An intelligent empirical correction method and system for ocean current numerical prediction based on a multimodal neural network

Abstract

The present invention discloses an intelligent empirical correction method and system for ocean current numerical prediction based on a multimodal neural network, which relates to the cross technical field of artificial intelligence and ocean prediction. The method steps include: collecting the observation data of the ocean current area to be predicted and constructing a label data set; preprocessing the label data set to construct a feature data set; constructing a correction model by using a multimodal neural network based on the label data set and the feature data set; and completing the correction of ocean current prediction based on the correction model. The present invention has the ability to adaptively capture the laws between multi-source heterogeneous data, improves the ocean current correction effect, thereby improving the accuracy of ocean current prediction, and provides more reliable prediction data for the development of the ocean economy, ocean research, and climate prediction.

Description

Technical Field

[0001] The present invention relates to the cross-technical field of artificial intelligence and ocean forecasting, and in particular to an intelligent empirical correction method and system for ocean current numerical forecasting based on a multimodal neural network. Background Art

[0002] Ocean current is a key dynamic process in the ocean. Accurate prediction of ocean current is one of the prerequisites for ensuring the reasonable selection of engineering parameters, navigation safety and rapid search and rescue of drowned targets in the ocean. At present, ocean current forecasting is mainly achieved through numerical models, which are technologies based on physical laws and numerical analysis. However, due to the complexity of the ocean system itself, humans have not yet fully understood the physical laws of the ocean, which makes the existing ocean current numerical forecasting still have large forecast errors and is difficult to meet the actual needs of ocean engineering design, drowned target search and rescue. To address this problem, empirical correction of the forecast results of the numerical model is an effective solution. The empirical correction method is a method that integrates historical observation data and numerical model forecast results to construct an empirical correction model. The correction model used by the traditional empirical correction method is relatively simple, which is insufficient in mapping ability and generalization ability. In addition, the input data type and structure of the correction model of these correction methods are single, and lack the ability to capture the laws between multi-source heterogeneous data.

[0003] In order to overcome the problems of insufficient mapping and generalization capabilities of traditional empirical correction models and the single type and structure of input data, improve the accuracy of numerical prediction of ocean currents, ensure the accuracy and efficiency of ocean navigation safety and search and rescue of distressed targets, there is an urgent need for a correction method with high complexity and the ability to capture the laws between multi-source heterogeneous data. This is the technical problem to be solved by the present invention. Summary of the invention

[0004] In order to solve the technical problems in the above background, the present invention provides an intelligent empirical correction method for ocean current numerical forecasting based on a multimodal neural network, aiming at the problems of low correction model complexity, poor fitting ability and inability to capture the laws between multi-source heterogeneous data in traditional empirical correction methods, so as to improve the accuracy of ocean current numerical forecasting results, thereby effectively ensuring the safety of ocean navigation, the rationality of marine engineering design, and the accuracy and efficiency of search and rescue of marine distress targets.

[0005] To achieve the above object, the present invention provides an intelligent empirical correction method for ocean current numerical forecast based on a multimodal neural network, the steps comprising:

[0006] Collecting observation data of the ocean current area to be forecasted and constructing a label data set, the steps include: setting a number of observation stations in the ocean current area to be forecasted, collecting station observation data and performing data preprocessing to obtain a label data set;

[0007] Preprocess the label dataset to construct a feature dataset. The steps include: obtaining the pattern data of the feature dataset through a numerical model, preprocessing the pattern data, and constructing the feature dataset;

[0008] Based on the label dataset and the feature dataset, construct a correction model using a multi-modal neural network. The constructed correction model includes: using Res-MLP+CNN-Net as the architecture, adopting the structure of an encoder plus a decoder; the encoder consists of ResMLP and Resnet50, and the decoder adopts Addition;

[0009] Based on the correction model, complete the correction of ocean current forecasting.

[0010] Preferably, the method for preprocessing the site observation data includes: removing samples where the zonal and meridional seawater velocities do not meet the continuity test conditions. The formula for the continuity test conditions is as follows:

[0011] ,

[0012] where x is the zonal and meridional seawater velocity, x i is the current observed value, and x i-1 is the previous observed value adjacent to x i , and H is the test threshold.

[0013] Preferably, the numerical model includes: an atmospheric model, a wave model, and an ocean current model. The variables of the pattern data include zonal seawater velocity at the sea surface, meridional seawater velocity at the sea surface, sea surface temperature, sea surface salinity, sea surface height, significant wave height, wave direction, 10 m zonal wind speed, and 10 m meridional wind speed.

