer, a backbone network, and an output decoding layer. Based on the input tensor of the trained long-term prediction model, it determines the wave field prediction data corresponding to the current prediction step, including: The input mapping layer maps the model input tensor into high-dimensional implicit features to obtain the initial latent space features. The initial latent space features are encoded using Gaussian Fourier features through the location encoding layer to determine the latent space encoded features; The backbone network is used to extract features based on the latent space coding features to determine the coupled evolution features. The coupled evolution features are decoded by the output decoding layer to determine the wave field prediction data corresponding to the current prediction step.

4. The long-term prediction method for wave fields according to claim 3, characterized in that, The backbone network decomposes the joint spatial attention computation into directional low-rank kernel attention along different directions, and uses a shared parameter multiple iterative update method to recursively update the latent space encoding features to determine the coupled evolution features.

5. The long-term prediction method for wave fields according to claim 1, characterized in that, A multi-step autoregressive rolling mechanism is employed to update the model input tensor based on the fluctuation field prediction data, including: Remove the first frame of the model input tensor in the current prediction step and add the fluctuation field prediction data as the last frame of the model input tensor to determine the updated model input tensor.

6. The long-term prediction method for wave fields according to claim 1, characterized in that, The training method for the long-term prediction model includes: Perform single-step prediction training on the long-term prediction model; Based on the long-term prediction model trained by single-step prediction, perform multi-step autoregressive training for a first preset number of time steps, and update the parameters of the long-term prediction model trained by single-step prediction based on the multi-step prediction loss. The long-term prediction model trained by multi-step autoregression is fine-tuned for a second preset number of time steps to determine the trained long-term prediction model.

7. The long-term prediction method for wave fields according to claim 6, characterized in that, The method further includes: When performing multi-step autoregressive expansion training on a long-term prediction model trained on a single-step prediction basis, random noise is added to the input window.

8. A long-term wave field prediction device, characterized in that, The device includes: The data acquisition module is used to construct the model input tensor based on historical sample data of the fluctuation field in the target area; The single-step prediction module is used to determine the fluctuation field prediction data corresponding to the current prediction step based on the input tensor of the trained long-term prediction model. The autoregressive rolling prediction module is used to update the model input tensor based on the fluctuation field prediction data using a multi-step autoregressive rolling mechanism, and to determine the fluctuation field prediction data for each prediction step based on the updated model input tensor using a trained long-term prediction model. The sequence output module is used to determine the short-term, medium-term, and long-term prediction sequences of the fluctuation field based on the fluctuation field prediction data of each prediction step, using the prediction span output heads of the trained long-term prediction model. The multi-scale fusion module is used to fuse the short-term prediction sequence, the medium-term prediction sequence, and the long-term prediction sequence through the fusion module of the trained long-term prediction model to determine the long-term fluctuation field prediction result.

9. A terminal, characterized in that, The terminal includes a memory and one or more processors; the memory stores one or more programs; the programs contain instructions for executing the long-term prediction method for wave fields as described in any one of claims 1-7; the processors are used to execute the programs.

10. A computer-readable storage medium storing a plurality of instructions thereon, characterized in that, The instructions are loaded and executed by the processor to implement the steps of the long-term prediction method for wave fields as described in any one of claims 1-7.

Images