Three-dimensional ocean dynamic environment intelligent forecasting method and device and electronic equipment

Abstract

The invention relates to the technical field of ocean dynamic environment intelligent forecasting, and provides a three-dimensional ocean dynamic environment intelligent forecasting method and device and electronic equipment, and the forecasting method comprises the steps: S1000, preparing 1 / 30-degree high-resolution ocean dynamic environment element data (three-dimensional temperature and salt flow and two-dimensional water level), and carrying out the normalization; s2000, selecting a corresponding model input sequence from the training set according to a forecasting strategy of the model; s3000, training a hierarchical feature projection network; and S4000: performing reverse normalization and generating future prediction data: performing reverse normalization on the feature-enhanced output data according to the maximum value and the minimum value of the ocean data, and finally generating a day-by-day prediction product of the sea surface height of the C sea, the three-dimensional sea temperature, the salinity, the latitudinal flow velocity and the longitude flow velocity with 1 / 30-degree high resolution in the future day, namely, the product is obtained by predicting the sea surface height of the C sea, the three-dimensional sea temperature, the salinity, the latitudinal flow velocity and the longitude flow velocity.

Description

Technical Field

[0001] This invention relates to the field of intelligent forecasting technology for marine dynamic environment, specifically to a three-dimensional intelligent forecasting method, device, and electronic equipment for marine dynamic environment. Background Technology

[0002] Current forecasting technologies for three-dimensional ocean temperature, salinity, currents, and water levels are core supporting technologies for marine environmental monitoring, disaster early warning, navigation safety, and marine resource development. These technologies are currently divided into two main categories: traditional numerical simulation methods and emerging data-driven intelligent forecasting methods. Both types of methods are developing towards higher resolution, higher accuracy, and longer forecasting time, and have been experimentally validated in scenarios such as global-scale ocean forecasting and regional-scale disaster early warning. However, their applicable scenarios differ significantly, and their technical characteristics have different focuses, such as:

[0003] 1. Traditional numerical simulation methods:

[0004] This type of method is currently the mainstream technology in the field of ocean forecasting. Specifically, it involves constructing complex ocean dynamic equations (such as the Navier-Stokes equations and heat conduction equations) based on ocean dynamics theory, and combining them with initial field data and given boundary conditions to achieve dynamic simulation and forecasting of three-dimensional temperature, salinity, current and water level.

[0005] However, this method still has three major flaws:

[0006] First, due to its reliance on complex physical equations and parameter settings, the computational cost is extremely high. Even with optimizations such as modal separation, it is still difficult to meet the real-time requirements of scenarios such as nearshore disaster early warning.

[0007] Secondly, the global model has low resolution, while regional models are mostly nested within the global model, resulting in insufficient capture of the detailed three-dimensional structure of the nearshore area, leading to significant nearshore forecast errors.

[0008] Third, it is highly sensitive to the initial field, and the deviation is easily amplified through the accumulation of dynamic equations. In addition, the parameterization scheme has inherent biases, resulting in high forecast uncertainty.

[0009] 2. Data-driven intelligent forecasting methods:

[0010] With the rapid development of artificial intelligence, this type of method has become a research hotspot in the field of marine forecasting. Its essence lies in integrating deep learning algorithms, with convolutional neural networks at its core. By mining the spatiotemporal correlation information hidden in massive amounts of historical marine reanalysis data and observational data, it constructs data-driven models to achieve rapid and intelligent forecasting of core marine variables such as temperature, salinity, current, and water level. Currently, this type of method is gradually transitioning from the theoretical research stage to operational applications. Leveraging its core advantage of rapid response, it primarily serves high-real-time scenarios such as short-term marine environmental early warning and maritime emergency support.

[0011] However, existing data-driven intelligent forecasting methods are insufficient in characterizing the synergistic coupling relationship among multiple variables such as temperature, salinity, and current, resulting in poor accuracy of multivariate joint forecasts; and they are difficult to adapt to the actual needs of regional high-resolution and refined forecasting. Summary of the Invention

[0012] This invention proposes a three-dimensional intelligent forecasting method for marine dynamic environment to solve the above-mentioned problems.