[0014] Preferably, combine the feature dataset and the label dataset to generate a feature-label dataset; then divide the feature-label dataset into a training set, a validation set, and a test set. The division steps include: taking the first 80% of the samples at each site and shuffling them as the training set, the subsequent 10% as the validation set, and the last 10% as the test set; among them, the training set is used to train the model, the validation set is used to supervise whether there is overfitting in the training process, and the test set is used to test the model performance.

[0015] Preferably, in the correction model:

[0016] The input data of ResMLP is the pattern single-point sequence data, and the output data is the scalar sequence data;

[0017] The input data of Resnet50 is the pattern spatial field sequence data, and the output data is the scalar sequence data;

[0018] The input data of Addition is several scalar sequence data, and the output data is scalar sequence data.

[0019] Preferably, a hyperparameter grid search strategy is adopted to find the optimal multi-modal neural network structure:

[0020] For ResMLP, the traversal grid of the hidden layer is set as an array with a range from 1 to 9 with a step size of 1, and the number of neurons in each hidden layer is set as an array with a range from 1 to 30 with a step size of 3;

[0021] Taking the pattern single-point sequence data in the training set as features and the observed data in the training set as labels, traverse the hyperparameter network for training;

[0022] Taking the root mean square error as the quantization index, select the ResMLP structure with the best performance on the test set, and based on the ResMLP, Resnet50 and Addition of the optimal structure, construct the optimal Res-MLP+CNN-Net multi-modal neural network structure.

[0023] Preferably, after the correction model is constructed, use the constructed feature-label data set to train the correction model, and the training strategies include: adjusting the number of training rounds, modifying the training batch size, modifying the learning rate, and modifying the optimizer.

[0024] The present invention also provides a sea current numerical prediction intelligent empirical correction system based on a multi-modal neural network. The system is used to implement the above method and includes: a collection module, a processing module, a construction module, and a correction module;

[0025] The collection module is used to collect the observed data of the sea current area to be predicted and construct a label data set. The steps include: setting a number of observation stations in the sea current area to be predicted, collecting the station observation data and performing data preprocessing to obtain the label data set;

[0026] The processing module is used to preprocess the label data set and construct a feature data set. The steps include: obtaining the pattern data of the feature data set through a numerical model, and performing data preprocessing on the pattern data to construct the feature data set;

[0027] The construction module is used to construct a correction model based on the label data set and the feature data set by using a multi-modal neural network. The constructed correction model includes: with Res-MLP+CNN-Net as the architecture, adopting the structure of an encoder plus a decoder; the encoder is composed of ResMLP and Resnet50, and the decoder adopts Addition;

[0028] The correction module is used to complete the correction of the sea current prediction based on the correction model.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] The present invention has the ability to adaptively capture the laws between multi-source heterogeneous data, improves the ocean current correction effect, thereby improving the accuracy of ocean current forecasting, and provides more reliable forecasting data for the development of the ocean economy, ocean research, and climate prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0032] Figure 1 It is the architecture diagram of the Res-MLP+CNN-Net multi-modal neural network for the embodiment of the present invention;

[0033] Figure 2 It is the overall process framework diagram for the embodiment of the present invention;

[0034] Figure 3 It is the performance distribution diagram of multiple models for correcting the zonal seawater velocity in the embodiment of the present invention;

[0035] Figure 4 It is the performance distribution diagram of multiple models for correcting the meridional seawater velocity in the embodiment of the present invention;

[0036] Figure 5 It is the performance display diagram of the optimal Res-MLP+CNN-Net for correcting the zonal seawater velocity in the embodiment of the present invention;

[0037] Figure 6 It is the performance display diagram of the optimal Res-MLP+CNN-Net for correcting the meridional seawater velocity in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0039] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] Embodiment 1

[0041] This embodiment provides an intelligent empirical correction method for ocean current numerical prediction based on a multimodal neural network. The steps include:

[0042] S1. Collect the observation data of the ocean current area to be predicted and construct a labeled data set.

