A method for predicting oil spill drift based on intelligent correction of ocean dynamic fields

Abstract

The present invention discloses a method for predicting oil spill drift based on intelligent correction of the ocean dynamic field, belonging to the technical field of oil spill drift prediction. The method comprises: a corrector and a discriminator respectively updating their loss function values ​​according to two scores; a loop of repeated iterations to obtain a trained model; the trained model is used to correct the numerically predicted ocean dynamic field during the oil spill accident to obtain a corrected ocean dynamic field; the oil spill drift numerical prediction model uses the corrected ocean dynamic field as a driver to obtain an oil spill drift prediction result; and the corrected ocean dynamic field for the next date is continuously updated to obtain a drift prediction result for the next date. The present invention's adversarial temporal convolutional network can effectively resist interference from noisy or incomplete data, significantly reducing computational complexity and enabling faster model training. The adversarial temporal convolutional network efficiently corrects wind flow field data, thereby significantly improving the accuracy of oil spill drift trajectory prediction.

Description

Technical Field

[0001] The invention discloses an oil spill drift prediction method based on ocean dynamic field intelligent correction, belonging to the technical field of oil spill drift prediction. Background Art

[0002] Numerical prediction models for oil spill drift rely on forecast wind and flow field data as inputs, then use numerical calculations to infer the horizontal trajectory of the oil film on the sea surface. However, the accuracy of the predicted ocean dynamic field is limited and cannot meet the precision requirements of oil spill drift prediction. Further analysis of the ocean dynamic field, while highly accurate, carries a time lag of several days, making it unsuitable for predicting oil spill drift in the future. These existing technologies, when using the ocean dynamic field as input for numerical prediction models of oil spill drift, produce inaccurate results. Summary of the Invention

[0003] The purpose of the present invention is to provide an oil spill drift prediction method based on intelligent correction of ocean dynamic field, so as to solve the problem in the prior art that the results of the oil spill drift numerical prediction model are not accurate enough when the ocean dynamic field is used as the input.

[0004] An oil spill drift prediction method based on intelligent correction of ocean dynamic fields, comprising:

[0005] S1 builds an intelligent correction model for ocean dynamic fields based on adversarial learning mechanism, which includes a corrector and a discriminator.

[0006] The S2 corrector learns the residual laws of the predicted ocean dynamic field and the residual laws of the reanalyzed ocean dynamic field to generate the corrected ocean dynamic field;

[0007] The S3 discriminator learns the difference between the corrected ocean dynamic field and the numerically predicted ocean dynamic field and generates a first score;

[0008] The S4 discriminator learns the difference between the reanalyzed ocean dynamic field and the numerically predicted ocean dynamic field and generates a second score;

[0009] The S5 corrector updates the loss function value of the corrector according to the two scores, and the discriminator updates the loss function value of the discriminator according to the two scores;

[0010] S6 imports the training set of the data to be processed into the ocean dynamic field intelligent correction model, and executes S2 to S5 in a loop to complete the learning of the training set;

[0011] S7 executes S2 to S6 cyclically until a preset number of cycles is reached, thereby obtaining a trained ocean dynamic field intelligent correction model;

[0012] S8 uses the intelligent correction model of the ocean dynamic field obtained in S7 to correct the numerically predicted ocean dynamic field during the oil spill accident and obtain the corrected ocean dynamic field;

[0013] S9 uses the corrected ocean dynamic field obtained in S8 as the driving force of the oil spill drift numerical prediction model to obtain the oil spill drift prediction results;

[0014] S10 cycles S7 to S9 multiple times, continuously obtaining the numerically predicted ocean dynamic field at the next time as the input of the correction model of S7, continuously obtaining the corrected ocean dynamic field at the next time, and using the corrected ocean dynamic field at the next time as the input of the numerical prediction model for oil spill drift, to obtain the oil spill drift prediction result at the next time, until the oil spill accident is completed.

[0015] The corrector uses a temporal convolutional network as its main structure. The temporal convolutional network consists of a cascade of L three-dimensional convolutional neural networks (3DCNNs) and a two-dimensional convolutional neural network (2DCNN). The corrector comprehensively extracts the spatial and temporal characteristics of the numerically predicted ocean dynamic field time series data, models the error characteristics between the numerically predicted ocean dynamic field and the reanalyzed ocean dynamic field, and realizes intelligent correction of the numerically predicted ocean dynamic field.

