Rapid forecasting method for environmental response of ship sailing in ice area

Abstract

The invention provides a rapid forecasting method for environmental response of a ship sailing in an ice region. The rapid forecasting method comprises the following steps: S1, establishing a motion response mathematical model of the ship on the sea under interference of the environment with ice, wind and waves; s2, establishing a ship navigation sea area ice condition and sea condition forecasting model; s3, enriching the environmental motion and structural response database of ship navigation in the ice area or performing data assimilation according to actual measurement information; and S4, according to the planned ship route of the ship and the forecasting result of the regional ice area environmental elements, carrying out refined real-time forecasting of the ice area environmental elements on the ship route region per hour within 24 hours, adopting multi-level decomposition and forecasting in the forecasting process, decomposing the forecasting into different time scales for forecasting, and finally, carrying out the forecasting of the ice area environmental elements in the ship route region. And summing different forecasts to obtain a final forecast result. According to the invention, the motion and structural response of the ship can be quickly and accurately obtained under the influence of a complex ice region environment.

Description

Technical Field

[0001] The present invention relates to the technical field of ice navigation, and in particular to a method for quickly predicting environmental response of a ship sailing in ice. Background Art

[0002] Real-time prediction of ship's attitude at sea plays an important role in ship turning in strong winds and waves, aircraft take-off and landing on deck, underwater vehicle retraction and deployment, offshore transfer, cable retraction and deployment and other offshore construction operations. However, when a ship is sailing at sea, it is affected by environmental factors such as wind, waves and currents. The swaying motion of the ship has complex characteristics such as nonlinearity and dynamic time-varying, which increases the difficulty of its prediction.

[0003] At present, the motion response of ships in ice areas is simulated by numerical simulation. This method is difficult to apply in real time due to the large amount of calculation, and cannot be applied to online assessment of ship risks in ice areas and online adjustment of routes. In addition, the current forecasting method has problems such as long general environmental forecast cycle and difficulty in real-time application. There are also problems such as the strong uncertainty of environmental factor forecasts, which leads to poor prediction accuracy of environmental factors. Summary of the invention

[0004] In view of this, the purpose of the present invention is to propose a method for quickly predicting the environmental response of ships sailing in ice areas, so as to solve the technical problem of poor accuracy of the existing environmental response prediction of ships sailing in ice areas.

[0005] The technical means adopted by the present invention are as follows:

[0006] A rapid prediction method for environmental response of a ship sailing in an ice area comprises the following steps:

[0007] S1. Establish a mathematical model of the ship's motion response and ship structure response under the environmental interference of ice, wind and waves at sea;

[0008] S2. Establish a forecast model for ice and sea conditions in the sea area where ships sail, use a spatiotemporal generative adversarial neural network model with an attention mechanism to achieve assimilation and fusion of observation data and high-resolution model forecast data, dynamically update the forecast of external ice, wind, wave and flow fields, and obtain the initial marine environment forecast results for the sea area where the ship passes;

[0009] S3. Based on the real-time measured ice conditions, wind, wave, current information, and ship motion and structural response information, the environmental motion and structural response database of ships sailing in ice areas is enriched, or based on the measured information of wind, wave and current elements and the ice load inversion information of ship structural response, the ice conditions and sea state forecast model database of ships sailing in the sea area is updated to obtain real-time marine environment forecast results;

[0010] S4. According to the planned route of the ship and the real-time marine environment forecast results, a refined real-time forecast of ice environment elements is made for the ship route area on an hourly basis within 24 hours. In the forecast process, multi-level decomposition and prediction are adopted, the forecast is decomposed into different time scales for forecasting, and different forecasts are summed to obtain the final forecast result;

[0011] S5. According to the sea area through which the ship's planned route passes, the motion response and structural response of the ship passing through the ice area, wind field and wave field of the sea area are calculated by using the motion response and structural response mathematical model; the input of the motion response and structural response mathematical model includes ship element information, environmental element information and dynamic element information, and the output of the motion response and structural response mathematical model is the speed, motion posture and structural response of the ship under the influence of the set ice, wind and waves;

[0012] S6. Based on the motion response and structural response of the ship such as speed and motion posture under the influence of ice, wind and waves, combined with the final forecast results, a comprehensive forecast of the response of the ship sailing in ice areas is carried out.

[0013] Furthermore, S1 specifically includes the following steps:

[0014] S11, setting the input of the training model and the output of the training model; the input of the training model includes ship element information, environmental element information and dynamic element information; the output of the training model includes the speed, motion posture and structural response of the ship under the influence of ice, wind and waves;

[0015] S12. Based on the model input and model output, a model input-output training sample is established by numerically simulating the motion and structural response of the ship in ice, and the response of a specific ship to environmental factors in terms of ship speed, motion posture and structure is calculated; assuming that there are m samples in the sample set, the input vector x has n independent variables, the input X is an m×n dimensional matrix, the output vector y has l dependent variables, and the output Y is an m×l dimensional matrix;

[0016] S13. Based on the input and output data of the samples obtained in S12, offline training of the neural network makes the neural network accurately approximate the input-output mapping, and establishes a neural network mapping model of the motion and structural response model of the ship's navigation.

[0017] Furthermore, the ship element information includes the main dimensions of the ship, structural parameters and hull parameters of the bow inclination, superstructure layout, parameters of the ship's ice class, ship draft and draft difference, and propeller dimensions; the environmental element information includes hydrological and meteorological environmental elements of ice, wind, and waves; and the dynamic element information includes the relative position and motion relationship between the ship and the environmental elements, propeller speed, and rudder angle.

[0018] Furthermore, S13 specifically includes the following steps:

[0019] S131, according to the uniform distribution sampling method, generate random numbers from the input and output sample sets X and Y to perform sampling, and extract the first group of data x in step 2 1 and 1 , the second step is to extract the second set of data x 2 and 2 , and so on, extract data x in the kth step k and k , construct a training sample set U∈R k×n and V∈R k×l ;

[0020] S132, constructing a radial basis function neural network using the input and output training sample sets U and V, and performing regression calculation of the hidden layer response matrix and the output using the partial least squares method:

[0021] In the RBF network, the Gaussian function is used as the radial basis function, and the vector u j (j=1,…,k) becomes the hidden node center; the hidden layer response matrix A∈R is obtained by radial basis function calculation. k×k , where the element in row i and column j is:

[0022]

[0023] Among them, u i is the i-th sample, where ||·|| represents the Euclidean distance, c j and σ j are the center and width of the jth basis function respectively;

[0024] S133, performing PLS principal component extraction and regression operation between the hidden layer response matrix A and the output matrix V;

[0025] After extracting the principal component T, project A and V onto the principal component matrix respectively, and we get:

[0026] V=TR+F=AWR+F

[0027] Among them, T is the principal component matrix of A, which is k×n T dimensional matrix; W is the transformation matrix of A, which is k×n T dimensional matrix; R is n T ×l-dimensional regression coefficient matrix; F is the k×l-dimensional residual matrix;

[0028] Get the generalized model:

[0029]

[0030] Among them A p1 and W 1are k×k and k×n respectively. T dimensional matrix, R 1 n T ×l-dimensional matrix, generalize the obtained network to the samples in the set U, and set the accuracy threshold to determine the number of principal component extraction times n T , when the change of the generalized root mean square error is less than the set accuracy change threshold e 1 Stop extracting principal components when:

[0031]

[0032] S134, generalize the obtained network to the training data in the entire data set X, and calculate the generalization error for the entire data set:

[0033]

[0034] in:

[0035]

[0036] S135, if RMSE O Greater than the set accuracy threshold e 2 , then repeat steps S231 to S234;

[0037] S136, if RMSE O Less than the set accuracy threshold e 2 , then it ends;

[0038] S137, the number of principal components obtained n T The number of loop steps k is the number of principal component extractions and the number of hidden nodes of the final radial basis function network, and the vectors in the set U are the hidden nodes of the obtained radial basis function network.