[0043] Determine the correction area and set several observation stations, collect the station observation data and perform data preprocessing to construct a labeled data set. Specifically, in this embodiment, the seawater velocity observation data of four stations are obtained through buoys equipped with ocean current detection sensors, and the required model data are obtained through the WRF atmospheric model, the WW3 wave model, and the ROMS ocean current model. The model data include the following variables: zonal seawater velocity at the sea surface, meridional seawater velocity at the sea surface, sea surface temperature, sea surface salinity, sea surface height, significant wave height, wave direction, 10 m zonal wind speed, and 10 m meridional wind speed. Among them, the sampling interval of the observation data at all stations is 10 minutes. The time span of Station 1 is from 00:10:00 on September 2, 2021 to 01:50:00 on March 25, 2022, and the longitude and latitude are (113.73°E, 21.21°N). The time span of Station 2 is from 00:00:00 on April 15, 2021 to 02:00:00 on August 18, 2021, and the longitude and latitude are (113.96°E, 21.35°N). The time span of Station 3 is from 00:00:00 on March 17, 2021 to 02:00:00 on September 2, 2021, and the longitude and latitude are (113.46°E, 21.56°N). The time span of Station 4 is from 00:00:00 on August 17, 2021 to 01:50:00 on May 1, 2022, and the longitude and latitude are (113.78°E, 21.88°N). The data interval corresponding to each model is 1 hour, and the time span is from February 1, 2021 to October 31, 2021 and from December 1, 2021 to May 31, 2022. The spatial resolution of the atmospheric model data is 0.1°×0.1°, and the longitude and latitude range is (109.93°E to 118.02°E, 14.95°N to 27.05°N). The spatial resolution of the wave model data is 0.05°×0.05°, and the longitude and latitude range is (112.5°E to 114.95°E, 19.90°N to 22.55°N). The spatial resolution of the ocean current model data is 0.03°×0.03°, and the longitude and latitude range is (112.46°E to 114.94°E, 19.91°N to 22.56°N).

[0044] After the observation data is collected, any abnormal samples with missing values are removed. The specific method is as follows:

[0045] Eliminate samples where the seawater velocity in the meridional and zonal directions does not meet the continuity test conditions where x is the seawater velocity in the meridional and zonal directions, x i is the current observed value, and x i-1 is the previous observed value adjacent to x i , and H is the test threshold, where H is 0.4 m / s;

[0046] After that, perform hourly average resampling on the data, and merge the observed data of all stations in the order of the station positions to form a labeled dataset.

[0047] S2. Preprocess the labeled dataset to construct a feature dataset.

[0048] Specifically, obtain the model data of the feature dataset through a numerical model, and perform data preprocessing on the model data to construct a feature dataset. The numerical model includes an atmospheric model, a wave model, and a current model. The variables of the model data include the zonal seawater velocity at the sea surface, the meridional seawater velocity at the sea surface, the sea surface temperature, the sea surface salinity, the sea surface height, the significant wave height, the wave direction, the zonal wind speed at 10 m, and the meridional wind speed at 10 m.

[0049] The steps for the above preprocessing include:

[0050] For duplicate data on the time axis, only retain the first data and eliminate the rest;

[0051] Assign all abnormal data as NaN;

[0052] Statistically calculate the global mean and standard deviation of each variable in the model data respectively, and then perform Z-score normalization on each variable to obtain the normalized model dataset.

[0053] Normalization can accelerate model convergence, improve optimization efficiency, avoid numerical instability, enhance model performance, strengthen model generalization ability, and simplify hyperparameter tuning. When training a neural network, the features of the input data may have different scales or distribution ranges. This inconsistency in data distribution can cause the model to take longer to adjust parameters during training, thereby reducing the convergence speed. By normalizing the input data, each feature can be brought to the same scale, thus accelerating model convergence. Neural networks usually use gradient-based optimization algorithms for training. If the distribution ranges of the input data vary significantly, the gradient updates may be too large in some directions and too small in others, leading to instability in the optimization process. Normalization can make the gradients more balanced in all directions, thereby improving the efficiency of the optimization algorithm. During training, a too large distribution range of the input data may cause numerical instability problems. Normalization can limit the data within a reasonable range, thus avoiding such problems and ensuring the stability of model training. Normalization can make the distribution of the input data more in line with the assumptions of the neural network, which helps the model better learn the internal laws of the data, thereby improving model performance. Normalization can reduce the overfitting phenomenon of the model to the training data distribution. By converting the input data to a unified distribution range, the model can better learn the essential features of the data instead of relying too much on the specific numerical ranges of certain features, thereby improving the model's generalization ability. Normalization can make the distribution of the input data more consistent, thus reducing the sensitivity to hyperparameters such as the learning rate. This makes the hyperparameter tuning process simpler and improves the robustness of the model.