[0016] The discriminator includes five fully convolutional network modules. The discriminator extracts the characteristic information of the input data through multi-layer convolution operations, learns the distribution differences between the corrected ocean dynamic field and the reanalyzed ocean dynamic field, and determines the type of input data based on the extracted data distribution characteristics. The input data type includes reanalyzed wind field data or corrected forecast wind field data.

[0017] S2 includes, let the feature map sequence output by the lth 3DCNN be:

[0018]

[0019] Where, It is the time series data of the feature map output by the Lth 3DCNN. Z is a tensor of all zeros, which is used to fill the feature tensor with zeros on one side when the 3DCNN performs causal convolution. represents the feature map of the lth 3DCNN output at time t, T1 and T2 represent two constants used to set the dimensions of the input time series data, The value of depends on

[0020] Will Input 2DCNN, integrate the feature time series data at different historical moments and output the wind field correction residual W t ':

[0021]

[0022] Corrected ocean dynamic field output by the corrector for:

[0023]

[0024] Where W t For numerical prediction of ocean dynamic fields.

[0025] S3 includes, the discriminator learns W t and The first score of different generation is, and the input of the discriminator is The output is W t and The discriminant score of

[0026] S4 includes, the discriminator learning W t Reanalysis of ocean dynamics The second score is generated by the different The output is W t and The discriminant score of

[0027] The difference in the discrimination scores obtained using S3 and S4 reflects and degree of similarity.

[0028] S5 includes the loss function for training the corrector for:

[0029]

[0030] Where λ1 is the equilibrium parameter, is the L1 norm loss term, is the adversarial loss term of the corrector, and D is the adversarial loss function;

[0031] The L1 norm loss term is constrained by the penalty and The distance between the corresponding values ​​​​prompts and The values ​​of are close to each other, and the adversarial loss term of the corrector stimulates the corrector output and make and Obtain mutually approximate discrimination scores in the discriminator, and continuously train the output of the corrector and the approaching each other Confuse the discriminator so that it cannot distinguish whether the input is or

[0032] The loss function for training the discriminator is:

[0033]

[0034] Where, The discriminator's adversarial loss term is used to enhance the discriminative power of the discriminator. is the gradient penalty term, λ2 is the balance parameter of the gradient penalty term, for The gradient of the gradient penalty term represent and Randomly sample values ​​between:

[0035]

[0036] Where ∈ is and The weight factor between is a random number between 0 and 1;

[0037] Minimize the loss function of the discriminator and continuously improve the discriminant ability of the discriminator so that it can distinguish and

[0038] S6 includes when the discriminator cannot correctly distinguish and When , the adversarial training terminates, at which point the rectifier and the discriminator simultaneously achieve optimal representations, and further adversarial training cannot improve performance;

[0039] The trained rectifier corrects W t , output and approaching each other

[0040] The Lagrangian oil particle model is used as the basic prediction framework of the oil spill drift numerical prediction model. The oil particle model regards the oil spill on the sea surface as an integration of independently analyzed oil spill particles or particles. The oil spill drift process is the superposition of the drift motion of all particles.

[0041] The corrected ocean dynamic field includes the corrected ocean wind field and the corrected ocean current field. The corrected ocean wind field provides ocean wind speed for the prediction model, and the corrected ocean current field provides ocean current speed for the prediction model.

[0042] The numerical prediction model of oil spill drift includes: in the drift process of any oil particle, let u drift and v drift Represent the drift velocity components of oil particles in the u direction and v direction respectively, let and Respectively represent the two wind speed components in the u and v directions at the location of the oil particles at time t, and They represent the two velocity components in the u and v directions at the location of the oil particle at time t. The two drift velocity components of the oil particle at time t are:

[0043]

[0044] Where, α w and α c are the weight coefficients of wind speed and flow velocity respectively;

[0045] Let Δt represent the minimum time resolution of the wind and flow forecasts, Δx drift and Δy drift They represent the drift distance components of oil particles in the u and v directions respectively. The drift distances of the two components in the time interval Δt are:

[0046] Δx drift =u drift Δt; Δy drift =v drift Δt.