[0039] Furthermore, S2 specifically includes:

[0040] According to the coverage of ice areas, wind fields, wave fields and flow fields selected according to the sea areas through which the ship's planned route passes, a high-resolution ice area ice condition and hydrological and meteorological environment forecast model is selected, and the ice condition information obtained by identifying the ship's measured video and image information and the real-time observed hydrological and meteorological environment data information are assimilated. The spatiotemporal generative adversarial neural network with an attention mechanism is used to realize the assimilation and fusion of the observed data and the high-resolution model forecast data, thereby improving the accuracy of the hydrological and meteorological forecast in the navigation area;

[0041] The principal component analysis method is used to reduce the dimension of meteorological field data and remove redundant noise signals, retain the main component feature vectors of the meteorological field, reconstruct the main feature vectors after dimensionality reduction and denoising as meteorological field data, and build a marine environment forecast model based on deep separable convolution; the deep separable convolution module is used to improve the prediction efficiency and performance, while reducing the calculation amount and hardware memory requirements of the model; the principal component analysis method reduces the data dimension and redundant noise effect to efficiently filter the evolution characteristics of external meteorological elements and improve the forecast accuracy;

[0042] The principal component analysis method is used to clean and reconstruct the external wind wave current and sea ice coverage area data through empirical orthogonal decomposition theory. The principal component analysis method reduces the data dimension and redundant noise effect, and effectively filters the evolution characteristics of meteorological elements to improve the forecast accuracy.

[0043] The calculation formula of the principal component analysis method is as follows:

[0044]

[0045] Among them, Meteroloty m,t represents the corresponding wind, wave, current and ice cover tensor field at time t, PC is the principal component variable, EOF is the corresponding spatial feature vector, t represents the external ice, wind and wave forecast field information of the sea area where the ship passes at time t, n=1,…k represents the number of principal component PCs, and m represents the number of tensors;

[0046] A deep separable 2D CNN layer is used to build a deep encoder-decoder. The 2D convolution kernel in the model is used to mine the coupling interaction characteristics between the tensor factors of the external meteorological environment field. The decoder module exports high-dimensional semantic information, and the encoder part extracts high-dimensional feature maps of tensor sequences with low-order nonlinear information. The integrated decoder-encoder constructs a spatiotemporal generative adversarial network model generator based on the deep separable convolution architecture.

[0047] Create a mathematical forecasting model of ice field, wind field, wave field and flow field in the sea area where the ship passes through, which is based on the deep separable module architecture of deep generative adversarial neural network. The model uses deep separable convolution as the skeleton and adopts the spatiotemporal generative adversarial neural network model with the introduction of attention mechanism to realize the assimilation and fusion of observation data and high-resolution model forecast data, and dynamically update the forecast of external ice conditions, wind and wave fields to obtain the forecast results of the marine environment in the sea area where the ship passes through.

[0048] The spatiotemporal generative adversarial network model includes two modules, a generator G(Z) and a discriminator D(x), wherein both the generator and the discriminator modules are constructed using a deep separable convolutional skeleton, the generator module G(Z) assimilates and fuses the observation data and the multimodal data of the model forecast field reconstructed after dimensionality reduction and denoising by the principal component analysis method, wherein Z represents the observation point information and the meteorological field; the meteorological field map generated by the generator module G and the real map decoded in the reanalysis data are fed back to the discriminator D, the input of the discriminator D is the generated meteorological field data and the reanalysis meteorological field data reconstructed after dimensionality reduction and denoising; the output of D is a scalar, which is used to judge the probability of whether the output map comes from the real input map; the stopping criterion for the training of the spatiotemporal generative adversarial network model is that the discriminator cannot distinguish whether the generated meteorological field forecast map conforms to the real map;

[0049] The loss function of the spatiotemporal generative adversarial network model is as follows:

[0050]

[0051] Among them, L D is the discriminator loss, L G is the generator loss function.

[0052] Furthermore, the external marine meteorological environment field forecast model parameters are trained and updated based on the following minimization and maximization formulas:

[0053]

[0054] Among them, P data (x) is the distribution of reanalysis meteorological field mapping data, P z (z) is the model input observation and the numerical simulation meteorological field data distribution reconstructed after dimensionality reduction and denoising by principal component analysis, and E represents the expectation operator;

[0055] The number of features of the time series sequential tensor reconstructed after dimensionality reduction and denoising of the multimodal meteorological factors input to the spatiotemporal generative adversarial neural network model by principal component analysis is C, H is the height of the model input tensor, W is the width of the model input tensor, Represents the input tensor of the model. The attention mechanism module uses dot-product to further calculate the three matrices K, Q and V of the input tensor to obtain three The corresponding mapping weight feature matrix value;

[0056] The dimensions of the K and Q feature matrices should be consistent, and the i-th query eigenvalue in the iterative operation is and the jth key eigenvalue The similarity between them is determined by the regularization function Calculate acquisition;

[0057] Assume that the shipboard instruments measure the external ice conditions, wind, current and wave meteorological elements respectively as S ice ,S wind ,S current ,S wave The ERA5 high-resolution numerical model forecasts ice conditions, wind, current and wave two-dimensional field elements after dimensionality reduction and denoising by principal component analysis, respectively: PCA-ERA ice ,PCA-ERA wind ,PCA-ERA current ,PCA-ERA wave ,Dynamically updated sea condition forecast model for ships sailing in the sea area:

[0058] P ice,t =STGAN(S ice,t ,PCAERA ice,t )

[0059] P wind,t =STGAN(S wind,t ,PCA-ERA wind,t )

[0060] P current,t =STGAN(S curren,t ,PCA-ERA current,t )

[0061] P wave,t =STGAN(S wave,t ,PCA-ERA wave,t )

[0062] Among them, STGAN is a spatiotemporal generative adversarial neural network model, and P*,t represents the external ice condition, wind and wave forecast information of the sea area where the ship passes at time t.