[0054] The Z-score normalization formula is as follows:

[0055] ,

[0056] where, is the normalized data, is the input data, represents the mean value of represents the standard deviation of

[0057] Based on the normalized model dataset, interpolation is performed based on the spatial axis of the ocean current model and the time axis of the label dataset, and all NaN values are assigned 0 to obtain the model spatial field sequence data;

[0058] Based on the normalized model dataset, interpolation is performed based on the spatial axis and time axis of the label dataset to obtain the model single-point sequence data;

[0059] Combine the model spatial field sequence data and the model single-point sequence data to obtain the feature dataset.

[0060] S3. Based on the labeled dataset and the feature dataset, a correction model is constructed using a multimodal neural network.

[0061] First, the feature dataset and the labeled dataset are combined to generate a feature-label dataset. Then, the feature-label dataset is divided into a training set, a validation set, and a test set. The division steps are as follows: Take the first 80% of the samples at each site and shuffle them as the training set, the subsequent 10% as the validation set, and the last 10% as the test set; among them, the training set is used to train the model, the validation set is used to monitor whether there is overfitting in the training process, and the test set is used to test the model performance.

[0062] After that, a correction model is constructed using a multimodal neural network, and a hyperparameter grid search strategy is used to find the optimal multimodal neural network structure. In the development and optimization process of neural network models, hyperparameter grid search is a systematic method to explore the optimal model structure. Its core significance lies in ensuring the maximization of model performance and enhancing the robustness of the technical solution by exhausting and validating hyperparameter combinations. The selection of hyperparameters such as the structure design of the neural network directly affects the model performance. Traditional methods rely on manual experience or local trial and error, which are difficult to cover the complex hyperparameter space. By defining the possible value ranges of hyperparameters, a multi-dimensional parameter space is constructed, and all combinations are systematically traversed to avoid missing potential optimal solutions.

[0063] Neural networks have significant technical advantages compared to traditional statistical models, mainly reflected in their modeling ability for complex data, automatic feature learning, capturing of non-linear relationships, and adaptability to large-scale data processing. Through the stacking of multiple layers of non-linear activation functions, neural networks can automatically model the high-order non-linear interactions between inputs and outputs without the need for manual pre-setting of interaction forms. Through the distributed representation of the hidden layer, neural networks can automatically extract high-dimensional abstract features from the original data. In addition, neural networks support end-to-end training, directly mapping the original input to the target output, reducing information loss in intermediate links.

[0064] Specifically, as Figure 1 shown, the model architecture is Res-MLP+CNN-Net. This neural network adopts an encoder+decoder structure. The encoder consists of a multi-layer perceptron with residual connections (ResMLP) and a Resnet50 convolutional neural network (Resnet50), and the decoder uses addition operation (Addition), where:

[0065] The input data of ResMLP is pattern single-point sequence data, and the output data is scalar sequence data;

[0066] The input data of Resnet50 is pattern spatial field sequence data, and the output data is scalar sequence data;

[0067] The input data of Addition is several scalar sequence data, and the output data is scalar sequence data;

[0068] Thus, the input data of Res-MLP+CNN-Net is pattern single-point sequence data and pattern space field sequence data, and the output is scalar sequence data.

[0069] The above-mentioned hyperparameter grid search strategy is specifically as follows:

[0070] For ResMLP, the traversal grid of the hidden layer is set as an array with a range from 1 to 9 with a step size of 1, and the number of neurons in each hidden layer is set as an array with a range from 1 to 30 with a step size of 3;

[0071] Using the pattern single-point sequence data in the training set as features and the observed data in the training set as labels, traverse the hyperparameter network for training;

[0072] Taking the root mean square error as the quantization index, select the ResMLP structure with the best performance on the test set, and based on the ResMLP, Resnet50 and Addition of this structure, construct the optimal Res-MLP+CNN-Net multimodal neural network structure.

[0073] The formulas for the root mean square error and the correlation coefficient are as follows:

[0074] ,

[0075] ,

[0076] Among them, represents the root mean square error, represents the correlation coefficient, represents the sequence, represents another sequence, represents the sample serial number, represents the sample size, represents the average value of represents the average value of

[0077] After the model is constructed, use the constructed feature-label data set to train the correction model. The training strategies include: adjusting the number of training epochs, modifying the training batch size, modifying the learning rate, and modifying the optimizer.