[0047] The numerical prediction model of oil spill drift includes the following: the oil particles drift on the sea surface accompanied by random diffusion process, let R n represents a random number with a mean of 0 and a variance of 1, μ represents the weighting parameter of the diffusion process, E r represents the horizontal diffusion rate, θ represents the direction angle uniformly distributed between 0 and 2π, Δx rand and Δy rand They represent the random diffusion distance components of oil particles in the u and v directions respectively. The diffusion distances of the two diffusion distance components at a time interval of Δt are:

[0048]

[0049] The two moving distance components of the oil particles at the time interval Δt are:

[0050] Δx=Δx drift +Δx rand ; Δy=Δy drift +Δy rand .

[0051] The numerical prediction model of oil spill drift includes: t ,y t ) represents the initial longitude and latitude coordinates of the oil particle, γ represents the distance corresponding to a 1° change in latitude under the same longitude conditions, and the predicted longitude and latitude coordinates (x t+Δt ,y t+Δt )for:

[0052]

[0053] Compared with the existing technology, the present invention has the following beneficial effects: the adversarial time convolution network can effectively resist the interference of noisy data or incomplete data, greatly reducing the computational complexity and completing model training faster; the adversarial time convolution network efficiently corrects the wind flow field data, thereby significantly improving the accuracy of oil spill drift trajectory prediction; accurate oil spill drift prediction can help relevant departments to formulate cleanup and prevention plans in a timely manner, reduce the harm of oil spills to the marine ecological environment, and reduce the economic losses caused by oil spill accidents; the present invention is not only suitable for oil spill drift prediction, but can also be extended to other application scenarios involving wind flow field correction, such as marine environment monitoring, shipping safety and weather forecasting, and has strong versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is the predicted result of the first drift of Symphony driven by the wind flow field;

[0055] Figure 2 This is the predicted result of the first drift of Symphony driven by the corrected wind flow field;

[0056] Figure 3 It is the predicted result of the second drift of Symphony driven by the wind flow field;

[0057] Figure 4 This is the predicted result of the second drift of Symphony driven by the corrected wind flow field;

[0058] Figure 5 It is the predicted drift result of Sanchi driven by the forecast wind flow field;

[0059] Figure 6 This is the predicted result of the Sanchi's drift driven by the corrected wind flow field. DETAILED DESCRIPTION

[0060] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0061] An oil spill drift prediction method based on intelligent correction of ocean dynamic fields, comprising:

[0062] S1 builds an intelligent correction model for ocean dynamic fields based on adversarial learning mechanism, which includes a corrector and a discriminator.

[0063] The S2 corrector learns the residual laws of the predicted ocean dynamic field and the residual laws of the reanalyzed ocean dynamic field to generate the corrected ocean dynamic field;

[0064] The S3 discriminator learns the difference between the corrected ocean dynamic field and the numerically predicted ocean dynamic field and generates a first score;

[0065] The S4 discriminator learns the difference between the reanalyzed ocean dynamic field and the numerically predicted ocean dynamic field and generates a second score;

[0066] The S5 corrector updates the loss function value of the corrector according to the two scores, and the discriminator updates the loss function value of the discriminator according to the two scores;

[0067] S6 imports the training set of the data to be processed into the ocean dynamic field intelligent correction model, and executes S2 to S5 in a loop to complete the learning of the training set;

[0068] S7 executes S2 to S6 cyclically until a preset number of cycles is reached, thereby obtaining a trained ocean dynamic field intelligent correction model;

[0069] S8 uses the intelligent correction model of the ocean dynamic field obtained in S7 to correct the numerically predicted ocean dynamic field during the oil spill accident and obtain the corrected ocean dynamic field;

[0070] S9 uses the corrected ocean dynamic field obtained in S8 as the driving force of the oil spill drift numerical prediction model to obtain the oil spill drift prediction results;

[0071] S10 cycles S7 to S9 multiple times, continuously obtaining the numerically predicted ocean dynamic field at the next time as the input of the correction model of S7, continuously obtaining the corrected ocean dynamic field at the next time, and using the corrected ocean dynamic field at the next time as the input of the numerical prediction model for oil spill drift, to obtain the oil spill drift prediction result at the next time, until the oil spill accident is completed.