[0013] The technical solution of the present invention is as follows:

[0014] A three-dimensional intelligent forecasting method for marine dynamic environment includes the following steps:

[0015] S1000: Prepare 1 / 30° high-resolution marine dynamic environment element data and normalize it;

[0016] S2000: Prediction strategy based on the model Select the corresponding model input sequence from the training set. ;

[0017] S3000: Training a hierarchical feature projection network;

[0018] S4000: Inverse normalization and generation of future prediction data:

[0019] Based on the maximum and minimum values ​​of marine dynamic environmental element data in the training set, the feature-enhanced output data is denormalized, and finally, the future data is generated. Daily forecasts of sea surface height and three-dimensional sea temperature, salinity, zonal current velocity, and meridional current velocity in the C Sea at a high resolution of 1 / 30°. .

[0020] Furthermore, step S1000 includes:

[0021] S1100: Assume the three-dimensional sea surface temperature data is... 3D salinity data Three-dimensional latitudinal velocity data are 3D meridional velocity data are and two-dimensional water level data ,in The training set consists of daily average ocean data samples;

[0022] S1200: The data ranges of the above five marine dynamic environmental elements are normalized to the range [0, 1] using the maximum-minimum normalization method. The specific calculation formula is as follows:

[0023] ;

[0024] in, , These represent the minimum and maximum values ​​in the current dynamic environmental data, respectively.

[0025] After normalizing the five marine dynamic environmental elements, they were concatenated to form the model input data, specifically represented as follows:

[0026] .

[0027] Furthermore, the training process of the hierarchical feature projection network in step S3000 includes: the core feature modeling module of the U-Transformer architecture, which includes a data preprocessing module, a two-layer convolution module, a feature enhancement module, and a latent space attention module;

[0028] A hierarchical collaborative mechanism is formed by a dual-layer convolution module for local fine-grained feature extraction, a Transformer module for capturing global correlation information, and a latent feature attention fusion mechanism.

[0029] Furthermore, the training process of the hierarchical feature projection network in step S3000 is as follows:

[0030] S3100:

[0031] The normalized three-dimensional ocean temperature, salinity, current, and water level data are formatted as [batchsize, height, width, depth]. 4+1, Construct the corresponding feature tensor;

[0032] Where batchsize is the batch size, height and width are the total number of grid points in the vertical and horizontal directions, respectively, and depth represents the depth layer of the four three-dimensional marine dynamic environmental elements (temperature, salinity, zonal current velocity and meridional current velocity).

[0033] The normalized features of three-dimensional ocean temperature, salinity, current and water level are divided into several training batches, which are then fed into the network in batches and iterative training is completed.

[0034] S3200: After the data preprocessing module performs three-dimensional coordinate position embedding, linear transformation, tensor dimension adjustment and filling operations on the feature tensor, the feature tensor is sequentially sent to three different scales of "local fine extraction-global association capture-latent feature mapping" fusion mechanism. Each scale fusion mechanism models ocean features with different spatial resolutions.

[0035] Furthermore, the fusion mechanism at each scale in step S3200 follows the logic of "local feature extraction - global association capture - latent feature enhancement," and the specific steps are as follows:

[0036] S3210: Local fine feature extraction: The extraction of local dynamic features of the ocean at the current scale is completed through a two-layer convolution module. This module adopts a serial structure of "three-dimensional convolutional layer → batch normalization → corrected linear unit" stacked twice in succession.

[0037] S3220: Global correlation information capture: Perform block operation on the local feature tensor output by the two-layer convolution module to generate several feature blocks, and complete the dimensional adaptation with a learnable token feature element Z through a three-dimensional convolution layer.

[0038] The dimension-aligned feature tensor and the token feature element are fed into the self-attention mechanism module.

[0039] S3230: Perform latent feature mapping learning;

[0040] S3240: Perform cross-scale feature fusion.

[0041] Furthermore, S3220 includes:

[0042] S3221: The association score between the feature block and the token feature element is calculated by matrix multiplication. After scaling and Softmax activation function processing, a normalized attention weight matrix is ​​generated.

