CN-121997046-A - Training method and device for polar sea ice prediction model

CN121997046ACN 121997046 ACN121997046 ACN 121997046ACN-121997046-A

Abstract

The embodiment of the application provides a training method and device for a polar sea ice prediction model, and relates to the technical field of deep learning. The method comprises the steps of extracting historical sea ice state data of a selected target training sample to obtain initial sample sea ice characteristics, determining a sample sea ice prediction result of the target training sample based on the initial sample sea ice characteristics, calculating a loss value of the sample sea ice prediction result through a plurality of loss functions, determining a loss value corresponding to the sample sea ice prediction result, and adjusting model parameters of a polar sea ice prediction model to be trained based on the loss value, wherein the plurality of loss functions comprise a first physical constraint loss function and a second physical constraint loss function, the first physical constraint loss function is used for measuring whether the change of the predicted sea ice concentration accords with the law of sea ice drifting in an actual environment, and the second physical constraint loss function is used for measuring whether the predicted sea ice thickness meets the buoyancy balance physical law. And the physical consistency and the prediction precision of the model are improved.

Inventors

MA LI
LIU HANGBO
LI YANG
ZHANG YONGMEI
MA DONGCHAO
FU YINGXUN
SONG LIHUA

Assignees

北方工业大学

Dates

Publication Date: 20260508
Application Date: 20260119

Claims (10)

