Three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity and flow velocity based on deep neural network
Through a deep neural network-based method, three-dimensional spatio-temporal field information of seawater temperature, salinity and flow velocity are extracted and fused, which solves the problem that it is difficult to achieve joint prediction of these elements in the prior art, and improves the accuracy and information completeness of marine environmental state prediction.
Patent Information
- Application Number
- CN202110789794.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-07-13
- Publication Date
- 2025-07-01
- Estimated Expiration
- 2041-07-13
AI Technical Summary
The prior art is difficult to achieve joint prediction of multiple marine environmental factors such as seawater temperature, salinity and flow velocity, and lacks the ability to jointly predict the three-dimensional space-time field of these factors.
Using a deep neural network-based method, three-dimensional spatial field information of seawater temperature, salinity and flow velocity is extracted from the historical time series data output from the ocean numerical mode, and after standardization, a deep neural network model is established for training. The ConvLSTM network and convolutional layer extract and fusion of deep features of the elements is achieved to achieve three-dimensional spatial field joint prediction of these elements.
It improves the accuracy of marine environmental state prediction, provides more complete information support, and provides better support for the optimal observation location and observation path planning of marine robots.
Smart Images
Figure CN115618988B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of marine robots and marine data-driven modeling, and specifically to a three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity, and flow velocity based on a deep neural network. Background Art
[0002] The prediction of the states of marine elements such as seawater temperature, salinity, and flow velocity is of great significance for many civil and military applications. In order to achieve accurate prediction of the marine element field, it is necessary to use Lagrangian platforms for in-situ marine observations. Marine observation platforms such as unmanned surface vehicles, autonomous underwater vehicles, and underwater gliders are several important Lagrangian platforms, and their applications in global and regional marine observation networks are becoming increasingly widespread. Compared with traditional in-situ marine observation platforms, Lagrangian platforms have the advantages of autonomy and controllable maneuverability. Therefore, optimizing the data collection locations and paths of marine robots is a key issue for the operation of marine observation and prediction systems based on Lagrangian platforms.
[0003] In order to optimize the observation locations and observation paths of marine robots in marine environmental state forecasting, in addition to optimization methods and algorithms, a marine model that can measure the utility of sampled data at any time and location is also required, which assimilates sparse data collected by mobile observation platforms and predicts the future marine environmental state, providing support for the optimal observation location and observation path planning of marine robots. In addition to numerical ocean models that need to solve geophysical fluid dynamics differential equations, data-driven models of local area marine environments are an important model that can support the prediction of marine environments in dynamically complex areas with high speed, high precision, and high resolution and the optimization of sampling strategies for marine robots.
[0004] Deep learning is an effective data-driven modeling and prediction method, suitable for local area marine environmental prediction. Currently, data-driven marine environmental modeling methods based on deep learning mainly target single marine environmental elements. There are interactions between multiple elements such as seawater temperature, salinity, and flow velocity, and marine phenomena are jointly affected by multiple marine elements. Therefore, joint modeling of multiple marine elements can provide more complete information for marine element field prediction, improve prediction accuracy, and is of great significance for marine phenomenon analysis. In the existing technologies of marine environmental observation and prediction systems based on marine robots, there is no dynamic model that jointly considers multiple marine environmental elements such as seawater temperature, salinity, and flow velocity, nor is there a technology for jointly predicting the three-dimensional spatio-temporal fields of each element. Summary of the Invention
[0005] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a new three-dimensional marine environmental field prediction model based on deep learning, which jointly establishes the above-mentioned multiple elements, and predicts by jointly considering the element field information such as seawater temperature, salinity, and flow velocity in the neural network.
[0006] The technical solution adopted by the present invention to achieve the above object is: a three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity and flow velocity based on a deep neural network, which is characterized by including the following steps:
[0007] Step 1: Obtain historical time-series data of the three-dimensional spatial field of seawater temperature, salinity and flow velocity output by the ocean numerical model from the dataset;
[0008] Step 2: Standardize the historical data of seawater temperature, salinity and flow velocity in the three-dimensional spatial field;
[0009] Step 3: Divide the standardized dataset into a training sample set, a validation sample set and a test sample set;
[0010] Step 4: Establish a deep neural network model and set the training stop condition;
[0011] Step 5: Use the training sample set and the validation sample set to train the artificial neural network model and select parameters. Terminate the training after reaching the training stop condition to prevent overfitting of the neural network;
[0012] Step 6: Use the test sample set to evaluate and calculate the prediction results of the neural network, and judge whether the neural network meets the optimization goal, so as to obtain an optimized neural network model for predicting seawater temperature, salinity and flow velocity;
[0013] Step 7: Obtain the output data of the ocean numerical model of seawater temperature, salinity and flow velocity at the previous K moments for standardization, input the trained neural network model, and automatically obtain the predicted values of each element corresponding to the three-dimensional spatial field at the next moment to realize the joint prediction of each element.
[0014] The HYCOM ocean numerical model is adopted, and the GLBv0.08-53.X dataset is adopted.