[0043] S3222: The weighted aggregation of the weight matrix and the original feature block is completed by matrix multiplication;

[0044] S3223: Finally, the aggregated feature tensor and the Token feature element are performed with residual connection. While preserving the original feature information, the deep fusion of global correlation features and Token features is achieved, thereby completing the global long-distance spatiotemporal correlation capture of ocean features at this scale.

[0045] Furthermore, step S3230 includes:

[0046] S3231: Using a combination function of sine and cosine basis operators, a spatiotemporal location coding feature subfunction adapted to the current scale is generated, which maps the three-dimensional ocean feature tensor into a periodic spatiotemporal location coding feature, thus completing the global long-distance spatiotemporal correlation capture of ocean features at this scale.

[0047] S3232: Next, the transformed position coding feature sub-function is multiplied element-wise with a learnable weight matrix W to complete the adaptive modulation of the position features;

[0048] S3233: Then, the modulated sub-features are residually connected to the original feature tensor of the branch, and finally, the latent feature mapping and enhancement with spatiotemporal location are realized while preserving local and global features.

[0049] S3234: Finally, the spatiotemporally enhanced ocean features and the original features aligned with the current scale dimension are fed into the second self-attention mechanism module. At the same time, the enhanced ocean feature tensor with spatiotemporal location awareness is output through residual connection, providing accurate latent feature support for the subsequent multi-scale fusion mechanism.

[0050] Furthermore, step S3240 includes:

[0051] S3241: First, perform a 2x upsampling operation on the small-scale marine latent features, and then perform feature splicing and fusion with the medium-scale latent features in the channel dimension;

[0052] S3242: Next, the fused feature tensor is fed into the two-layer convolution module for secondary local feature refinement extraction.

[0053] S3243: Then, perform a 2x upsampling operation on the local feature tensor of the small- and medium-scale fused features, and similarly perform a tensor concatenation operation with the large-scale latent features, and feed it into the two-layer convolution module to finally obtain the cross-scale fused feature tensor.

[0054] S3244: Feed the fused feature tensor into the feature enhancement module:

[0055] First, adjust its dimensions to precisely match and align with the input dimensions of the data preprocessing module using Conv3d, clipping, and replacement operations;

[0056] Then, a linear transformation layer, a modified linear unit activation layer, and a quadratic linear transformation layer are sequentially connected. Through the synergistic effect of linear mapping and nonlinear activation, the ability to represent fused features is deeply enhanced.

[0057] Another objective of this invention is to provide a three-dimensional intelligent forecasting device for marine dynamic environment, which executes the above-mentioned forecasting method during operation.

[0058] Another objective of this invention is to provide a three-dimensional intelligent forecasting electronic device for marine dynamic environment, which executes the aforementioned forecasting method during operation.

[0059] The working principle and beneficial effects of this invention are as follows:

[0060] (1) A hierarchical feature projection network is constructed based on the U-Transformer architecture. Through the fusion mechanism of "local fine extraction - global association capture - latent feature mapping", the core contradiction that a single architecture cannot effectively take into account both local fine dynamic processes and global long-distance spatiotemporal associations is solved.

[0061] (2) It can quickly generate three-dimensional temperature, salinity, current and water level daily forecast products with a spatial resolution of up to 1 / 30° for several days in the future, which solves the problems of insufficient accuracy and weak timeliness of traditional forecast models in regional refined element forecasts.

[0062] (3) The three-dimensional marine dynamic environment intelligent forecasting model based on hierarchical feature projection network proposed in this invention has better overall results than existing numerical model simulation results and reanalysis data (such as system A, system B, etc.), proving the effectiveness and feasibility of hierarchical feature projection network in intelligent forecasting tasks of three-dimensional marine dynamic environment (such as temperature, salinity and current). Attached Figure Description

[0063] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0064] Figure 1 This is a flowchart illustrating the overall training process of the present invention.

[0065] Figure 2 The box plot of RMSE between the prediction results (sea surface temperature) of the model on day 1 of this invention and other prediction results;

[0066] Figure 3 This is a box plot of the RMSE of the prediction results (sea level anomaly) of the model on day 1 of this invention and other prediction results;

[0067] Figure 4 This is a box plot of the RMSE of the prediction results (sea surface temperature) on day 1 of the model of this invention with other prediction results;

[0068] Figure 5 This is a box plot of the RMSE of the salinity prediction result (salinity) on day 1 of the model of this invention with other prediction results;

[0069] Figure 6 A comparison of the RMSE of this invention and other prediction results on the vertical profile (sea surface temperature) at different depths;

[0070] Figure 7 A comparison of the RMSE of this invention and other prediction results on the vertical profile (salinity) at different depths;

[0071] Figure 8 This is a box plot of the RMSE of the 7-day prediction results (sea surface temperature) for the whole year of 2021 and other prediction results.