[0063] Furthermore, S4 specifically includes the following steps:

[0064] The empirical mode decomposition method is used to decompose the time series of sea ice thickness variables:

[0065] I = IMF 1 +IMF 2 +…+IMF j +…+IMF n +r

[0066] For each component IMF j Both use neural networks to identify and predict them;

[0067] Specifically, for the IMF jA q-step-ahead forecast is performed, and the input order of the prediction model is determined to be p using the Lipschitz quotient method. Then the structure of the q-step-ahead forecast is p-input-1-output, and the input and output of the neural network identification process are:

[0068]

[0069] Among them, p j is the number of neural network inputs to the jth IMF component, which is determined by the Lipschitz quotient method based on the data to increase the adaptability of the method;

[0070] The input and output in the prediction process are:

[0071]

[0072] The order of time series forecast for each component is determined by calculating the Lipschitz coefficient method. After the input and output orders are determined, the variable structure radial basis function neural network is used to predict the time series of each component.

[0073] Furthermore, the steps of establishing the variable structure neural network prediction model are as follows:

[0074] S41. Establishing a sliding data window

[0075] In order to reflect the latest ship motion status, a sliding data window is established to observe the ship motion status, and the fitting model based on the radial basis function neural network is dynamically adjusted using the real-time updated input and output data;

[0076] The sliding window is a fixed-width first-in-first-out data sample sequence. When a new set of input-output data is received, the new data set is added to the sliding window, and the earliest set of data is removed from the sliding window. The sliding window W at time t is SD It is expressed as:

[0077] W SD =[(p t-L+1 ,q t-L+1 ),L,(p t ,q t )]

[0078] Where L is the width of the sliding window;

[0079] The sliding window is adjusted in real time to reflect the latest movement of the ship and the response of the structure. The sliding window adjustment method based on time series segmentation is adopted, that is, the sliding window input with a width of L is segmented into a time series, and a sliding data window with an adaptive width is formed from the earliest data point to the mutation point, and the window width is kept not less than the set minimum width Lmin;

[0080] For the time series \(x(t)\) within a sliding window consisting of \(L\) data inputs, calculate the mean \(\mu\) of the data before each data point 1 (i) and the standard deviation \(s\) 1 (i), as well as the mean \(\mu\) of the data after 2 (i) and the standard deviation \(s\) 2 (i), then the combined deviation \(S\) at point \(i\) D is:

[0081]

[0082] where \(L\) 1 , \(L\) 2 are the number of points in the data part before and after point \(i\) respectively. Use the t-test value \(T(i)\) to represent the difference in the means of the two parts of data before and after point \(i\):

[0083]

[0084] Repeat the above calculation for each data in the sequence \(x(t)\) to obtain a sequence of statistical test values \(T(t)\) corresponding to the data values of \(x(t)\), and calculate the maximum value \(T\) in \(T(t)\) max of the statistical significance \(P(T\) max ):

[0085] \(P(T\) max ) = Prob(T ≤ T max )

[0086] where \(P(T\) max ) represents the probability of obtaining a \(T\) value less than or equal to \(T\) max in a random process. Generally, \(P(T\) max ) can be approximately expressed as:

[0087]

[0088] The values of \(\eta\) and \(\delta\) are obtained by Monte Carlo simulation. \(L\) is the sliding window width, i.e., the length of \(x(t)\), \(v = L - 2\), \(I\) x (a, b) is the incomplete beta function. Set the \(P\) 0 threshold. If \(P(T\) max ) ≥ \(P\) 0 , the two sub - sequences obtained by splitting \(x(t)\) are different sub - sequences, and this point is a mutation point;

[0089] Form a new sliding data window \(L_a\) from the mutation point and the earliest data point within the sliding window. If the width \(L_a\) of the sliding data window is less than the set minimum width \(L_{min}(L_a < L_{min})\), then the sliding data window is \(L_{min}\);

[0090] Using the input-output data set in the sliding window, that is, using the input matrix P and the corresponding output vector Q to represent the real-time dynamics of the mapping relationship:

[0091]

[0092] Q=[q t-L+1 L t ]

[0093] Where n p is the dimension of the input matrix;

[0094] The input matrix P and the corresponding output Q are used as the input and output of the radial basis function neural network respectively to train and dynamically adjust the neural network;

[0095] S42, after receiving a new data sample at each step, update the sliding data window, add the latest sample to the window, delete the earliest sample from the window, and directly add the new data sample to the hidden layer as a new hidden node;

[0096] S43, calculate the response matrix Φ of the hidden layer, where:

[0097]

[0098] Among them, c j is the center of the jth hidden node, p i is the i-th sample, ||·|| represents the Euclidean distance, σ is the basis function width; M is the number of hidden nodes, and the vectors in Φ are orthogonally decomposed using the Gram-Schmidt rule. Φ=WA, and W=[w 1 ,L,w M ];

[0099] Calculate the error reduction rate:

[0100]

[0101] Standardized error reduction rate:

[0102]

[0103] S44, select those hidden nodes whose sum of output contribution is less than the set value, until the sum of the standardized error reduction rates of the selected hidden nodes is Select k 1 ,...,k S Construct a set of hidden nodes to be deleted

[0104] Take the past continuous M S The intersection I of the hidden node sets selected in the step and delete the hidden nodes in I:

[0105]

[0106] S45. After the hidden nodes are determined at each step, the connection weights from the hidden layer to the output layer are updated using the least squares method:

[0107] Θ=Φ + Y=(Φ T Φ) -1 Φ T Y

[0108] Neural network predictions for residual values It is obtained by reorganizing the predicted values ​​of each component:

[0109]

[0110] Thus, a neural network capable of online prediction is obtained.

[0111] The present invention also provides a storage medium, which includes a stored program, wherein when the program is run, any of the above-mentioned methods for rapid prediction of environmental response of ships sailing in ice areas is executed.

[0112] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes any of the above-mentioned methods for rapid prediction of environmental response of ships sailing in ice areas through the computer program.

[0113] Compared with the prior art, the present invention has the following advantages:

[0114] The present invention uses an offline neural network fitting model to calculate the motion response and structural response of the ship to the environment to increase the rapidity of the ship response calculation; comprehensively considers the motion response and structural response of the ship's motion when considering the risk of the ship sailing in ice areas; in the establishment of the neural network, the partial least squares method is used to solve the problem of too high input dimension, and a simple network structure is selected by gradually increasing the number of network nodes to improve the generalization ability of the obtained neural network. Differentiated processing is performed on the wind field, wave field and flow field.

[0115] When a ship is sailing in ice areas, ice loads not only have a great impact on the ship's motion posture, but also pose a great threat to the ship's structure. Therefore, this patent divides environmental responses into motion responses and structural responses. In view of the many unexpected situations when a ship is sailing in ice areas, such as frequent turns and speed changes of the ship, and frequent interactions with ice blocks, the ship's motion state and structural load will change dramatically accordingly. In this case, the sliding window needs to be adjusted in real time to reflect the latest motion of the ship and the response of the structure. The present invention adopts a sliding window adjustment method based on time series segmentation.

[0116] The structural response and motion response of ships to ice have their own unique difficulties. The motion and structural response of ships to wind, waves and ice are caused by wind and wave loads and ice loads. However, ice loads are different from wave and wind loads. Wave and wind elements can be directly measured to calculate wind and wave loads, but ice loads cannot be directly measured for the time being. Therefore, the present invention inverts ice loads based on the measured data of the structural response of the ship.