1. The training method of the polar sea ice prediction model is characterized by comprising the following steps of: Acquiring a training sample set, wherein each training sample in the training sample set comprises historical sea ice state data of a sample historical period and real sea ice state data of a sample prediction period; performing iterative training on the polar sea ice prediction model to be trained based on the training sample set to obtain a target polar sea ice prediction model, wherein in one iterative training process, the following operations are executed: Extracting characteristics of historical sea ice state data of a selected target training sample to obtain initial sample sea ice characteristics, and determining a sample sea ice prediction result of the target training sample based on the initial sample sea ice characteristics, wherein the sample sea ice prediction result comprises a predicted sea ice concentration and a predicted sea ice thickness; And calculating a loss value of the sample sea ice prediction result through a plurality of loss functions, determining a loss value corresponding to the sample sea ice prediction result, and adjusting model parameters of the polar sea ice prediction model to be trained based on the loss value, wherein the plurality of loss functions comprise a first physical constraint loss function and a second physical constraint loss function, the first physical constraint loss function is used for measuring whether the change of the predicted sea ice concentration accords with the law of sea ice drifting motion in an actual environment, and the second physical constraint loss function is used for measuring whether the predicted sea ice thickness meets the buoyancy balance physical law.
2. The method of claim 1, wherein the calculating the loss value of the sample sea ice prediction result by using a plurality of loss functions, and determining the loss value corresponding to the sample sea ice prediction result, comprises: Calculating a plurality of sub-loss values corresponding to the sample sea ice prediction result based on the plurality of loss functions; And summing the plurality of sub-loss values to obtain the loss value.
3. The method of claim 2, wherein the predicted sea ice concentration comprises a predicted sea ice concentration value for each predicted sample point, the predicted sea ice thickness comprises a predicted sea ice thickness value for each predicted sample point, the sample sea ice prediction result further comprises a predicted surface snow thickness comprising a predicted sea ice thickness value for each predicted sample point; Calculating a plurality of sub-loss values corresponding to the sample sea ice prediction result based on the plurality of loss functions, including: determining respective predicted concentration constraint values of the predicted sample points based on a first change rate of the respective predicted sea ice concentration values of the predicted sample points over time, a second change rate of the respective predicted sea ice concentration values over speed and a irradiance, and calculating errors between the respective predicted concentration constraint values of the predicted sample points and corresponding real concentration constraint values based on the first physical constraint loss function to obtain first sub-loss values, wherein the real concentration constraint values are determined based on real sea ice state data of the target training sample; And determining the respective predicted free plate height of each predicted sample point based on the respective predicted sea ice thickness value and the predicted surface snow thickness value of each predicted sample point in combination with the sea water density, the sea ice density and the snow accumulation density, and calculating the error between the respective predicted free plate height of each predicted sample point and the corresponding real free plate height based on the second physical constraint loss function to obtain a second sub-loss value, wherein the real free plate height is determined based on the real sea ice state data of the target training sample.
4. The method of claim 3, wherein the plurality of loss functions further comprises a mean square error loss function, the method further comprising: And calculating an error between the sample sea ice prediction result and the real sea ice state data of the target training sample based on the mean square error loss function to obtain a third sub-loss value.
5. The method of claim 3, wherein the sample sea ice prediction result further comprises a predicted east speed and a predicted north speed, the predicted east speed comprising a predicted east speed value for each predicted sample point, the predicted north speed comprising a predicted north speed value for each predicted sample point; The method further comprises the following steps of before determining the predicted concentration constraint value of each predicted sample point based on the first change rate of the predicted sea ice concentration value of each predicted sample point along with the time change, the second change rate of the predicted sea ice concentration value along with the speed change and the irradiance: For each predicted sample point, the following operations are respectively executed: Determining a first rate of change of a predicted sea ice concentration value for one predicted sample point over time based on the predicted sea ice concentration value for the one predicted sample point, a predicted sea ice concentration value for a time instant preceding the one predicted sample point, and a time step; Determining a concentration longitude partial derivative and a concentration latitude partial derivative of the predicted sea ice concentration value of the one predicted sample point in a longitude direction and a concentration latitude partial derivative in a latitude direction based on the predicted sea ice concentration value of each adjacent predicted sample point which is east-west adjacent to the one predicted sample point, and determining a second change rate of the predicted sea ice concentration value of the one predicted sample point with the speed based on the predicted east-direction speed value and the predicted north-direction speed value of the one predicted sample point and the concentration longitude partial derivative and the concentration latitude partial derivative; Determining a speed longitude partial derivative of the predicted east speed value of the one predicted sample point in a longitude direction and a speed latitude partial derivative of the predicted north speed value in a latitude direction based on the predicted east speed value of each adjacent predicted sample point adjacent to the one predicted sample point and the predicted north speed value of each adjacent predicted sample point adjacent to the one predicted sample point north-south, and determining a speed field divergence of the one predicted sample point based on the speed longitude partial derivative and the speed latitude partial derivative, and determining a predicted sea ice concentration value of the one predicted sample point and the speed field divergence.
6. A method as defined in claim 5, wherein the determining a second rate of change of the predicted sea ice concentration value for the one predicted sample point as a function of speed based on the predicted east and north velocity values for the one predicted sample point, and the concentration longitude and latitude partial derivatives, comprises: determining a first rate of change component of the predicted sea ice concentration value of the one predicted sample point as a function of speed in the longitudinal direction based on the predicted east speed value of the one predicted sample point and the concentration longitude partial derivative; Determining a second rate of change component of the predicted sea ice concentration value of the one predicted sample point in the latitudinal direction as a function of speed based on the predicted north velocity value of the one predicted sample point and the partial derivative of the concentration latitude; A second rate of change of the predicted sea ice concentration value of the one predicted sample point with speed is determined based on the first rate of change component and the second rate of change component.
7. The method of claim 1, wherein the historical sea ice status data for the target training sample comprises historical sea ice concentration values, historical sea ice thickness values, historical east speed values, historical north speed values, historical surface snow thickness values for each historical sample point; Extracting the characteristic of the historical sea ice state data of the selected target training sample to obtain the sea ice characteristic of the initial sample, wherein the method comprises the following steps: Extracting data required by physical constraint from historical sea ice state data of the target training sample, wherein the data required by physical constraint comprises historical sea ice concentration values of all the historical adjacent sample points which are adjacent to each other in the north and the east, corresponding to each historical sample point, historical sea ice concentration values of all the historical sample points which are adjacent to each other in the west, corresponding to each historical sample point, historical northeast speed values of all the historical adjacent sample points which are adjacent to each other in the east, corresponding to each historical sample point, and historical northbound speed values of all the historical adjacent sample points which are adjacent to each other in the north and the south; Splicing the data required by the physical constraint with the historical sea ice state data of the target training sample to form an enhanced input vector; And carrying out fusion mapping of spatial information and time information on the enhanced input vector through an embedding layer to generate the sea ice features of the initial sample.
8. The method of claim 1, wherein the determining the sample sea ice prediction result for the target training sample based on the initial sample sea ice characteristics comprises: Inputting the initial sample sea ice characteristics into a predictor module sequence comprising n predictor modules, gradually generating output results corresponding to the n predictor modules respectively, and taking the output result corresponding to the n predictor modules as the sample sea ice prediction result; And each predictor module outputs a predicted result component and a residual characteristic based on input data of the predictor module, wherein the residual characteristic is used as input data of a later predictor module, the initial sample sea ice characteristic is used as input data of a1 st predictor module, the output result corresponding to the nth predictor module is obtained by subtracting the output result corresponding to the n-1 st predictor module from the predicted result component of the nth predictor module, and the output result corresponding to the 1 st predictor module is the predicted result component of the 1 st predictor module.
9. The method of claim 8, wherein each predictor module outputs a prediction result component and a residual feature based on its input data, comprising: applying an attention mechanism to input data corresponding to one predictor module to generate an attention feature; Carrying out random discarding treatment on the attention features to obtain discarded features; subtracting the discarded characteristics from the input data corresponding to the predictor module to obtain first intermediate characteristics; And performing first branch processing on the first intermediate feature to obtain a residual feature output by the predictor module, and performing second branch processing on the first intermediate feature to obtain a predicted result component output by the predictor module.
10. The utility model provides a training device of polar region sea ice prediction model which characterized in that includes: The acquisition module is used for acquiring a training sample set, wherein each training sample in the training sample set comprises historical sea ice state data of a sample historical period and real sea ice state data of a sample prediction period; The training module is used for carrying out iterative training on the polar sea ice prediction model to be trained based on the training sample set to obtain a target polar sea ice prediction model, wherein in the one-time iterative training process, the following operations are executed: Extracting characteristics of historical sea ice state data of a selected target training sample to obtain initial sample sea ice characteristics, and determining a sample sea ice prediction result of the target training sample based on the initial sample sea ice characteristics, wherein the sample sea ice prediction result comprises a predicted sea ice concentration and a predicted sea ice thickness; And calculating a loss value of the sample sea ice prediction result through a plurality of loss functions, determining a loss value corresponding to the sample sea ice prediction result, and adjusting model parameters of the polar sea ice prediction model to be trained based on the loss value, wherein the plurality of loss functions comprise a first physical constraint loss function and a second physical constraint loss function, the first physical constraint loss function is used for measuring whether the change of the predicted sea ice concentration accords with the law of sea ice drifting motion in an actual environment, and the second physical constraint loss function is used for measuring whether the predicted sea ice thickness meets the buoyancy balance physical law.