[0078] Specifically, using the feature dataset in the training set as features and the label dataset in the training set as labels, select multiple model training strategies to train the model, find the optimal training strategy, repeat the training based on this strategy to obtain multiple pre-trained models, and select the optimal pre-trained model as the final correction model.

[0079] S4. Based on the correction model, complete the correction of the ocean current forecast.

[0080] Input the future feature dataset into the correction model to obtain the corrected forecast product. The overall process architecture of this embodiment is as Figure 2 shown.

[0081] Embodiment 2

[0082] To verify the effectiveness of the present invention compared with the prior art, a comparative embodiment is specifically set here.

[0083] Conduct experiments according to the method provided by the present invention, and construct correction models based on multiple other non-multimodal traditional models. Compare with the Res-MLP+CNN-Net involved in the present invention. Other models include linear regression model (LR), MLP with residual connection (ResMLP), residual convolutional neural network (Resnet50), and vision Transformer (ViT). The experimental results are as follows:

[0084] In the test set, for the zonal seawater velocity, the root mean square error between the ocean current model data and the observed data is 0.147 m / s, and the correlation coefficient is 0.487; for the meridional seawater velocity, the root mean square error between the ocean current model data and the observed data is 0.137 m / s, and the correlation coefficient is 0.050;

[0085] Figure 3 、 Figure 4 are the performance distribution diagrams of multiple models for correcting the zonal and meridional seawater velocities respectively; Table 1 and Table 2 respectively represent the optimal performance information tables of multiple models for correcting the zonal and meridional seawater velocities. Among them, Res-MLP+CNN-Net is the multimodal neural network involved in the present invention, and other models include the ocean current numerical model (ROMS), linear regression model (LR), MLP with residual connection (ResMLP), residual convolutional neural network (Resnet50), and vision Transformer (ViT).

[0086] From Figure 3 and Table 1, it can be seen that for the zonal seawater velocity, the pre-trained model of Res-MLP+CNN-Net has the lowest root mean square error and the highest correlation coefficient compared with other models. The corresponding root mean square error of its optimal pre-trained model is 0.088 m / s, and the correlation coefficient is 0.613.

[0087] Table 1

[0088] 。

[0089] From Figure 4 Table 2, it can be seen that for the meridional seawater velocity, the pre-trained model of Res-MLP+CNN-Net has the lowest root mean square error and the highest correlation coefficient compared with other models. The corresponding root mean square error of its optimal pre-trained model is 0.076 m / s, and the correlation coefficient is 0.319.

[0090] Table 2

[0091] 。

[0092] As Figure 5 、 Figure 6 shown, compared with the zonal and meridional seawater velocities of the ocean current model data, the correction values given by the model provided by the present invention are closer to the true values.

[0093] Based on the powerful fitting ability and generalization ability of the multi-modal neural network, and its ability to adaptively capture the laws between multi-source heterogeneous data, the present invention improves the ocean current correction effect, thereby improving the correction accuracy and providing more reliable forecast data for ocean economic development, ocean research, and climate prediction.

[0094] Embodiment 3

[0095] This embodiment also provides an intelligent empirical correction system for ocean current numerical forecasting based on a multi-modal neural network, including: a collection module, a processing module, a construction module, and a correction module; the collection module is used to collect the observation data of the ocean current area to be forecasted and construct a label data set; the processing module is used to preprocess the label data set and construct a feature data set; the construction module is used to construct a correction model based on the label data set and the feature data set by using a multi-modal neural network; the correction module is used to complete the correction of the ocean current forecast based on the correction model.

[0096] Next, in combination with this embodiment, it will be described in detail how the present invention solves the technical problems in actual work.

[0097] First, use the collection module to collect the observation data of the ocean current area to be forecasted and construct a label data set.