[0117] The present invention uses environmental forecast information to perform multi-time scale response forecasting. On the basis of dynamically updating and forecasting external ice, wind, waves, currents and other ice zone environmental information every 24 hours, a sliding window-based neural network is used to perform multi-scale decomposition and forecasting within 24 hours, which can identify and extract long-term, medium-term and short-term change information of environmental factors and accurately forecast environmental factors.

[0118] In the ice zone hydrometeorological prediction method, compared with the traditional generative adversarial model, the present invention introduces a dual attention mechanism module to map the high-dimensional feature information containing multiple meteorological factors to the multi-dimensional feature matrix of the encoder and decoder. And the basic skeleton of the generative adversarial model is replaced with a two-dimensional convolutional neural network module for directly inputting the two-dimensional ice zone environmental element field information. A 2D deep CNN layer is used to build a deep encoder and decoder of the generative adversarial model. The 2D convolution kernel is used to mine the deep coupling interaction characteristics between various meteorological factors; the deep convolution block of this method can provide a method to improve the prediction efficiency and performance, while reducing its computational and memory requirements, and can better perform feature extraction and aggregation. The encoder part extracts the high-dimensional feature map of the tensor sequence with low-order nonlinear information, and the decoder module derives high-dimensional semantic information; improves the forecast accuracy of data fusion field elements. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0120] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0121] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0122] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0123] like Figure 1 As shown, the present invention provides a method for rapid prediction of environmental response of ships sailing in ice areas. The method makes full use of the existing forecast information of ice area environmental factors and the actual measured ice area environmental factors and ship motion response of the ship, and performs the numerical simulation of ship motion and structural response of ships sailing in ice areas with large calculation amount offline, and uses neural network to fit the database, and adopts hierarchical prediction in the forecast link of ice area environmental factors to ensure the stability of prediction while improving the accuracy of real-time prediction. This method can quickly and accurately obtain the motion and structural response of the ship under the influence of the complex environment of the ice area.

[0124] (1) Establish a mathematical model of the ship's motion response and structural response to environmental disturbances such as ice, wind, and waves at sea.

[0125] First, set the training model input and model output. Set the model training input as ① information on ship elements such as the main dimensions, structural parameters and bow inclination, superstructure layout, ship ice class, ship draft and draft difference, propeller dimensions, etc.; ② information on parameters of hydrometeorological environmental elements such as ice, wind, and waves; ③ dynamic information such as the relative position and motion relationship between the ship and environmental elements, propeller speed, and rudder angle. Set the training model output as the ship's speed, motion posture, and structural response under the influence of ice, wind, and waves. Secondly, establish the model input-output training samples by numerically simulating the movement of the ship in the ice, and calculate the response of a specific ship to environmental elements in terms of ship speed, motion posture, and structure. Finally, based on the input and output data of the obtained samples, train the neural network offline so that the neural network accurately approximates the input-output mapping, and establishes a neural network mapping model for the motion and structural response model of the ship's navigation. Specifically,

[0126] Suppose there are m samples in the sample set, the input vector x has n independent variables, the input X is an m×n dimensional matrix, the output vector y has l dependent variables, and the output Y is an m×l dimensional matrix.

[0127] The algorithm steps for establishing the neural network calculation model of ship marine motion response are as follows:

[0128] I. Starting from step 1, in step k, use the uniform distribution sampling method to generate random numbers from the input and output sample sets X and Y to implement sampling and extract a set of data x from the population. k and k , respectively construct training sample sets U∈R k×n and V∈R k×l .

[0129] II. Use the input and output training sample sets U and V to construct the RBF neural network, and use the partial least squares method to perform regression calculations on the hidden layer response matrix and output:

[0130] In the RBF network, the Gaussian function is used as the radial basis function, and the vector u j (j=1,…,k) becomes the hidden node center. The hidden layer response matrix A∈R is obtained by radial basis function calculation. k×k . The element in row i and column j is:

[0131]

[0132] where u i is the i-th sample, where ||·|| represents the Euclidean distance, c j and σ j are the center and width of the jth basis function respectively.

[0133] III. Perform PLS principal component extraction and regression operations between the response matrix A and the output matrix V.

[0134] After extracting the principal component T, project A and V onto the principal component matrix respectively, and we get:

[0135] V=TR+F=AWR+F

[0136] Among them, T is the principal component matrix of A, which is k×n T dimensional matrix; W is the transformation matrix of A, which is k×n T dimensional matrix; R is n T ×l-dimensional regression coefficient matrix; F is the k×l-dimensional residual matrix.

[0137] Get the generalized model:

[0138]

[0139] Among them A p1 and W 1 are k×k and k×n respectively. T dimensional matrix. 1 n T ×l-dimensional matrix. The obtained network is generalized to the samples in the set U, and an accuracy threshold is set to determine the number of principal component extractions n T , when the change of the generalized root mean square error (RMSE) is less than the set accuracy change threshold e 1 Stop extracting principal components when:

[0140]

[0141] IV. Generalize the obtained network to the training data in the entire dataset X and calculate the generalization error for the entire dataset:

[0142]

[0143] in

[0144]

[0145] V. If RMSE O Greater than the accuracy threshold e 2 , then repeat steps I-IV.

[0146] VI. Algorithm in RMSE O Less than the accuracy threshold e 2 When it ends.

[0147] The number of principal components obtained is n TThe number of loop steps k is the number of principal component extractions and the number of hidden nodes of the final RBF network, and the vectors in the set U are the hidden nodes of the obtained RBF network.

[0148] The principal component analysis method is introduced to construct a marine environment forecast model based on deep separable convolution; the deep separable convolution module can improve the prediction efficiency and performance, while reducing the model's computational workload and hardware memory requirements. Compared with traditional convolution, deep separable convolution greatly reduces the number of model parameters and computational requirements. This is highly practical and feasible for realizing shipboard weather forecasts when ships are sailing at sea. The principal component analysis method can effectively filter the evolution characteristics of external meteorological elements and improve forecast accuracy by reducing data dimensions and redundant noise effects.

[0149] The principal component analysis method is used in the data preprocessing module of the marine environmental element forecast module to perform preliminary dimensionality reduction and denoising preprocessing on the collected meteorological data. After the meteorological data is processed by principal component analysis, it is used as the input of the external marine environmental field forecast model.

[0150] A principal component analysis method is introduced to clean and reconstruct the external wind wave current and sea ice coverage area data. The principal component analysis method can effectively filter the evolution characteristics of meteorological elements and improve the forecast accuracy by reducing data dimensions and redundant noise effects.

[0151] The reconstruction calculation formula of the meteorological element tensor based on empirical orthogonal decomposition is as follows:

[0152]

[0153] Among them, PC is the principal component variable, EOF is the corresponding spatial eigenvector, t represents the external ice, wind and wave forecast field information of the sea area where the ship passes through at time t, n=1,…k represents the number of principal component PCs, and m represents the number of tensors.