[0072] The corrector uses a temporal convolutional network as its main structure. The temporal convolutional network consists of a cascade of L three-dimensional convolutional neural networks (3DCNNs) and a two-dimensional convolutional neural network (2DCNN). The corrector comprehensively extracts the spatial and temporal characteristics of the numerically predicted ocean dynamic field time series data, models the error characteristics between the numerically predicted ocean dynamic field and the reanalyzed ocean dynamic field, and realizes intelligent correction of the numerically predicted ocean dynamic field.

[0073] The discriminator includes five fully convolutional network modules. The discriminator extracts the characteristic information of the input data through multi-layer convolution operations, learns the distribution differences between the corrected ocean dynamic field and the reanalyzed ocean dynamic field, and determines the type of input data based on the extracted data distribution characteristics. The input data type includes reanalyzed wind field data or corrected forecast wind field data.

[0074] S2 includes, let the feature map sequence output by the lth 3DCNN be:

[0075]

[0076] Where, It is the time series data of the feature map output by the Lth 3DCNN. Z is a tensor of all zeros, which is used to fill the feature tensor with zeros on one side when the 3DCNN performs causal convolution. represents the feature map of the lth 3DCNN output at time t, T1 and T2 represent two constants used to set the dimensions of the input time series data, The value of depends on

[0077] Will Input 2DCNN, integrate the feature time series data at different historical moments and output the wind field correction residual W t ':

[0078]

[0079] Corrected ocean dynamic field output by the corrector for:

[0080]

[0081] Where W t For numerical prediction of ocean dynamic fields.

[0082] S3 includes, the discriminator learns W t and The first score of different generation is, and the input of the discriminator is The output is W t and The discriminant score of

[0083] S4 includes, the discriminator learning W t Reanalysis of ocean dynamics The second score is generated by the different The output is W t and The discriminant score of

[0084] The difference in the discrimination scores obtained using S3 and S4 reflects and degree of similarity.

[0085] S5 includes the loss function for training the corrector for:

[0086]

[0087] Where λ1 is the equilibrium parameter, is the L1 norm loss term, is the adversarial loss term of the corrector, and D is the adversarial loss function;

[0088] The L1 norm loss term is constrained by the penalty and The distance between the corresponding values ​​​​prompts and The values ​​of are close to each other, and the adversarial loss term of the corrector stimulates the corrector output and make and Obtain mutually approximate discrimination scores in the discriminator, and continuously train the output of the corrector and the approaching each other Confuse the discriminator so that it cannot distinguish whether the input is or

[0089] The loss function for training the discriminator is:

[0090]

[0091] Where, The discriminator's adversarial loss term is used to enhance the discriminative power of the discriminator. is the gradient penalty term, λ2 is the balance parameter of the gradient penalty term, for The gradient of the gradient penalty term represent and Randomly sample values ​​between:

[0092]

[0093] Where ∈ is and The weight factor between is a random number between 0 and 1;

[0094] Minimize the loss function of the discriminator and continuously improve the discriminant ability of the discriminator so that it can distinguish and

[0095] S6 includes when the discriminator cannot correctly distinguish and When , the adversarial training terminates, at which point the rectifier and the discriminator simultaneously achieve optimal representations, and further adversarial training cannot improve performance;

[0096] The trained rectifier corrects W t , output and approaching each other

[0097] The Lagrangian oil particle model is used as the basic prediction framework of the oil spill drift numerical prediction model. The oil particle model regards the oil spill on the sea surface as an integration of independently analyzed oil spill particles or particles. The oil spill drift process is the superposition of the drift motion of all particles.

[0098] The corrected ocean dynamic field includes the corrected ocean wind field and the corrected ocean current field. The corrected ocean wind field provides ocean wind speed for the prediction model, and the corrected ocean current field provides ocean current speed for the prediction model.

[0099] The numerical prediction model of oil spill drift includes: in the drift process of any oil particle, let u drift and v drift Represent the drift velocity components of oil particles in the u direction and v direction respectively, let and Respectively represent the two wind speed components in the u and v directions at the location of the oil particles at time t, and They represent the two velocity components in the u and v directions at the location of the oil particle at time t. The two drift velocity components of the oil particle at time t are:

[0100]

[0101] Where, α w and α c are the weight coefficients of wind speed and flow velocity respectively;

[0102] Let Δt represent the minimum time resolution of the wind and flow forecasts, Δx drift and Δy drift They represent the drift distance components of oil particles in the u and v directions respectively. The drift distances of the two components in the time interval Δt are:

[0103] Δx drift =u drift Δt; Δy drift =v drift Δt.