[0072] Figure 9 This is a box plot of the RMSE of the 7-day prediction results (sea level anomaly) for the whole year of 2021 and other prediction results of this invention.

[0073] Figure 10 This is a box plot of the average RMSE of the 7-day sea surface temperature forecast for the whole year of 2021 and other forecast results. Figure 11 This is a box plot of the average RMSE of the 7-day salinity prediction results for the whole year of 2021 and other prediction results.

[0074] Figure 12 Vertical profile curves showing the average root mean square error of temperature forecast for 7 days throughout 2021 under different depths according to this invention;

[0075] Figure 13 Vertical profile curves showing the average root mean square error of salinity for the 7-day forecast of the whole year of 2021 at different depths according to this invention.

[0076] Figure 14 The vertical profile curves of the average root mean square error of the zonal current velocity for the 7-day forecast of the whole year of 2021 under different depths are shown.

[0077] Figure 15 Vertical profile curves showing the average root mean square error of meridional current velocity for the 7-day forecast of the entire year of 2021 at different depths;

[0078] Figure 16 This is a diagram of the overall structure of the hierarchical feature projection network. Detailed Implementation

[0079] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0080] Example 1

[0081] Compared to common three-dimensional marine environmental element prediction models, the problem this invention aims to solve is that existing models cannot simultaneously capture both detailed local dynamic processes and long-range global spatiotemporal correlations, making it difficult to adapt to the complex nature of multi-scale, nonlinear dynamic evolution of marine data. Specifically, this includes:

[0082] (1) The single model architecture has inherent defects. If the focus is on modeling local features, the inherent laws of ocean time series data are easily overlooked; if the focus is on simulating global features, local details will be lost.

[0083] (2) Due to the weak multi-scale feature coupling modeling capability, it is impossible to accurately explore the nonlinear characteristics of ocean processes at different scales, and thus it is difficult to output high-resolution, long-term three-dimensional marine environmental element forecast products.

[0084] Therefore, this embodiment provides a three-dimensional intelligent forecasting method for marine dynamic environment, such as... Figures 1 to 16 As shown, it includes the following steps:

[0085] S1000: Prepare 1 / 30° high-resolution marine dynamic environmental data (three-dimensional temperature, salinity, and current, and two-dimensional water level), and normalize them;

[0086] Step S1000 includes:

[0087] S1100: Assume the three-dimensional sea surface temperature data is... 3D salinity data Three-dimensional latitudinal velocity data are 3D meridional velocity data are and two-dimensional water level data ,in The training set consists of daily average ocean data samples;

[0088] Temperature, salinity, and current data are all in three-dimensional format, i.e., [height, width, depth]; while water level data are in two-dimensional format, i.e., [height, width, 1], where height and width are the vertical and horizontal resolutions, respectively, and depth represents the depth layer of each feature.

[0089] S1200: The data ranges of the above five marine dynamic environmental elements are normalized to the range [0, 1] using the maximum-minimum normalization method. The specific calculation formula is as follows:

[0090] ;

[0091] in, , These represent the minimum and maximum values ​​in the current dynamic environmental data, respectively.

[0092] After normalization, the five marine dynamic environmental elements are spliced ​​together to form the model input data, specifically represented as follows: .

[0093] The data format is [height, width, depth] 4+1];

[0094] Among them, 4 represents four types of ocean dynamic elements: sea temperature, salinity, zonal current velocity, and meridional current velocity, and 1 represents water level element, which together constitute the characteristic dimension of a single spatiotemporal point.

[0095] S2000: Prediction strategy based on the model Select the corresponding model input sequence from the training set. ;

[0096] The format is [height, width, depth] 4+1 ];

[0097] S3000: Training a hierarchical feature projection network;

[0098] The training process of the above-mentioned hierarchical feature projection network includes: the core feature modeling module of the U-Transformer architecture, which includes a data preprocessing module, a double convolution module, a feature enhancement module (FEM), and a latent attention block.