Description

Training method and device for polar sea ice prediction model Technical Field The application relates to the technical field of deep learning, in particular to a training method and device for a polar sea ice prediction model. Background Under the background that the current global climate change and polar exploration activities are increasingly frequent, the accurate prediction of the polar sea ice state is important for channel planning, resource exploration and safety guarantee. The prediction of sea ice concentration (Sea Ice Concentration, SIC) and sea ice thickness (SEA ICE THICKNESS, SIT) is a core index for assessing ice conditions and aiding in decision making, wherein sea ice concentration is also referred to as sea ice concentration. Currently, polar sea ice prediction methods mainly rely on two types of technologies, namely a method based on physical driving and a method based on data driving. The method based on the data driving is to identify trend and periodical change by analyzing historical data, and to predict polar sea ice, and the method based on the data driving comprises methods such as statistical model, traditional machine learning, deep learning and the like. However, the method based on physical driving is limited by the accuracy and model adaptability of data acquisition, complex and nonlinear environment interaction cannot be effectively processed, so that prediction accuracy is low, the statistical model method and the traditional machine learning method in the method based on data driving are difficult to capture nonlinear relations of sea ice changes, important features are easy to ignore, model performance is possibly limited, the deep learning method in the method based on data driving is often regarded as a 'black box' although features can be automatically extracted and high-dimensional data can be processed, model generalization capability is insufficient and interpretability is poor, and particularly physical correlation among variables cannot be fully utilized in multi-variable time sequence data processing, so that prediction accuracy is insufficient. Disclosure of Invention The embodiment of the application provides a training method and a training device for a polar sea ice prediction model, which are used for improving the prediction precision of polar sea ice prediction and enhancing the generalization capability, the interpretability and the physical consistency of the model. In a first aspect, an embodiment of the present application provides a method for training a polar sea ice prediction model, where the method includes: Acquiring a training sample set, wherein each training sample in the training sample set comprises historical sea ice state data of a sample historical period and real sea ice state data of a sample prediction period; Performing iterative training on the polar sea ice prediction model to be trained based on the training sample set to obtain a target polar sea ice prediction model, wherein in the one-time iterative training process, the following operations are executed: Extracting characteristics of historical sea ice state data of a selected target training sample to obtain initial sample sea ice characteristics, and determining a sample sea ice prediction result of the target training sample based on the initial sample sea ice characteristics, wherein the sample sea ice prediction result comprises a predicted sea ice concentration and a predicted sea ice thickness; And calculating a loss value of the sample sea ice prediction result through a plurality of loss functions, determining a loss value corresponding to the sample sea ice prediction result, and adjusting model parameters of a polar sea ice prediction model to be trained based on the loss value, wherein the plurality of loss functions comprise a first physical constraint loss function and a second physical constraint loss function, the first physical constraint loss function is used for measuring whether the change of the predicted sea ice concentration accords with the law of sea ice drifting motion in an actual environment, and the second physical constraint loss function is used for measuring whether the predicted sea ice thickness meets the buoyancy balance physical law. In an alternative embodiment, calculating the loss value of the sample sea ice prediction result through a plurality of loss functions, and determining the loss value corresponding to the sample sea ice prediction result includes: calculating a plurality of sub-loss values corresponding to the sample sea ice prediction result based on the plurality of loss functions; and summing the plurality of sub-loss values to obtain a loss value. In an alternative embodiment, the predicted sea ice concentration comprises a predicted sea ice concentration value of each predicted sample point, the predicted sea ice thickness comprises a predicted sea ice thickness value of each predicted sample point, the sample sea ice predic