[0154] Using a deep separable 2D CNN layer, a deep encoder-decoder is built, and a deep separable convolutional neural network model is constructed. The 2D convolution kernel in the model is used to mine the coupling interaction characteristics between the tensor factors of the extreme wind field. The decoder module exports high-dimensional semantic information, and the encoder part extracts high-dimensional feature maps of tensor sequences with low-order nonlinear information. The deep separable convolution block introduced in this method can significantly reduce the model calculation and memory requirements. Compared with traditional convolution, the number of parameters and calculations required is greatly reduced. This makes them more efficient and faster to train and run on devices with limited computing resources.

[0155] A mathematical forecasting model based on deep generative adversarial neural network is created for the ice field, wind field, wave field and flow field in the sea area through which the ship passes. The model uses deep separable convolution as the skeleton and adopts the spatiotemporal generative adversarial neural network model with attention mechanism to realize the assimilation and fusion of observation data and high-resolution model forecast data, and dynamically update the forecast of external ice conditions, wind, waves and flow fields to obtain the forecast results of the marine environment of the sea area through which the ship passes.

[0156] The marine environment element forecasting module of the present invention is a deep generative adversarial neural network model, and the convolution skeleton in the model adopts deep separable convolution. The decoder and encoder are two submodules of the generator in the generative adversarial network model.

[0157] The spatiotemporal generative adversarial network model includes two modules, a generator G(Z) and a discriminator D(x), wherein both the generator and the discriminator modules are constructed using a deep separable convolutional skeleton, the generator module G(Z) assimilates and fuses the observation data and the multimodal data of the model forecast field reconstructed after dimensionality reduction and denoising by the principal component analysis method, wherein Z represents the observation point information and the meteorological field; the meteorological field map generated by the generator module G and the real map decoded in the reanalysis data are fed back to the discriminator D, the input of the discriminator D is the generated meteorological field data and the reanalysis meteorological field data reconstructed after dimensionality reduction and denoising; the output of D is a scalar, which is used to judge the probability of whether the output map comes from the real input map; the stopping criterion for the training of the spatiotemporal generative adversarial network model is that the discriminator cannot distinguish whether the generated meteorological field forecast map conforms to the real map;

[0158] The loss function of the spatiotemporal generative adversarial network model is as follows:

[0159]

[0160] Among them, L D is the discriminator loss, L G is the generator loss function;

[0161] The generative adversarial network model is the external ocean environment meteorological field forecast model in the present invention, where S2 is the model's skeleton and structure creation module, and S3 is the model's forecast application module.

[0162] The parameters of the external marine meteorological environment forecast model are trained and updated based on the following minimization and maximization formulas:

[0163]

[0164] Among them, P data (x) is the distribution of reanalysis meteorological field mapping data, P z (z) is the model input observation and the numerical simulation meteorological field data distribution reconstructed after dimensionality reduction and denoising by principal component analysis, and E represents the expectation operator;

[0165] The number of features of the time series sequential tensor reconstructed after dimensionality reduction and denoising of the multimodal meteorological factors input to the spatiotemporal generative adversarial network model by principal component analysis is C, H is the height of the model input tensor, W is the width of the model input tensor, Represents the input tensor of the model. The attention mechanism module uses dot-product to further calculate the three matrices K, Q and V of the input tensor to obtain three The corresponding mapping weight feature matrix value;

[0166] The dimensions of the K and Q feature matrices should be consistent, and the i-th query eigenvalue in the iterative operation is and the jth key eigenvalue The similarity between them is determined by the regularization function Calculate acquisition;

[0167] Furthermore, it is assumed that the shipboard instruments measure the external ice conditions, wind, current and wave meteorological elements respectively as S ice ,S wind ,S current ,S wave , the ERA5 high-resolution numerical model forecasts ice conditions, wind, current and wave two-dimensional field elements after dimensionality reduction and denoising by principal component analysis, respectively PCA-ERA ice ,PCA-ERA wind ,PCA-ERA current ,PCA-ERA wave ,Dynamically updated sea condition forecast model for ships sailing in the sea area:

[0168] P ice,t =STGAN(S ice,t ,PCAERA ice,t )

[0169] P wind,t =STGAN(S wind,t ,PCA-ERA wind,t )

[0170] P current,t =STGAN(S curren,t ,PCA-ERA current,t )

[0171] P wave,t =STGAN(S wave,t ,PCA-ERA wave,t )

[0172] Among them, STGAN is a spatiotemporal generative adversarial neural network model, and P*,t represents the external ice condition, wind and wave forecast information of the sea area where the ship passes at time t.

[0173] Based on the generative adversarial ocean environment forecast model created above, accurate forecast results of the ocean environment such as ice coverage, wind field, wave field and flow field of the sea area through which the ship passes are obtained in real time and dynamically, providing real-time guidance for ship route planning.

[0174] (2) Based on the real-time measured ice conditions and wind, wave, current information as well as ship motion and structural response information, the database of environmental motion and structural response of ships sailing in ice areas is enriched, or the database is updated based on the measured information of wind, wave, current and other elements and the ice load inversion information of ship structural response;

[0175] The ship's motion and structural response information is obtained through ship stress measurement sensors (fiber grating sensors). Because it is difficult to directly measure ice loads, ice loads are indirectly inverted through the measurement data of local or overall responses of the ship's structure, such as ship structural strain and ship's six-degree-of-freedom motion parameters.

[0176] (3) Provide detailed real-time forecasts of ice environment factors in waters where ships are sailing;

[0177] According to the planned route of the ship and the forecast results of the regional ice environment factors, a refined real-time forecast of the ice environment factors in the ship route area is made hourly within 24 hours. In the forecast process, multi-level decomposition and prediction are adopted, the forecast is decomposed into different time scales for forecasting, and different forecasts are summed to obtain the final forecast result. Different environmental factors are forecasted separately, and the sea ice thickness is used as an example to illustrate as follows:

[0178] The empirical mode decomposition (EMD) method is used to decompose the time series of sea ice thickness variables:

[0179] I = IMF 1 +IMF 2 +…+IMF j +…+IMF n +r

[0180] For each component IMF j Neural networks are used to identify and predict them.

[0181] Specifically, for the IMF j A q-step-ahead forecast is performed, and the input order of the prediction model is determined to be p using the Lipschitz quotient method. Then the structure of the q-step-ahead forecast is p-input-1-output, and the input and output of the neural network identification process are:

[0182]

[0183] where p jis the number of neural network inputs to the jth IMF component, which is determined by the Lipschitz quotient method based on the data to increase the adaptability of the method.

[0184] The input and output in the prediction process are:

[0185]

[0186] The order of time series prediction for each component is determined by calculating the Lipschitz coefficient and other methods. After the input and output orders are determined, the variable structure radial basis function neural network is used to predict the time series of each component.