[0099] A hierarchical collaborative mechanism is formed by a dual-layer convolution module for local fine-grained feature extraction, a Transformer (deep learning model architecture) module for capturing global correlation information, and a latent feature attention fusion mechanism.

[0100] The training process of the hierarchical feature projection network in step S3000 is as follows:

[0101] S3100:

[0102] The normalized three-dimensional ocean temperature, salinity, current, and water level data are formatted as [batchsize, height, width, depth]. 4+1, Construct the corresponding feature tensor;

[0103] Where batchsize is the batch size, height and width are the total number of grid points in the vertical and horizontal directions, respectively, and depth represents the depth layer of the four three-dimensional marine dynamic environmental elements (temperature, salinity, zonal current velocity and meridional current velocity).

[0104] The normalized features of three-dimensional ocean temperature, salinity, current, and water level are divided into several training batches, which are then fed into the network in batches for iterative training. ;

[0105] S3200: After the data preprocessing module performs three-dimensional coordinate position embedding, linear transformation, tensor dimension adjustment and filling operations on the feature tensor, the feature tensor is sequentially sent to three different scales of "local fine extraction-global correlation capture-latent feature mapping" fusion mechanism. Each scale fusion mechanism models ocean features (such as large-scale circulation, meso-scale eddies and small-scale nearshore disturbances) with different spatial resolutions.

[0106] The fusion mechanism at each scale in step S3200 follows the logic of "local feature extraction - global association capture - latent feature enhancement", and the specific steps are as follows:

[0107] S3210: Local fine feature extraction: The extraction of local dynamic features of the ocean at the current scale is completed through a two-layer convolution module. This module adopts a serial structure of "three-dimensional convolutional layer (Conv3d) → batch normalization (BatchNorm) → corrected linear unit (ReLU)" stacked twice in succession.

[0108] S3220: Global correlation information capture: Perform block operation on the local feature tensor output by the two-layer convolution module to generate several feature blocks, and complete the dimensional adaptation with a learnable token feature element Z through a three-dimensional convolution layer.

[0109] The dimension-aligned feature tensor and the token feature element are fed into the self-attention mechanism module.

[0110] S3220 includes:

[0111] S3221: The association score between the feature block and the token feature element is calculated by matrix multiplication. After scaling and softmax activation, a normalized attention weight matrix is ​​generated.

[0112] S3222: The weighted aggregation of the weight matrix and the original feature block is completed by matrix multiplication;

[0113] S3223: Finally, the aggregated feature tensor and the Token feature element are performed with residual connection. While preserving the original feature information, the deep fusion of global correlation features and Token features is achieved, thereby completing the global long-distance spatiotemporal correlation capture of ocean features at this scale.

[0114] S3230: Perform latent feature mapping learning;

[0115] Step S3230 includes:

[0116] S3231: Using a combination function of sine-cosine basis operators (SinCos module), a spatiotemporal location coding feature subfunction adapted to the current scale is generated, which maps the three-dimensional ocean feature tensor into periodic spatiotemporal location coding features, thus completing the global long-distance spatiotemporal correlation capture of ocean features at this scale.

[0117] Since the self-attention mechanism module cannot effectively identify the spatial and temporal location of ocean data features, such as the differences in ocean dynamic features between nearshore / offshore and between upper and lower thermoclines, after processing by the SinCos module, each ocean spatiotemporal location generates a corresponding feature vector, thereby enabling accurate mining of ocean feature differences at different latitudes and longitudes, vertical layers, and time steps.

[0118] S3232: Next, the transformed position coding feature sub-function is multiplied element-wise with a learnable weight matrix W to complete the adaptive modulation of the position features;

[0119] S3233: Then, the modulated sub-feature is compared with this branch ( Figure 16 middle The original feature tensors of (location) are subjected to residual connections, and finally, while preserving local and global features, latent feature mapping and enhancement with spatiotemporal location are achieved.