[0015] The dimension of the historical time-series observation data is (K, M, N, D), where K is the number of time steps in the input data time series, M is the length of the input data in the latitude direction, N is the length of the input data in the longitude direction, and D is the number of layers in the three-dimensional spatial field corresponding to the input data.
[0016] The data standardization processing method is:
[0017] z = (x - μ) / σ (1) where x is the original data, z is the data after standardization processing, μ is the mean of the original data, and σ is the standard deviation of the original data.
[0018] The training sample set is 70% of the entire dataset, the validation sample set is 15% of the entire dataset, and the test sample set is 15% of the entire dataset;
[0019] The described artificial neural network model includes four parallel ConvLSTM networks with an encoding–forecasting structure, a fusion feature node, and four parallel convolutional layers; the four parallel ConvLSTM networks with an encoding–forecasting structure are used to extract the deep features of the historical time series of each element in seawater temperature, salinity, and flow velocity; the fusion feature node is used to fuse the extracted deep features of seawater temperature, salinity, and flow velocity to obtain joint features; the four parallel convolutional layers are used to perform convolutional operations on the joint features respectively to obtain the prediction results of the next moment of the three-dimensional spatial field of each element.
[0020] The output dimension of the described ConvLSTM network with an encoding–forecasting structure is (M, N, P), where P is the number of hidden layer states of the ConvLSTM network; the dimension of the described joint features is (M, N, 4×P), and the model prediction output dimension is (M, N, D).
[0021] The training stop conditions of the described artificial neural network model include reaching the maximum number of iterations, the best training time, the minimum cost function value, and the minimum training gradient.
[0022] The cost function is:
[0023]
[0024] where G = M×N×D, (x t , x s , x un , x ue ) and are the true values and model prediction values of seawater temperature, seawater salinity, north-south direction flow velocity, and east-west direction flow velocity respectively.
[0025] The present invention has the following beneficial effects and advantages:
[0026] 1. In the process of neural network modeling, multiple marine element field information is combined, providing more complete information support for element field prediction;
[0027] 2. In the process of neural network modeling, three-dimensional marine element field information is considered, providing information support for the interaction between different layers of ocean water bodies for element field prediction;
[0028] 3. The prediction accuracy of the marine environmental state is improved. Description of the Drawings
[0029] Figure 1 is the overall flowchart of the present invention;
[0030] Figure 2 is the model structure diagram of the present invention; Specific embodiments
[0031] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe in detail the specific implementation methods of the present invention with reference to the accompanying drawings. Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the invention. Therefore, the present invention is not limited by the specific implementations disclosed below.
[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.
[0033] As Figure 1 shown, a three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity, and flow velocity based on a deep neural network according to the present invention includes the following steps:
[0034] Step 1: Obtain historical time-series data of the three-dimensional spatial field to be processed output by the HYCOM ocean numerical model from the GLBv0.08-53.X dataset.
[0035] Step 2: Standardize the historical time-series observation data of seawater temperature, salinity, and flow velocity in the three-dimensional spatial field with dimensions (K, M, N, D), where K is the number of time steps in the input data time series, M is the length of the input data in the latitude direction, N is the length of the input data in the longitude direction, and D is the number of layers in the corresponding three-dimensional spatial field of the input data. The standardization method used is z = (x - μ) / σ, where x is the original data, z is the data after standardization, and n is the number of samples. In the experiment, K = 16, M = 32, N = 32, D = 10;
[0036] Step 3: Divide the standardized dataset into a training sample set, a validation sample set, and a test sample set, where the training sample set is 70% of the entire dataset, the validation sample set is 15% of the entire dataset, and the test sample set is 15% of the entire dataset;
[0037] Step 4: Set the training termination conditions for the artificial neural network model, including the maximum number of iterations, the optimal training time, the minimum cost function value, the minimum training gradient value, etc. The cost function used is where G = M × N × D, (x t , x s, x un , x ue ) and are the true values and model predicted values of seawater temperature, seawater salinity, north-south direction flow velocity, and east-west direction flow velocity respectively. In the experiment, the maximum number of iterations is 100, the optimal training time is 1 hr, the minimum cost function value is 1e -2 , and the minimum training gradient value is 1e -6 ;
[0038] Step Five: Use the training sample set and the validation sample set to train the artificial neural network model and select parameters. The artificial neural network model first uses the encoding–forecasting structure ConvLSTM network to extract the deep features of the historical time series of each element such as seawater temperature, salinity, and flow velocity; secondly, combine the deep features corresponding to each element in the neural network to provide complete information of the fused associated element features for the prediction of the spatial field of each element; then obtain the prediction results of the three-dimensional spatial field of each element through convolutional operations on the combined features. During the training process, when the maximum number of iterations, or the optimal training time, or the minimum cost function value, or the minimum training gradient value is reached, the training is terminated to prevent overfitting;
[0039] Step Six: Use the test sample set to evaluate the prediction performance of the neural network. Use the cost function evaluation index to evaluate the generalization ability of the neural network through the prediction results of the test samples not used during the training process;
[0040] Step Seven: Use the μ and σ values calculated by the standardization processing method in Step One to perform standardization processing on the data to be predicted. The neural network inputs the three-dimensional spatio-temporal field sequence after standardization processing of each element and outputs the predicted values of the three-dimensional spatial field corresponding to each element at the next moment to achieve the joint prediction of each element.