[0187] The variable structure neural network prediction model is established through the following steps:

[0188] ① Establish a sliding data window

[0189] The movement of ships at sea has dynamic and time-varying characteristics. In order to reflect the latest ship movement status, a sliding data window is established to observe the ship movement status, and the fitting model based on the radial basis function neural network is dynamically adjusted using the real-time updated input and output data;

[0190] The sliding window is a fixed-width, first-in, first-out data sample sequence. When a new set of input-output data is received, the new data set is added to the sliding window, and the earliest set of data is removed from the sliding window. SD It is expressed as:

[0191] W SD =[(p t-L+1 ,q t-L+1 ),L,(p t ,q t )]

[0192] Where L is the width of the sliding window;

[0193] In view of the many unexpected situations when ships are sailing in ice areas, such as frequent ship turns and speed changes, frequent interactions with ice blocks, etc., the ship's motion state and structural load will change dramatically accordingly. In this case, the sliding window needs to be adjusted in real time to reflect the latest movement of the ship and the response of the structure. The present invention adopts a sliding window adjustment method based on time series segmentation, that is, the sliding window input with a width of L is segmented into a time series, and a sliding data window with an adaptive width is formed from the earliest data point to the variation point, and the window width is kept not less than the set minimum width Lmin.

[0194] For the time series x(t) in the sliding window consisting of L data inputs, calculate the average value μ of the data before each data point 1(i) and the standard deviation s 1 (i), and the mean value μ of the subsequent data 2 (i) and the standard deviation s 2 (i), then the combined deviation S at point i D (i) is

[0195]

[0196] where L 1 , L 2 are the number of points in the data before and after point i respectively, and the difference in the mean values of the two parts of data before and after point i is represented by the t-test value T(i):

[0197]

[0198] Repeat the above calculation for each data in the sequence x(t) to obtain the sequence of statistical test values T(t) corresponding to the data values of x(t), and calculate the statistical significance P(Tmax) of the maximum value Tmax in T(t):

[0199] P(T max ) = Prob(T ≤ T max )

[0200] where P(Tmax) represents the probability of obtaining a T value less than or equal to Tmax in the random process. Generally, P(Tmax) can be approximately expressed as

[0201]

[0202] The values of η and δ can be obtained by Monte Carlo simulation. L is the sliding window width, i.e., the length of x(t), v = L - 2, and Ix(a,b) is the incomplete beta function. Set the P0 threshold. If P(Tmax) ≥ P0, the two sub-sequences obtained by splitting x(t) are different sub-sequences, and this point is a mutation point.

[0203] Form a new sliding data window La from the mutation point and the earliest data point in the sliding window. If the width La of the sliding data window is less than the set minimum width Lmin (La < Lmin), then the sliding data window is Lmin.

[0204] Use the input-output data group within the sliding window, that is, represent the real-time dynamics of the mapping relationship with the input matrix P and the corresponding output vector Q respectively:

[0205]

[0206] Q = [q t-L+1 L q t ​

[0207] Where n p is the dimension of the input matrix;

[0208] The input matrix P and the corresponding output Q are used as the input and output of the radial basis function neural network respectively to train and dynamically adjust the neural network.

[0209] ② After receiving a new data sample at each step, update the sliding data window, add the latest sample to the window, and delete the earliest sample from the window. Directly add the new data sample to the hidden layer as a new hidden node.

[0210] ③ Calculate the response matrix Φ of the hidden layer, where

[0211]

[0212] where c j is the center of the jth hidden node, p i is the i-th sample, ||·|| represents the Euclidean distance, σ is the basis function width; M is the number of hidden nodes. The vectors in Φ are orthogonally decomposed using the Gram-Schmidt rule Φ = WA, and W = [w 1 ,L,w M ].

[0213] Calculate the error reduction rate:

[0214]

[0215] Standardized error reduction rate:

[0216]

[0217] ③ Select those hidden nodes whose sum of contribution to the output is less than the set value. Until the sum of the standardized error reduction rates of the selected hidden nodes is Select k 1 ,...,k S Construct a set of hidden nodes to be deleted

[0218] Take the past continuous M S The intersection I of the hidden node sets selected in the step and delete the hidden nodes in I:

[0219]

[0220] ④ After the hidden nodes are determined at each step, for convenience, the least squares method is used to update the connection weights from the hidden layer to the output layer.

[0221] Θ=Φ + Y=(Φ T Φ) -1 ΦT Y.

[0222] Neural network predictions for residual values It is obtained by reorganizing the predicted values ​​of each component:

[0223]

[0224] Thus, a neural network capable of online prediction is obtained.

[0225] (4) According to the sea area through which the ship's planned route passes, calculate the motion response and structural response of the ship passing through the ice area, wind field and wave field in the sea area. The input of the calculation model is ① the main dimensions of the ship and the hull parameters such as the bow inclination angle, superstructure layout, ice class of the ship and other parameters, current ship draft and draft difference, propeller size and other ship elements; ② the location of the ship and the parameters of ice, wind, waves and other hydrological and meteorological environmental elements in the future; ③ the relative position and relative motion relationship between the ship and ice, wind, waves and other environmental elements, propeller speed, rudder angle and other dynamic information. The training model output is set to be the speed, motion posture and structural response of the ship under the influence of the set ice, wind and waves.

[0226] Based on the above-mentioned ship's speed, motion attitude and other motion responses and differential structural responses under the influence of ice, wind and waves, combined with the prediction results of the flow field, the response forecast of the ship's navigation in the ice area is comprehensively carried out. On this basis, real-time risk forecast of ship navigation in ice areas and real-time route optimization can be carried out.

[0227] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. 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. However, 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 rapid prediction method for environmental response of ships sailing in ice areas, characterized in that: The steps include: S1. Establish a mathematical model of the ship's motion response and ship structure response under the environmental interference of ice, wind and waves at sea; S2. Establish a forecast model for ice and sea conditions in the sea area where ships sail, use a spatiotemporal generative adversarial neural network model with an attention mechanism to achieve assimilation and fusion of observation data and high-resolution model forecast data, dynamically update the forecast of external ice, wind, wave and flow fields, and obtain the initial marine environment forecast results for the sea area where the ship passes; S3. Based on the real-time measured ice conditions, wind, wave, current information, and ship motion and structural response information, the environmental motion and structural response database of ships sailing in ice areas is enriched, or based on the measured information of wind, wave and current elements and the ice load inversion information of ship structural response, the ice conditions and sea state forecast model database of ships sailing in the sea area is updated to obtain real-time marine environment forecast results; S4. According to the planned route of the ship and the real-time marine environment forecast results, a refined real-time forecast of ice environment elements is made for the ship route area on an hourly basis within 24 hours. In the forecast process, multi-level decomposition and prediction are adopted, the forecast is decomposed into different time scales for forecasting, and different forecasts are summed to obtain the final forecast result; S5. According to the sea area through which the ship's planned route passes, the motion response and structural response of the ship passing through the ice area, wind field and wave field of the sea area are calculated by using the motion response and structural response mathematical models; The input of the mathematical model of motion response and structural response includes ship element information, environmental element information and dynamic element information, and the output of the mathematical model of motion response and structural response is the speed, motion attitude and structural response of the ship under the influence of set ice, wind and waves; S6. Based on the motion response and structural response of the ship such as speed and motion posture under the influence of ice, wind and waves, combined with the final forecast results, a comprehensive forecast of the response of the ship sailing in ice areas is carried out.