[0120] S3234: Finally, the spatiotemporally enhanced ocean features and the original features aligned with the current scale dimension are fed into the second self-attention mechanism module. At the same time, the enhanced ocean feature tensor with spatiotemporal location awareness is output through residual connection, providing accurate latent feature support for the subsequent multi-scale fusion mechanism.

[0121] S3240: Perform cross-scale feature fusion.

[0122] The mechanism employs a fusion of three different scales—"local fine-grained extraction - global correlation capture - latent feature mapping"—to generate enhanced ocean latent feature tensors for the corresponding scales. These three scales correspond to small-scale nearshore disturbances, mesoscale eddies, and large-scale circulation features, respectively. There are certain correlations between the features at each scale regarding ocean dynamic processes; that is, large-scale circulation provides background field constraints for mesoscale eddies, while mesoscale eddies cause physical laws governing small-scale nearshore disturbances. Step S3240 includes:

[0123] S3241: First, perform a 2x upsampling operation on the small-scale marine latent features, and then perform feature splicing and fusion with the medium-scale latent features in the channel dimension;

[0124] S3242: Next, the fused feature tensor is fed into the two-layer convolution module for secondary local feature refinement extraction.

[0125] S3243: Then, perform a 2x upsampling operation on the local feature tensor of the small- and medium-scale fused features, and similarly perform a tensor concatenation operation with the large-scale latent features, and feed it into the two-layer convolution module to finally obtain the cross-scale fused feature tensor.

[0126] Upsampling is performed layer by layer to restore the spatial resolution of the feature tensor, i.e., a high-resolution ocean grid scale consistent with the original input, compensating for the spatial resolution loss caused by feature extraction during multi-scale modeling. Tensor concatenation achieves the initial aggregation of multi-scale enhanced ocean features, allowing the fused feature tensor to simultaneously contain large-scale global circulation features, meso-scale eddy evolution features, and small-scale refined features such as nearshore upwelling, forming multi-scale fused ocean features. The fused features are then fed a second time into a two-layer convolutional module to eliminate the feature ambiguity caused by upsampling and to enhance the local detail representation of the cross-scale fused features.

[0127] After completing the cross-scale fusion of multi-scale ocean features, the resulting fused feature tensor integrates hierarchical ocean dynamic information from large-scale circulation, mesoscale eddies, and small-scale nearshore disturbances.

[0128] S3244: Feed the fused feature tensor into the Feature Enhancement Module (FEM):

[0129] First, adjust its dimensions to precisely match and align with the input dimensions of the data preprocessing module using Conv3d, clipping, and replacement operations;

[0130] Then, a linear transformation layer, a modified linear unit activation layer, and a quadratic linear transformation layer are sequentially connected. Through the synergistic effect of linear mapping and nonlinear activation, the ability to represent fused features is deeply enhanced.

[0131] S4000: Inverse normalization and generation of future prediction data:

[0132] Based on the maximum and minimum values ​​of the centralized ocean dynamic data in S1200, the feature-enhanced output data is denormalized, and finally, the future... Daily forecasts of sea surface height and three-dimensional sea temperature, salinity, zonal current velocity, and meridional current velocity in the C Sea at a high resolution of 1 / 30°. .

[0133] Specifically, the pseudocode steps of this embodiment are as follows:

[0134]

[0135] Furthermore, the calculation formulas for the sine and cosine basis operators in the above combined function (SinCos module) are as follows:

[0136] ; ;

[0137] in, These are the initial global association features at three scales, while This refers to the number of sine or cosine basis operators selected. The results of each basis operator function are concatenated to form a mapping feature tensor at the corresponding scale:

[0138] ;

[0139] Therefore, after strengthening the latent features with spatiotemporal location information, we can obtain the enhanced feature tensor. The calculation formula is as follows:

[0140] ;

[0141] in, This represents the weight tensor to be learned during the training process of the hierarchical feature projection network, corresponding to each sine or cosine basis operator function.

[0142] Example 2

[0143] Based on Embodiment 1, a three-dimensional intelligent marine dynamic environment forecasting device executes the forecasting method in Embodiment 1 during operation.

[0144] Example 3

[0145] Based on Embodiment 1, a three-dimensional intelligent forecasting electronic device for marine dynamic environment executes the forecasting method in Embodiment 1 during operation.