[0104] The numerical prediction model of oil spill drift includes the following: the oil particles drift on the sea surface accompanied by random diffusion process, let R n represents a random number with a mean of 0 and a variance of 1, μ represents the weighting parameter of the diffusion process, E r represents the horizontal diffusion rate, θ represents the direction angle uniformly distributed between 0 and 2π, Δx rand and Δy rand They represent the random diffusion distance components of oil particles in the u and v directions respectively. The diffusion distances of the two diffusion distance components at a time interval of Δt are:

[0105]

[0106] The two moving distance components of the oil particles at the time interval Δt are:

[0107] Δx=Δx drift +Δx rand ; Δy=Δy drift +Δy rand .

[0108] The numerical prediction model of oil spill drift includes: t ,y t ) represents the initial longitude and latitude coordinates of the oil particle, γ represents the distance corresponding to a 1° change in latitude under the same longitude conditions, and the predicted longitude and latitude coordinates (x t+Δt ,y t+Δt )for:

[0109]

[0110] The present invention uses ocean dynamic field data, SAR images, and optical remote sensing images for experiments and analysis. The two oil spill comprehensive data sets used are both self-built and released, namely the Sanchi oil spill comprehensive data set and the Symphony oil spill comprehensive data set. For the Symphony oil spill comprehensive data set, it includes 1460 numerically predicted ocean wind fields, 1460 numerically predicted ocean current fields, 1460 reanalyzed ocean wind fields, 1460 analyzed ocean current fields from 00:00 on January 1, 2021 to 18:00 on December 31, 2021, and four remote sensing images on May 1, 2021, May 2, 2021, May 5, 2021, and May 7, 2021. The test set includes 30 numerically predicted ocean wind fields, 30 reanalysis ocean wind fields, 30 numerically predicted ocean current fields, and 30 analytical ocean current fields from 00:00 on May 1, 2021, to 06:00 on May 8, 2021, and the remaining data are used as the training set. The Sanchi oil spill comprehensive dataset includes 1,460 numerically predicted ocean wind fields, 1,460 reanalysis ocean wind fields, 12 numerically predicted ocean current fields, and two remote sensing images from January 18, 2018, and January 20, 2018, from 00:00 on January 18, 2018, to 18:00 on January 20, 2018, as the test set, and the remaining data are used as the training set.

[0111] Table 1 shows a comparison of the proposed method and other comparative methods in their ability to correct ocean wind field predictions using the Symphony oil spill dataset. It can be seen that the proposed method achieves the lowest root mean square error (RMS) and mean absolute error (MAE).

[0112] Table 1 Comparison of calibration capabilities using the Symphony ocean wind field dataset

[0113]

[0114] Table 2 shows a comparison of the proposed method and other comparative methods in their ability to correct ocean currents using the Symphony oil spill dataset. It can be seen that the proposed method achieves the lowest root mean square error (RMS) and mean absolute error (MAE).

[0115] Table 2 Comparison of correction capabilities using the Symphony ocean current dataset

[0116]

[0117] Table 3 shows a comparison of the proposed method and other comparative methods in their ability to numerically predict ocean dynamic field corrections using the Sanchi oil spill dataset. It can be seen that the proposed method achieves the lowest root mean square error (RMS) and mean absolute error (MAE).

[0118] Table 3 Comparison of correction capabilities using the Sanchi ocean wind field dataset

[0119]

[0120] The oil spill drift prediction was performed using the comprehensive data set of the Symphony oil spill. The prediction results of the first drift of the Symphony driven by the wind flow field are as follows: Figure 1 As shown in the figure, the predicted drift results of the Symphony No. 2 driven by the wind flow field are as follows: Figure 3 The predicted drift results of the first section of Symphony driven by the corrected wind flow field are shown in Figure 2 As shown in the figure, the predicted drift results of the Symphony No. 2 driven by the corrected wind flow field are as follows: Figure 4 As shown in the figure. The cyan color is the result of numerical prediction of oil spill drift, and the gold color is the result of actual oil spill drift. It can be seen that the method proposed in this invention is closer to the actual oil spill drift result. The oil spill drift prediction is performed using the Sanchi oil spill comprehensive dataset. Figure 5 The results of oil spill drift prediction driven by numerical prediction of ocean dynamics are presented. Figure 6 This figure shows the oil spill drift prediction results obtained using the method proposed in this paper. The cyan color represents the numerical prediction results of the oil spill drift, while the gold color represents the actual oil spill drift results. As can be seen, the proposed method is closer to the actual oil spill drift results.