[0098] Determine the correction area and set up several observation stations, collect the observation data of the stations and perform data preprocessing to construct a labeled data set. Specifically, in this embodiment, the seawater velocity observation data of four stations are obtained through buoys equipped with ocean current detection sensors, and the required model data are obtained through the WRF atmospheric model, the WW3 wave model, and the ROMS ocean current model. The model data includes the following variables: zonal seawater velocity at the sea surface, meridional seawater velocity at the sea surface, sea surface temperature, sea surface salinity, sea surface height, significant wave height, wave direction, 10 m zonal wind speed, and 10 m meridional wind speed. Among them, the sampling interval of the observation data of all stations is 10 minutes. The time span of Station 1 is from 00:10:00 on September 2, 2021 to 01:50:00 on March 25, 2022, and the longitude and latitude are (113.73°E, 21.21°N). The time span of Station 2 is from 00:00:00 on April 15, 2021 to 02:00:00 on August 18, 2021, and the longitude and latitude are (113.96°E, 21.35°N). The time span of Station 3 is from 00:00:00 on March 17, 2021 to 02:00:00 on September 2, 2021, and the longitude and latitude are (113.46°E, 21.56°N). The time span of Station 4 is from 00:00:00 on August 17, 2021 to 01:50:00 on May 1, 2022, and the longitude and latitude are (113.78°E, 21.88°N). The data interval corresponding to each model is 1 hour, and the time span is from February 1, 2021 to October 31, 2021 and from December 1, 2021 to May 31, 2022. The spatial resolution of the atmospheric model data is 0.1°×0.1°, and the longitude and latitude range is (109.93°E to 118.02°E, 14.95°N to 27.05°N). The spatial resolution of the wave model data is 0.05°×0.05°, and the longitude and latitude range is (112.5°E to 114.95°E, 19.90°N to 22.55°N). The spatial resolution of the ocean current model data is 0.03°×0.03°, and the longitude and latitude range is (112.46°E to 114.94°E, 19.91°N to 22.56°N).

[0099] After the observation data is collected, any abnormal samples with missing values are removed. The specific method is as follows:

[0100] Remove the samples where the zonal and meridional seawater velocities do not meet the continuity test conditions where x is the zonal and meridional seawater velocity, x i is the current observed value, and x i-1 is the previous observed value adjacent to x i and H is the test threshold, where H is 0.4 m / s here;

[0101] After that, the data is resampled by hourly averaging, and the observed data of all stations are merged in the order of station positions to form a labeled data set.

[0102] The processing module is used to preprocess the labeled data set to construct a feature data set.

[0103] Specifically, the model data of the feature data set is obtained through a numerical model, and the model data is preprocessed to construct a feature data set. The numerical model includes an atmospheric model, a wave model, and a current model. The variables of the model data include the zonal sea water velocity at the sea surface, the meridional sea water velocity at the sea surface, the sea surface temperature, the sea surface salinity, the sea surface height, the significant wave height, the wave direction, the 10 m zonal wind speed, and the 10 m meridional wind speed.

[0104] The steps for the above preprocessing include:

[0105] For duplicate data on the time axis, only the first data is retained, and the rest are excluded;

[0106] All abnormal data are combined and assigned as NaN;

[0107] The global mean and standard deviation of each variable in the model data are respectively calculated, and then each variable is normalized by Z-score to obtain a normalized model data set.

[0108] Normalization can accelerate model convergence, improve optimization efficiency, avoid numerical instability, enhance model performance, strengthen model generalization ability, and simplify hyperparameter tuning. When training a neural network, the features of the input data may have different scales or distribution ranges. This inconsistency in data distribution can cause the model to take longer to adjust parameters during training, thereby reducing the convergence speed. By normalizing the input data, each feature can be brought to the same scale, thus accelerating model convergence. Neural networks usually use gradient-based optimization algorithms for training. If the distribution ranges of the input data vary significantly, the gradient updates may be too large in some directions and too small in others, resulting in an unstable optimization process. Normalization can make the gradients more balanced in all directions, thereby improving the efficiency of the optimization algorithm. During training, a too large distribution range of the input data may lead to numerical instability problems. Normalization can limit the data within a reasonable range, thus avoiding such problems and ensuring the stability of model training. Normalization can make the distribution of the input data more in line with the assumptions of the neural network, which helps the model better learn the internal laws of the data, thereby improving model performance. Normalization can reduce the overfitting phenomenon of the model to the training data distribution. By converting the input data to a unified distribution range, the model can better learn the essential features of the data rather than relying too much on the specific numerical ranges of certain features, thereby improving the model's generalization ability. Normalization can make the distribution of the input data more consistent, thereby reducing the sensitivity to hyperparameters such as the learning rate. This makes the hyperparameter tuning process simpler and improves the robustness of the model.

[0109] The Z-score normalization formula is as follows:

[0110] ,

[0111] where, is the normalized data, is the input data, represents the mean value of represents the standard deviation of

[0112] Based on the normalized model dataset, interpolation is performed based on the spatial axis of the ocean current model and the time axis of the label dataset, and all NaN values are assigned as 0 to obtain the model spatial field sequence data;

[0113] Based on the normalized model dataset, interpolation is performed based on the spatial axis and time axis of the label dataset to obtain the model single-point sequence data;

[0114] Combine the model spatial field sequence data and the model single-point sequence data to obtain the feature dataset.