[0146] Furthermore, the above-described implementation methods can be verified using the following evaluation metrics:

[0147] Unnormalized The original truth data at time point is represented as ,in This is a daily average ocean data sample for the test set. and They represent The first moment The Primitive Heaven True value data of a marine dynamic environmental element and prediction results of a certain model, where 1≤ ≤ ,1≤ ≤5. And and These correspond to the grid points. The value at position, where 1≤ ≤ ,1≤ ≤ , and These represent the number of grid points along the latitude and longitude directions, respectively.

[0148] 1. Root Mean Square Error (RMSE):

[0149] ;

[0150] in, This represents the total number of samples in the test set.

[0151] 2. Peak Signal-to-Noise Ratio (PSNR):

[0152] ;

[0153] in, This represents the theoretical maximum value of five marine dynamic environmental factors (i.e., sea temperature, salinity, zonal and meridional flow velocity, and water level).

[0154] 3. Natural Image Quality Evaluator (NIQE):

[0155] ;

[0156] in, and These are the mean vector and covariance matrix of the image feature library, respectively; Indicates the prediction result The mean vector.

[0157] Using measured data from sea area C as comparative data, the following comparison table of predicted sea surface temperatures is obtained (in the table, "↓" indicates that the smaller the value, the better, and "↑" indicates that the larger the value, the better):

[0158]

[0159] The above-described operating methods can be arbitrarily adjusted according to the design, provided that the hardware module supports them; this embodiment will not elaborate further. The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-dimensional intelligent forecasting method for marine dynamic environment, characterized in that, Includes the following steps: S1000: Prepare 1 / 30° high-resolution marine dynamic environment element data and normalize it; S2000: Prediction strategy based on the model Select the corresponding model input sequence from the training set. ; S3000: Training a hierarchical feature projection network; S4000: Inverse normalization and generation of future prediction data: Based on the maximum and minimum values ​​of the marine dynamic environment data in the training set, the feature-enhanced output data is denormalized, and finally, the future... Daily forecasts of sea surface height and three-dimensional sea temperature, salinity, zonal current velocity, and meridional current velocity in the C Sea at a high resolution of 1 / 30°. .

2. The intelligent forecasting method for three-dimensional marine dynamic environment according to claim 1, characterized in that, Step S1000 includes: S1100: Assume the three-dimensional sea surface temperature data is... 3D salinity data 3D latitudinal velocity data are 3D meridional velocity data are and two-dimensional water level data ,in Daily average ocean data samples for the training set ; S1200: The data ranges of the above five marine dynamic environmental elements are normalized to the range [0, 1] using the maximum-minimum normalization method. The specific calculation formula is as follows: ； in, , These represent the minimum and maximum values ​​in the current dynamic environmental data, respectively. After normalizing the five marine dynamic environmental elements, they were concatenated to form the model input data, specifically represented as follows: 。 3. The intelligent forecasting method for three-dimensional marine dynamic environment according to claim 2, characterized in that, The training process of the hierarchical feature projection network in step S3000 includes: the core feature modeling module of the U-Transformer architecture, which includes a data preprocessing module, a two-layer convolution module, a feature enhancement module, and a latent space attention module; A hierarchical collaborative mechanism is formed by a dual-layer convolution module for local fine-grained feature extraction, a Transformer module for capturing global correlation information, and a latent feature attention fusion mechanism.

4. The intelligent forecasting method for three-dimensional marine dynamic environment according to claim 3, characterized in that, The training process of the hierarchical feature projection network in step S3000 is as follows: S3100: For normalized three-dimensional ocean temperature, salinity, current, and water level data, formatted as [batchsize, height, width, depth] 4+1, Construct the corresponding feature tensor; Where batchsize is the batch size, height and width are the total number of grid points in the vertical and horizontal directions, respectively, and depth represents the depth layer of the four three-dimensional ocean dynamic environmental elements (temperature, salinity, zonal current velocity and meridional current velocity); that is, the normalized features of three-dimensional ocean temperature, salinity and current and water level are tensorized into several training batches, fed into the network in batches, and iterative training is completed. S3200: After the data preprocessing module performs three-dimensional coordinate position embedding, linear transformation, tensor dimension adjustment and filling operations on the feature tensor, the feature tensor is sequentially sent to three different scales of "local fine extraction-global association capture-latent feature mapping" fusion mechanism. Each scale fusion mechanism models ocean features with different spatial resolutions.