[0121] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting oil spill drift based on intelligent correction of ocean dynamic field, characterized in that: include: S1 builds an intelligent correction model for ocean dynamic fields based on adversarial learning mechanism, which includes a corrector and a discriminator. The S2 corrector learns the residual laws of the predicted ocean dynamic field and the residual laws of the reanalyzed ocean dynamic field to generate the corrected ocean dynamic field; The S3 discriminator learns the difference between the corrected ocean dynamic field and the numerically predicted ocean dynamic field and generates a first score; The S4 discriminator learns the difference between the reanalyzed ocean dynamic field and the numerically predicted ocean dynamic field and generates a second score; The S5 corrector updates the loss function value of the corrector according to the two scores, and the discriminator updates the loss function value of the discriminator according to the two scores; S6 imports the training set of the data to be processed into the ocean dynamic field intelligent correction model, and executes S2 to S5 in a loop to complete the learning of the training set; S7 executes S2 to S6 cyclically until a preset number of cycles is reached, thereby obtaining a trained ocean dynamic field intelligent correction model; S8 uses the intelligent correction model of the ocean dynamic field obtained in S7 to correct the numerically predicted ocean dynamic field during the oil spill accident and obtain the corrected ocean dynamic field; S9 uses the corrected ocean dynamic field obtained in S8 as the driving force of the oil spill drift numerical prediction model to obtain the oil spill drift prediction results; S10 cycles S7 to S9 multiple times, continuously obtaining the numerically predicted ocean dynamic field at the next time as the input of the correction model of S7, continuously obtaining the corrected ocean dynamic field at the next time, and using the corrected ocean dynamic field at the next time as the input of the numerical prediction model for oil spill drift, to obtain the oil spill drift prediction result at the next time, until the oil spill accident is completed.

2. The oil spill drift prediction method based on intelligent correction of ocean dynamic field according to claim 1 is characterized in that: The corrector uses a temporal convolutional network as its main structure. The temporal convolutional network consists of a cascade of L three-dimensional convolutional neural networks (3DCNNs) and a two-dimensional convolutional neural network (2DCNN). The corrector comprehensively extracts the spatial and temporal characteristics of the numerically predicted ocean dynamic field time series data, models the error characteristics between the numerically predicted ocean dynamic field and the reanalyzed ocean dynamic field, and realizes intelligent correction of the numerically predicted ocean dynamic field. The discriminator includes five fully convolutional network modules. The discriminator extracts the characteristic information of the input data through multi-layer convolution operations, learns the distribution differences between the corrected ocean dynamic field and the reanalyzed ocean dynamic field, and determines the type of input data based on the extracted data distribution characteristics. The input data type includes reanalyzed wind field data or corrected forecast wind field data.

3. The oil spill drift prediction method based on intelligent correction of ocean dynamic field according to claim 2 is characterized in that: S2 includes, let the feature map sequence output by the lth 3DCNN be: Where, It is the time series data of the feature map output by the Lth 3DCNN. Z is a tensor of all zeros, which is used to fill the feature tensor with zeros on one side when the 3DCNN performs causal convolution. represents the feature map of the lth 3DCNN output at time t, T1 and T2 represent two constants used to set the dimensions of the input time series data, The value of depends on Will Input 2DCNN, integrate the feature time series data at different historical moments and output the wind field correction residual W t ': Corrected ocean dynamic field output by the corrector for: Where W t For numerical prediction of ocean dynamic fields.

4. The oil spill drift prediction method based on intelligent correction of ocean dynamic field according to claim 3 is characterized in that: S3 includes, the discriminator learns W t and The first score of different generation is, and the input of the discriminator is The output is W t and The discriminant score of S4 includes, the discriminator learning W t Reanalysis of ocean dynamics The second score is generated by the different The output is W t and The discriminant score of The difference in the discrimination scores obtained using S3 and S4 reflects and degree of similarity.