[0041] As Figure 2 shown, the specific model structure of the model in a deep learning model for joint prediction of seawater temperature, salinity, and flow velocity in a three-dimensional spatial field of the present invention is as follows:
[0042] The time series of the three-dimensional spatial fields of the previous K consecutive time steps of each element are respectively input into an encoding–forecasting structure ConvLSTM network, and the network input data dimension of each element is (K, M, N, D). Among them, the size of the convolutional kernel used by each ConvLSTM during convolutional operations is 3×3×D, and the number of hidden layer states of each ConvLSTM network used is P. The output dimension of the ConvLSTM network is (M, N, P). In the experiment, P = 32.
[0043] In order to combine the features of multiple ocean elements during the prediction process, the deep features output by the ConvLSTM network corresponding to each of the aforementioned elements are then concatenated along the feature channel direction to obtain a joint feature with a dimension of (M, N, 4×P). Then, for each element, the prediction output of the three-dimensional spatial field at the next moment is obtained through a convolutional layer. The constructed joint feature is used as the input of the convolutional layer. The size of the convolutional kernel used during the convolutional operation is 3×3×(4×P), and the number of convolutional filters used is D. The output dimension of the convolutional layer corresponding to each element is (M, N, D). In this way, an autoregressive neural network for time series prediction is constructed.
[0044] The embodiments described in the above description will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, several transformations and improvements can be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.
Claims
1. A three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity and flow velocity based on a deep neural network, characterized in that, It includes the following steps: Step 1: Obtain the historical time series data of the three-dimensional spatial fields of seawater temperature, salinity, and flow velocity output by the ocean numerical model from the dataset; the dimension of the historical time series data is (K, M, N, D), where K is the number of time steps in the input data time series, M is the length of the data in the latitude direction, N is the length of the data in the longitude direction, and D is the number of layers in the corresponding three-dimensional spatial field of the input data; Step 2: Standardize the historical data of seawater temperature, salinity, and flow velocity in the three-dimensional spatial field; Step 3: Divide the standardized dataset into a training sample set, a validation sample set, and a test sample set; Step 4: Establish a deep neural network model and set the training stop conditions; the deep neural network model includes four parallel ConvLSTM networks with an encoding–forecasting structure, a fusion feature node, and four parallel convolutional layers; the four parallel ConvLSTM networks with an encoding–forecasting structure are used to extract the deep features of the historical time series of each element in seawater temperature, salinity, and flow velocity; the fusion feature node is used to fuse the extracted deep features of seawater temperature, salinity, and flow velocity to obtain joint features; the four parallel convolutional layers are used to perform convolutional operations on the joint features respectively to obtain the prediction results of the next moment of each element in the three-dimensional spatial field; the output dimension of the ConvLSTM network with an encoding–forecasting structure is (M, N, P), where P is the number of hidden layer states of the ConvLSTM network; the dimension of the joint features is (M, N, 4×P), and the model prediction output dimension is (M, N, D); Step 5: Use the training sample set and the validation sample set to train the deep neural network model and select parameters, and terminate the training after reaching the training stop conditions to prevent overfitting of the deep neural network; Step 6: Use the test sample set to evaluate and calculate the prediction results of the deep neural network, and judge whether the deep neural network meets the optimization goal, so as to obtain an optimized deep neural network model for predicting seawater temperature, salinity, and flow velocity; Step 7: Obtain the output data of the ocean numerical model of seawater temperature, salinity, and flow velocity at the previous K moments for standardization, input the trained deep neural network model, and automatically obtain the predicted values of each element in the corresponding three-dimensional spatial field at the next moment to achieve the joint prediction of each element.
2. The three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity and flow velocity based on a deep neural network according to claim 1, wherein It uses the HYCOM ocean numerical model and the GLBv0.08 - 53.X dataset.
3. The three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity and flow velocity based on a deep neural network according to claim 1, characterized in that, The data standardization processing method is: z = (x - μ) / σ (1) where x is the original data, z is the data after standardization processing, μ is the mean of the original data, and σ is the standard deviation of the original data.
4. The three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity and flow velocity based on a deep neural network according to claim 1, characterized in that The training sample set is 70% of the entire dataset, the validation sample set is 15% of the entire dataset, and the test sample set is 15% of the entire dataset.
5. The three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity and flow velocity based on a deep neural network according to claim 1, wherein The training stop conditions of the deep neural network model include reaching the maximum number of iterations, the best training time, the minimum cost function value, and the minimum training gradient.
6. The three-dimensional spatio-temporal field joint prediction method for seawater temperature, salinity and flow velocity based on a deep neural network according to claim 5, characterized in that The cost function mentioned above is as follows: where i is the spatial position serial number, G = M × N × D, (x ti , x si , x uni , x uei ) and are the true values and model predicted values of seawater temperature, seawater salinity, north-south direction flow velocity, and east-west direction flow velocity, respectively.
Citation Information
Patent Citations
Crime prediction model training method and device and electronic equipment
CN111598059A
Medical image lesion segmentation method
CN112258488A