2. The rapid prediction method for environmental response of ships sailing in ice areas according to claim 1 is characterized in that: S1 specifically includes the following steps: S11, setting the input of the training model and the output of the training model; the input of the training model includes ship element information, environmental element information and dynamic element information; the output of the training model includes the speed, motion posture and structural response of the ship under the influence of ice, wind and waves; S12. Based on the model input and model output, a model input-output training sample is established by numerically simulating the motion and structural response of the ship in ice, and the response of a specific ship to environmental factors in terms of ship speed, motion posture and structure is calculated; assuming that there are m samples in the sample set, the input vector x has n independent variables, the input X is an m×n dimensional matrix, the output vector y has l dependent variables, and the output Y is an m×l dimensional matrix; S13. Based on the input and output data of the samples obtained in S12, offline training of the neural network makes the neural network accurately approximate the input-output mapping, and establishes a neural network mapping model of the motion and structural response model of the ship's navigation.

3. The rapid prediction method for environmental response of ships sailing in ice areas according to claim 2 is characterized in that: The ship element information includes the main dimensions of the ship, structural parameters and hull parameters of the bow inclination, superstructure layout, parameters of the ship's ice class, ship draft and draft difference, and propeller dimensions; the environmental element information includes hydrological and meteorological environmental elements of ice, wind, and waves; the dynamic element information includes the relative position and movement relationship between the ship and the environmental elements, propeller speed, and rudder angle.

4. The rapid prediction method for environmental response of ships sailing in ice areas according to claim 2 is characterized in that: S13 specifically includes the following steps: S131, according to the uniform distribution sampling method, generate random numbers from the input and output sample sets X and Y to perform sampling, extract the first group of data x1 and y1 in the second step, extract the second group of data x2 and y2 in the second step, and so on, extract data x in the kth step. k and k , construct a training sample set U∈R k×n and V∈R k×l ; S132, constructing a radial basis function neural network using the input and output training sample sets U and V, and performing regression calculation of the hidden layer response matrix and the output using the partial least squares method: In the RBF network, the Gaussian function is used as the radial basis function, and the vector u j (j=1,…,k) becomes the hidden node center; the hidden layer response matrix A∈R is obtained by radial basis function calculation. k×k , where the element in row i and column j is: Among them, u i is the i-th sample, where ||·|| represents the Euclidean distance, c j and σ j are the center and width of the jth basis function respectively; S133, performing PLS principal component extraction and regression operation between the hidden layer response matrix A and the output matrix V; After extracting the principal component T, project A and V onto the principal component matrix respectively, and we get: V=TR+F=AWR+F Among them, T is the principal component matrix of A, which is k×n T dimensional matrix; W is the transformation matrix of A, which is k×n T dimensional matrix; R is n T ×l-dimensional regression coefficient matrix; F is the k×l-dimensional residual matrix; Get the generalized model: Among them, A p1 and W1 are k×k and k×n respectively. T dimensional matrix, R1 is n T ×l-dimensional matrix, generalize the obtained network to the samples in the set U, and set the accuracy threshold to determine the number of principal component extraction times n T , stop extracting the principal components when the change in the generalized root mean square error is less than the set accuracy change threshold e1: S134, generalize the obtained network to the training data in the entire data set X, and calculate the generalization error for the entire data set: in: S135, if RMSE O If it is greater than the set accuracy threshold e2, steps S231 to S234 are repeated; S136, if RMSE O If it is less than the set accuracy threshold e2, then the process ends; S137, the number of principal components obtained n T The number of loop steps k is the number of principal component extractions and the number of hidden nodes of the final radial basis function network, and the vectors in the set U are the hidden nodes of the obtained radial basis function network.

5. The rapid prediction method for environmental response of ships sailing in ice areas according to claim 1 is characterized in that: S2 specifically includes: According to the coverage of ice areas, wind fields, wave fields and flow fields selected according to the sea areas through which the ship's planned route passes, a high-resolution ice area ice condition and hydrological and meteorological environment forecast model is selected, and the ice condition information obtained by identifying the ship's measured video and image information and the real-time observed hydrological and meteorological environment data information are assimilated. The spatiotemporal generative adversarial neural network with an attention mechanism is used to realize the assimilation and fusion of the observed data and the high-resolution model forecast data, thereby improving the accuracy of the hydrological and meteorological forecast in the navigation area; The principal component analysis method is used to reduce the dimension of meteorological field data and remove redundant noise signals, retain the main component feature vectors of the meteorological field, reconstruct the main feature vectors after dimensionality reduction and denoising as meteorological field data, and build a marine environment forecast model based on deep separable convolution; the deep separable convolution module is used to improve the prediction efficiency and performance, while reducing the calculation amount and hardware memory requirements of the model; the principal component analysis method reduces the data dimension and redundant noise effect to efficiently filter the evolution characteristics of external meteorological elements and improve the forecast accuracy; The principal component analysis method is used to clean and reconstruct the external wind wave current and sea ice coverage area data through empirical orthogonal decomposition theory. The principal component analysis method reduces the data dimension and redundant noise effect, and effectively filters the evolution characteristics of meteorological elements to improve the forecast accuracy. The calculation formula of the principal component analysis method is as follows: Among them, Meterology m,t represents the corresponding wind, wave, current and ice cover tensor field at time t, PC is the principal component variable, EOF is the corresponding spatial feature vector, t represents the external ice, wind and wave forecast field information of the sea area where the ship passes at time t, n=1,…k represents the number of principal component PCs, and m represents the number of tensors; A deep separable 2D CNN layer is used to build a deep encoder-decoder. The 2D convolution kernel in the model is used to mine the coupling interaction characteristics between the tensor factors of the external meteorological environment field. The decoder module exports high-dimensional semantic information, and the encoder part extracts high-dimensional feature maps of tensor sequences with low-order nonlinear information. The integrated decoder-encoder constructs a spatiotemporal generative adversarial network model generator based on the deep separable convolution architecture. Create a mathematical forecasting model of ice field, wind field, wave field and flow field in the sea area where the ship passes through, which is based on the deep separable module architecture of deep generative adversarial neural network. The model uses deep separable convolution as the skeleton and adopts the spatiotemporal generative adversarial neural network model with the introduction of attention mechanism to realize the assimilation and fusion of observation data and high-resolution model forecast data, and dynamically update the forecast of external ice conditions, wind and wave fields to obtain the forecast results of the marine environment in the sea area where the ship passes through. The spatiotemporal generative adversarial network model includes two modules, a generator G(Z) and a discriminator D(x), wherein both the generator and the discriminator modules are constructed using a deep separable convolutional skeleton, the generator module G(Z) assimilates and fuses the observation data and the multimodal data of the model forecast field reconstructed after dimensionality reduction and denoising by the principal component analysis method, wherein Z represents the observation point information and the meteorological field; the meteorological field map generated by the generator module G and the real map decoded in the reanalysis data are fed back to the discriminator D, the input of the discriminator D is the generated meteorological field data and the reanalysis meteorological field data reconstructed after dimensionality reduction and denoising; the output of D is a scalar, which is used to judge the probability of whether the output map comes from the real input map; the stopping criterion for the training of the spatiotemporal generative adversarial network model is that the discriminator cannot distinguish whether the generated meteorological field forecast map conforms to the real map; The loss function of the spatiotemporal generative adversarial network model is as follows: Among them, L D is the discriminator loss, L G is the generator loss function.