[0115] The building block constructs a correction model using a multi-modal neural network based on a labeled dataset and a feature dataset.

[0116] First, the feature dataset and the labeled dataset are combined to generate a feature-label dataset. Then, the feature-label dataset is divided into a training set, a validation set, and a test set. The division steps are as follows: Take the first 80% of the samples at each site and shuffle them as the training set, the subsequent 10% as the validation set, and the last 10% as the test set; among them, the training set is used to train the model, the validation set is used to monitor whether there is overfitting in the training process, and the test set is used to test the model performance.

[0117] After that, a correction model is constructed using a multi-modal neural network, and a hyperparameter grid search strategy is used to find the optimal multi-modal neural network structure. In the development and optimization process of a neural network model, hyperparameter grid search is a systematic method to explore the optimal model structure. Its core significance lies in ensuring the maximization of model performance and enhancing the robustness of the technical solution by exhausting and validating hyperparameter combinations. The selection of hyperparameters such as the structure design of the neural network directly affects the model performance. Traditional methods rely on manual experience or local trial and error, which are difficult to cover the complex hyperparameter space. By defining the possible value ranges of hyperparameters, constructing a multi-dimensional parameter space, and systematically traversing all combinations, potential optimal solutions can be avoided from being missed.

[0118] Neural networks have significant technical advantages compared to traditional statistical models, mainly reflected in their modeling ability for complex data, automatic feature learning, capturing of non-linear relationships, and adaptability to large-scale data processing. Through the stacking of multiple layers of non-linear activation functions, neural networks can automatically model the high-order non-linear interactions between inputs and outputs without the need for manual pre-setting of interaction forms. Through the distributed representation of the hidden layer, neural networks can automatically extract high-dimensional abstract features from the original data. In addition, neural networks support end-to-end training, directly mapping the original input to the target output and reducing information loss in the intermediate links.

[0119] Specifically, as Figure 1 shown, the model architecture is Res-MLP+CNN-Net. This neural network adopts an encoder+decoder structure. The encoder consists of a multi-layer perceptron with residual connections (ResMLP) and a Resnet50 convolutional neural network (Resnet50), and the decoder uses addition operation (Addition), where:

[0120] The input data of ResMLP is pattern single-point sequence data, and the output data is scalar sequence data;

[0121] The input data of Resnet50 is pattern spatial field sequence data, and the output data is scalar sequence data;

[0122] The input data of Addition is several scalar sequence data, and the output data is scalar sequence data;

[0123] Therefore, the input data of Res-MLP+CNN-Net is pattern single-point sequence data and pattern space field sequence data, and the output is scalar sequence data.

[0124] The above-mentioned hyperparameter grid search strategy is specifically as follows:

[0125] For ResMLP, the traversal grid of the hidden layer is set as an array with a range from 1 to 9 with a step size of 1, and the number of neurons in each hidden layer is set as an array with a range from 1 to 30 with a step size of 3;

[0126] Using the pattern single-point sequence data in the training set as features and the observed data in the training set as labels, traverse the hyperparameter network for training;

[0127] Taking the root mean square error as the quantization index, select the ResMLP structure with the best performance on the test set, and based on the ResMLP, Resnet50 and Addition of this structure, construct the optimal Res-MLP+CNN-Net multi-modal neural network structure.

[0128] The formulas for the root mean square error and the correlation coefficient are as follows:

[0129] ,

[0130] ,

[0131] Among them, represents the root mean square error, represents the correlation coefficient, represents a sequence, represents another sequence, represents the sample serial number, represents the sample size, represents the average value of represents the average value of

[0132] After the model is constructed, use the constructed feature-label data set to train the correction model. The training strategies include: adjusting the number of training epochs, modifying the training batch size, modifying the learning rate, and modifying the optimizer.

[0133] Specifically, using the feature dataset in the training set as features and the label dataset in the training set as labels, select multiple model training strategies and train the model to find the optimal training strategy. Based on this strategy, repeat the training to obtain multiple pre-trained models, and select the optimal pre-trained model as the final correction model.

[0134] Finally, the correction module completes the correction of the ocean current forecast based on the correction model.

[0135] Input the future feature dataset into the correction model to obtain the corrected forecast product. The overall process architecture of this embodiment is as Figure 2 shown.