5. The intelligent forecasting method for three-dimensional marine dynamic environment according to claim 4, characterized in that, The fusion mechanism at each scale in step S3200 follows the logic of "local feature extraction - global association capture - latent feature enhancement", and the specific steps are as follows: S3210: Local fine feature extraction: The extraction of local dynamic features of the ocean at the current scale is completed through a two-layer convolution module. This module adopts a serial structure of "three-dimensional convolutional layer → batch normalization → corrected linear unit" stacked twice in succession. S3220: Global correlation information capture: Perform block operation on the local feature tensor output by the two-layer convolution module to generate several feature blocks, and complete the dimensional adaptation with a learnable token feature element Z through a three-dimensional convolution layer. The dimension-aligned feature tensor and the token feature element are fed into the self-attention mechanism module. S3230: Perform latent feature mapping learning; S3240: Perform cross-scale feature fusion.

6. The intelligent forecasting method for three-dimensional marine dynamic environment according to claim 5, characterized in that, S3220 includes: S3221: The association score between the feature block and the token feature element is calculated by matrix multiplication. After scaling and Softmax activation function processing, a normalized attention weight matrix is ​​generated. S3222: The weighted aggregation of the weight matrix and the original feature block is completed by matrix multiplication; S3223: Finally, the aggregated feature tensor and the Token feature element are performed with residual connection. While preserving the original feature information, the deep fusion of global correlation features and Token features is achieved, thereby completing the global long-distance spatiotemporal correlation capture of ocean features at this scale.

7. The intelligent forecasting method for three-dimensional marine dynamic environment according to claim 6, characterized in that, Step S3230 includes: S3231: Using a combination function of sine and cosine basis operators, a spatiotemporal location coding feature subfunction adapted to the current scale is generated, which maps the three-dimensional ocean feature tensor into a periodic spatiotemporal location coding feature, thus completing the global long-distance spatiotemporal correlation capture of ocean features at this scale. S3232: Next, the transformed position coding feature sub-function is multiplied element-wise with a learnable weight matrix W to complete the adaptive modulation of the position features; S3233: Then, the modulated sub-features are joined with the original feature tensors of the branches using residual connections. Finally, while preserving local and global features, latent feature mapping and enhancement with spatiotemporal location are achieved. S3234: Finally, the spatiotemporally enhanced ocean features and the original features aligned with the current scale dimension are fed into the second self-attention mechanism module. At the same time, the enhanced ocean feature tensor with spatiotemporal location awareness is output through residual connection, providing accurate latent feature support for the subsequent multi-scale fusion mechanism.

8. The intelligent forecasting method for three-dimensional marine dynamic environment according to claim 7, characterized in that, Step S3240 includes: S3241: First, perform a 2x upsampling operation on the small-scale marine latent features, and then perform feature splicing and fusion with the medium-scale latent features in the channel dimension; S3242: Next, the fused feature tensor is fed into the two-layer convolution module for secondary local feature refinement extraction. S3243: Then, perform a 2x upsampling operation on the local feature tensor of the small- and medium-scale fused features, and similarly perform a tensor concatenation operation with the large-scale latent features, and feed it into the two-layer convolution module to finally obtain the cross-scale fused feature tensor. S3244: Feed the fused feature tensor into the feature enhancement module: First, adjust its dimensions to precisely match and align with the input dimensions of the data preprocessing module using Conv3d, clipping, and replacement operations; Then, a linear transformation layer, a modified linear unit activation layer, and a quadratic linear transformation layer are sequentially connected. Through the synergistic effect of linear mapping and nonlinear activation, the ability to represent fused features is deeply enhanced.

9. A three-dimensional intelligent forecasting device for marine dynamic environment, characterized in that, During runtime, the forecasting method described in any one of claims 1-8 is executed.

10. A three-dimensional intelligent forecasting electronic device for marine dynamic environment, characterized in that, During runtime, the forecasting method described in any one of claims 1-8 is executed.

Images