5. The method for predicting oil spill drift based on intelligent correction of ocean dynamic field according to claim 4, characterized in that: S5 includes the loss function L of the training corrector Ct for: Where λ1 is the equilibrium parameter, is the L1 norm loss term, is the adversarial loss term of the corrector, and D is the adversarial loss function; The L1 norm loss term is constrained by the penalty and The distance between the corresponding values ​​​​prompts and The values ​​of are close to each other, and the adversarial loss term of the corrector stimulates the corrector output and make and Obtain mutually approximate discrimination scores in the discriminator, and continuously train the output of the corrector and the approaching each other Confuse the discriminator so that it cannot distinguish whether the input is or The loss function for training the discriminator is: Where, The discriminator's adversarial loss term is used to enhance the discriminative power of the discriminator. is the gradient penalty term, λ2 is the balance parameter of the gradient penalty term, for The gradient of the gradient penalty term represent and Randomly sample values ​​between: Where ∈ is and The weight factor between is a random number between 0 and 1; Minimize the loss function of the discriminator and continuously improve the discriminant ability of the discriminator so that it can distinguish and 6. The method for predicting oil spill drift based on intelligent correction of ocean dynamic field according to claim 5, characterized in that: S6 includes when the discriminator cannot correctly distinguish and When , the adversarial training terminates, at which point the rectifier and the discriminator simultaneously achieve optimal representations, and further adversarial training cannot improve performance; The trained rectifier corrects W t , output and approaching each other 7. The method for predicting oil spill drift based on intelligent correction of ocean dynamic field according to claim 6, characterized in that: The Lagrangian oil particle model is used as the basic prediction framework of the oil spill drift numerical prediction model. The oil particle model regards the oil spill on the sea surface as an integration of independently analyzed oil spill particles or particles. The oil spill drift process is the superposition of the drift motion of all particles. The corrected ocean dynamic field includes the corrected ocean wind field and the corrected ocean current field. The corrected ocean wind field provides ocean wind speed for the prediction model, and the corrected ocean current field provides ocean current speed for the prediction model.

8. The method for predicting oil spill drift based on intelligent correction of ocean dynamic field according to claim 7, characterized in that: The numerical prediction model of oil spill drift includes: in the drift process of any oil particle, let u drift and v drift Represent the drift velocity components of oil particles in the u direction and v direction respectively, let and Respectively represent the two wind speed components in the u and v directions at the location of the oil particles at time t, and They represent the two velocity components in the u and v directions at the location of the oil particle at time t. The two drift velocity components of the oil particle at time t are: Where, α w and α c are the weight coefficients of wind speed and flow velocity respectively; Let Δt represent the minimum time resolution of the wind and flow forecasts, Δx drift and Δy drift They represent the drift distance components of oil particles in the u and v directions respectively. The drift distances of the two components in the time interval Δt are: Δx drift =u drift Δt;Δy drift =v drift Δt。 9. The method for predicting oil spill drift based on intelligent correction of ocean dynamic field according to claim 8, characterized in that: The numerical prediction model of oil spill drift includes the following: the oil particles drift on the sea surface accompanied by random diffusion process, let R n represents a random number with a mean of 0 and a variance of 1, μ represents the weighting parameter of the diffusion process, E r represents the horizontal diffusion rate, θ represents the direction angle uniformly distributed between 0 and 2π, Δx rand and Δy rand They represent the random diffusion distance components of oil particles in the u and v directions respectively. The diffusion distances of the two diffusion distance components at a time interval of Δt are: The two moving distance components of the oil particles at the time interval Δt are: Δx=Δx drift +Δx rand ;Δy=Δy drift +Δy rand 。 10. The method for predicting oil spill drift based on intelligent correction of ocean dynamic field according to claim 9, characterized in that: The numerical prediction model of oil spill drift includes: t ,y t ) represents the initial longitude and latitude coordinates of the oil particle, γ represents the distance corresponding to a 1° change in latitude under the same longitude conditions, and the predicted longitude and latitude coordinates (x t+Δt ,y t+Δt )for:

Images