[0136] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An intelligent empirical correction method for ocean current numerical forecast based on multimodal neural network, characterized in that: The steps include: Collecting observation data of the ocean current area to be forecasted and constructing a label data set, the steps include: setting a number of observation stations in the ocean current area to be forecasted, collecting station observation data and performing data preprocessing to obtain a label data set; Preprocessing the label data set to construct a feature data set, the steps comprising: obtaining pattern data of the feature data set through a numerical pattern, and preprocessing the pattern data to construct the feature data set; Based on the label data set and the feature data set, a correction model is constructed using a multimodal neural network. The constructed correction model includes: taking Res-MLP+CNN-Net as the architecture, and adopting an encoder plus decoder structure; the encoder is composed of ResMLP and Resnet50, and the decoder adopts Addition; in the correction model: the input data of ResMLP is pattern single-point sequence data, and the output data is scalar sequence data; the input data of Resnet50 is pattern space field sequence data, and the output data is scalar sequence data; the input data of Addition is several scalar sequence data, and the output data is scalar sequence data; Use hyperparameter grid search strategy to find the optimal multimodal neural network structure: for ResMLP, the traversal grid of the hidden layer is set to an array ranging from 1 to 9 with a step size of 1, and the number of neurons in each hidden layer is set to an array ranging from 1 to 30 with a step size of 3; use the single-point sequence data in the training set as features and the observation data in the training set as labels to traverse the hyperparameter network for training; use the root mean square error as a quantitative indicator to select the ResMLP structure with the best performance on the test set, and form the optimal Res-MLP+CNN-Net multimodal neural network structure based on the optimal structure of ResMLP, Resnet50 and Addition; Based on the correction model, the correction of the ocean current forecast is completed.

2. The intelligent empirical correction method for ocean current numerical forecast based on multimodal neural network according to claim 1 is characterized in that: The method for preprocessing the station observation data includes: removing samples whose latitude and longitude seawater flow velocity does not meet the continuity test conditions. The formula of the continuity test condition is as follows: ， Among them, x is the seawater velocity in the longitude and latitude directions, x i is the current observation value, x i-1 For x i The adjacent previous observation value, H is the test threshold.

3. The intelligent empirical correction method for ocean current numerical forecast based on multimodal neural network according to claim 1 is characterized in that: The numerical model includes: atmospheric model, wave model, and ocean current model. The variables of the model data include sea surface latitudinal seawater flow velocity, sea surface meridional seawater flow velocity, sea surface temperature, sea surface salinity, sea surface height, significant wave height, wave direction, 10 m latitudinal wind speed, and 10 m meridional wind speed.

4. The intelligent empirical correction method for ocean current numerical forecast based on multimodal neural network according to claim 1 is characterized in that: The feature data set and the label data set are combined to generate a feature-label data set; then the feature-label data set is divided into a training set, a validation set, and a test set, and the division step includes: taking the first 80% of the samples of each site and shuffling them as the training set, the subsequent 10% as the validation set, and the last 10% as the test set; wherein the training set is used to train the model, the validation set is used to supervise whether there is overfitting in the training process, and the test set is used to test the model performance.

5. The intelligent empirical correction method for ocean current numerical forecast based on multimodal neural network according to claim 1 is characterized in that: After the correction model is constructed, the correction model is trained using the constructed feature-label data set, and the training strategy includes: adjusting the number of training rounds, modifying the training batch size, modifying the learning rate, and modifying the optimizer.

6. An intelligent empirical correction system for ocean current numerical forecast based on a multimodal neural network, the system being used to implement the method described in any one of claims 1 to 5, characterized in that: include: Acquisition module, processing module, construction module and correction module; The acquisition module is used to collect observation data of the ocean current area to be forecasted and construct a label data set, and the steps include: setting a number of observation stations in the ocean current area to be forecasted, collecting the observation data of the stations and performing data preprocessing to obtain the label data set; The processing module is used to preprocess the label data set to construct a feature data set, and the steps include: obtaining pattern data of the feature data set through a numerical pattern, and performing data preprocessing on the pattern data to construct a feature data set; The construction module is used to construct a correction model based on the label data set and the feature data set using a multimodal neural network. The constructed correction model includes: using Res-MLP+CNN-Net as the architecture and adopting an encoder plus decoder structure; the encoder is composed of ResMLP and Resnet50, and the decoder adopts Addition; The correction module is used to complete the correction of the ocean current forecast based on the correction model.

Images