6. The rapid prediction method for environmental response of ships sailing in ice areas according to claim 5 is characterized in that: The parameters of the external marine meteorological environment forecast model are trained and updated based on the following minimization and maximization formulas: Among them, P data (x) is the distribution of reanalysis meteorological field mapping data, P z (z) is the model input observation and the numerical simulation meteorological field data distribution reconstructed after dimensionality reduction and denoising by principal component analysis, and E represents the expectation operator; The number of features of the time series sequential tensor reconstructed after dimensionality reduction and denoising of the multimodal meteorological factors input to the spatiotemporal generative adversarial neural network model by principal component analysis is C, H is the height of the model input tensor, W is the width of the model input tensor, Represents the input tensor of the model. The attention mechanism module uses dot-product to further calculate the three matrices K, Q and V of the input tensor to obtain three The corresponding mapping weight feature matrix value; The dimensions of the K and Q feature matrices should be consistent, and the i-th query eigenvalue in the iterative operation is and the jth key eigenvalue The similarity between them is determined by the regularization function Calculate acquisition; Assume that the shipboard instruments measure the external ice conditions, wind, current and wave meteorological elements respectively as S ice , S wind , S current , S wave The ERA5 high-resolution numerical model forecasts ice conditions, wind, current and wave two-dimensional field elements after dimensionality reduction and denoising by principal component analysis, respectively: PCA-ERA ice , PCA-ERA wind , PCA-ERA current , PCA-ERA wave , dynamically updated sea condition forecast model for ships sailing in the sea area: P ice,t =STGAN(S ice，t ,PCAERA ice，t ) P wind，t =STGAN(S wind，t ,PCA-ERA wind，t ) P current，t =STGAN(S curren，t ,PCA-ERA current，t ) P wave，t =STGAN(S wave，t ,PCA-ERA wave，t ) Among them, STGAN is a spatiotemporal generative adversarial neural network model, and P*,t represents the external ice condition, wind and wave forecast field information of the sea area where the ship passes at time t.

7. The rapid prediction method for environmental response of ships sailing in ice areas according to claim 1 is characterized in that: S4 specifically includes the following steps: The empirical mode decomposition method is used to decompose the time series of sea ice thickness variables: I=IMF1+IMF2+…+IMF j +…+IMF n +r For each component IMF j Both use neural networks to identify and predict them; Specifically, for the IMF j A q-step-ahead forecast is performed, and the input order of the prediction model is determined to be p using the Lipschitz quotient method. Then the structure of the q-step-ahead forecast is p-input-1-output, and the input and output of the neural network identification process are: Among them, p j is the number of neural network inputs to the jth IMF component, which is determined by the Lipschitz quotient method based on the data to increase the adaptability of the method; The input and output in the prediction process are: The order of time series forecast for each component is determined by calculating the Lipschitz coefficient method. After the input and output orders are determined, the variable structure radial basis function neural network is used to predict the time series of each component.

8. The rapid prediction method for environmental response of ships sailing in ice areas according to claim 7 is characterized in that: The steps to establish a variable structure neural network prediction model are as follows: S41. Establishing a sliding data window In order to reflect the latest ship motion status, a sliding data window is established to observe the ship motion status, and the fitting model based on the radial basis function neural network is dynamically adjusted using the real-time updated input and output data; The sliding window is a fixed-width first-in-first-out data sample sequence. When a new set of input-output data is received, the new data set is added to the sliding window, and the earliest set of data is removed from the sliding window. The sliding window W at time t is SD It is expressed as: W SD =[(p t-L+1 ,q t-L+1 ),…,(p t ,q t )] Where L is the width of the sliding window; The sliding window is adjusted in real time to reflect the latest movement of the ship and the response of the structure. The sliding window adjustment method based on time series segmentation is adopted, that is, the sliding window input with a width of L is segmented into a time series, and a sliding data window with an adaptive width is formed from the earliest data point to the mutation point, and the window width is kept not less than the set minimum width Lmin; For a time series x(t) in a sliding window consisting of L data inputs, the mean μ1(i) and standard deviation s1(i) of the data before each data point, as well as the mean μ2(i) and standard deviation s2(i) of the data after each data point are calculated, then the combined deviation S of point i is D (i) is: Among them, L1 and L2 are the number of points before and after point i, respectively. The t-test value T(i) represents the difference in the mean of the two parts of data before and after point i: Repeat the above calculation for each data in the sequence x(t) to obtain the statistical test value sequence T(t) corresponding to the data value of x(t), and calculate the maximum value T in T(t). max The statistical significance of P(T max ): P(T max )=Prob(T≤T max ) Among them, P(T max ) indicates that the value of T is less than or equal to T in the random process max The probability of max ) can be approximately expressed as: The values ​​of η and δ are obtained by Monte Carlo simulation, where L is the width of the sliding window, i.e., the length of x(t), v = L-2, and I x (a, b) is an incomplete β function, set the P0 threshold, if P(T max )≥P0, if there is a subsequence that is different between the two subsequences into which x(t) is divided, then this point is a mutation point; A new sliding data window La is formed by the mutation point and the earliest data point within the sliding window. If the width La of the sliding data window is less than the set minimum width Lmin (La < Lmin), the sliding data window is Lmin; Using the input-output data group within the sliding window, that is, the real-time dynamics of the mapping relationship are represented by the input matrix P and the corresponding output vector Q respectively; Q=[q t-L+1 …q t ] Where n p is the dimension of the input matrix; The input matrix P and the corresponding output Q are used as the input and output of the radial basis function neural network respectively to train and dynamically adjust the neural network; S42. After receiving a new data sample at each step, update the sliding data window, add the latest sample to the window, and delete the earliest sample from the window. Directly add the new data sample to the hidden layer as a new hidden node; S43. Calculate the response matrix Φ of the hidden layer, where: Among them, c j is the center of the jth hidden node, p i is the i-th sample, ||·|| represents the Euclidean distance, σ is the basis function width; M is the number of hidden nodes, and the vectors in Φ are orthogonally decomposed using the Gram-Schmidt rule. Φ=WA, and we get W=[w1,…,w M ]; Calculate the error descent rate: Normalize the error descent rate: S44, select those hidden nodes whose sum of output contribution is less than the set value, until the sum of the standardized error reduction rates of the selected hidden nodes is Choose k1,...,k S Construct a set of hidden nodes to be deleted Take the past continuous M S The intersection I of the hidden node sets selected in the step and delete the hidden nodes in I: I={S k ∩S k-1 ∩…∩S k-MS+1 } S45. After determining the hidden node at each step, use the least squares method to update the connection weights from the hidden layer to the output layer: I = F + Y=(Φ T F) -1 F T Y Neural network predictions for residual values It is obtained by reorganizing the predicted values ​​of each component: Thus, a neural network capable of online prediction is obtained.

9. A storage medium, characterized in that: The storage medium includes a stored program, wherein when the program runs, it executes the rapid prediction method for the environmental response of an ice-going ship described in any one of claims 1 to 8.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The processor runs through the computer program to execute the rapid prediction method for the environmental response of an ice-going ship described in any one of claims 1 to 8.

Images