Sea fog and low cloud coverage two-dimensional distribution field prediction method and system

Abstract

The application discloses a sea fog and low cloud coverage two-dimensional distribution field prediction method and system, relates to the technical field of artificial intelligence, and constructs a sea fog and low cloud space-time change two-dimensional field prediction analysis model based on a U-shaped deep two-dimensional convolutional neural network, adopts a 2D CNN layer, builds a deep encoding decoder, establishes a U-shaped deep two-dimensional convolutional neural network model, introduces a residual learning mechanism, extracts and retains potential physical disturbance information in the original high-dimensional mapping space of each variable, introduces a multi-head attention mechanism module, maps high-dimensional feature information containing multiple meteorological factors to a multi-dimensional feature matrix of the encoding decoder, and uses a trained model to predict two-dimensional sea fog field information. The application can dynamically predict the space-time change of the spatial distribution of sea fog and low clouds, predict and master the birth and death changes of sea fog and low clouds in advance, can provide a large perspective sea fog and low cloud distribution prediction for water, land and air transportation, and further guarantees the safety of transportation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for forecasting the two-dimensional distribution field of sea fog and low cloud cover. Background Technology

[0002] The appearance of sea fog drastically reduces both vertical and horizontal visibility, obstructing human vision. It poses serious safety hazards to maritime construction operations, water transportation, and coastal air and land traffic, potentially leading to major accidents. Statistics show that 80% of collisions between ships at sea are caused by reduced visibility due to sea fog. Furthermore, the formation and spread of sea fog also impacts agricultural development in coastal areas. However, statistics indicate that the identification and forecasting of sea fog remains at a relatively low level compared to other hazardous weather events. Therefore, intelligent monitoring and forecasting of the formation and dissipation of sea fog along nearshore coasts is of significant guiding and practical importance for maritime disaster prevention and mitigation. Sea fog occurs in different seasons along my country's coastal areas from south to north. The formation and dissipation of sea fog involve complex atmospheric and oceanic physical processes and their interactions, including turbulent transport, droplet deposition, radiation, fog top entrainment, and wind shear. Existing atmospheric and oceanic numerical models often struggle to accurately characterize the complex boundary layers involved in the formation and dissipation of sea fog.

[0003] Currently, the main methods for monitoring and forecasting sea fog include meteorological station observations and satellite remote sensing. Meteorological observation stations typically provide high-precision, continuous single-point visibility data. However, observations from sparse stations along the coast and at sea, such as buoys, cannot meet the needs of large-scale sea fog forecasting and monitoring on a spatiotemporal scale. Furthermore, the spatiotemporal distribution patterns of sea fog are highly complex, and using observation stations to roughly represent the area is insufficient to accurately characterize the heterogeneity of sea fog in continuous space, introducing significant uncertainty to regional-scale forecasting analysis. The continuous spatial coverage of remote sensing network observation technology provides reliable data support for the spatial distribution of sea fog. However, existing research has not yet established a high-precision estimation model for the two-dimensional field forecasting of sea fog that incorporates multi-source heterogeneous reanalysis data. In addition, the parameterization of theoretical sea fog models is extremely complex, making the construction of numerical models very difficult. Moreover, numerical models often fail to adequately consider the interactive coupling effects of multimodal factors, easily leading to the loss of spatiotemporal heterogeneous coupling information of meteorological and hydrological elements.

[0004] Although modern watercraft and aircraft are mostly equipped with advanced radar and other environmental monitoring and navigation equipment, maritime and air accidents caused by sea fog still occur frequently, resulting in considerable losses. Sea fog is indeed a very harmful marine hazard. First, the extreme reduction in visibility caused by sea fog severely impacts water transport, air transport, fishing and aquaculture, and marine production activities. Second, sea fog refracts sunlight, greatly reducing the amount of sunshine on the Earth's surface, and the resulting low temperatures and high humidity can severely damage crop growth. Third, the water droplets in sea fog have a high salt content, which has a significant corrosive effect on structures and equipment on the water.

[0005] Domestic and international research on sea fog and low cloud forecasting based on machine learning and deep learning mainly focuses on sea fog identification, such as using convolutional neural networks and various ensemble machine learning methods to identify sea fog occurrences. Recent research on sea fog based on convolutional neural networks also primarily focuses on the forecasting analysis of the probability or frequency of sea fog and low cloud occurrences. It fails to integrate and analyze the spatiotemporal characteristics of sea fog and low cloud formations, nor does it delve into the spatiotemporal heterogeneity of sea fog and low cloud distribution and the multimodal, multi-factor spatiotemporal coupling factors influencing their formation and dissipation. Furthermore, it neglects the coupling and continuity between accurately identifying sea fog formation and dissipation and the subsequent reconstruction and forecasting of sea fog spatiotemporal distribution characteristics. Summary of the Invention

[0006] In view of this, this invention provides a method and system for forecasting the two-dimensional distribution field of sea fog and low cloud cover. Utilizing multi-source heterogeneous reanalysis data, and fully considering the multimodal and multi-factor factors influencing the formation and dissipation of sea fog, a U-shaped encoder-decoder based on residual learning and multi-head attention mechanisms is constructed. Furthermore, based on the sum of multiple forecasting factors for sea fog, the spatiotemporal dynamic distribution of the two-dimensional field of sea fog and low clouds is predicted. This fills a research gap in the spatiotemporal forecasting of two-dimensional field information for sea fog and low clouds.

[0007] Therefore, the present invention provides the following technical solution:

[0008] This invention provides a method for forecasting the two-dimensional distribution field of sea fog and low cloud cover, comprising the following steps:

[0009] S1. Collect and summarize the two-dimensional field information of various multimodal meteorological factors that affect the formation and spatial variation of sea fog, and select the two-dimensional field distribution information of fog and low clouds in the corresponding sea area.

[0010] S2. Using the correlation function method, calculate the autocorrelation coefficients of each meteorological factor with the two-dimensional field of fog and low clouds, and obtain the weight factors of each meteorological factor; select effective meteorological factors based on the weight factors.

[0011] S3. Construct a two-dimensional field forecast analysis model for the spatiotemporal variation of sea fog and low clouds based on a U-shaped deep two-dimensional convolutional neural network. The model takes spatial meteorological factors as input and outputs the two-dimensional field forecast results for the spatiotemporal variation of sea fog and low clouds. The model construction includes: using 2D CNN layers, building a deep encoder-decoder, and constructing a U-shaped deep two-dimensional convolutional neural network model. The 2D convolutional kernel is used to mine the deep coupling and interaction characteristics between various meteorological factors. The encoder part extracts high-dimensional feature maps of tensor sequences with low-order nonlinear information, while the decoder module derives high-dimensional semantic information. A residual learning mechanism is introduced into the U-shaped deep two-dimensional convolutional neural network model. The residual learning mechanism controls the feature extraction and fidelity fusion of the model information flow, extracts potential physical disturbance information and retains it in the original high-dimensional mapping space of each variable. The output information flow of each residual module is the sum of the original information flow and the output information flow after convolution mapping. A multi-head attention mechanism module is further introduced into the U-shaped deep two-dimensional convolutional neural network model. The multi-head attention mechanism module maps the high-dimensional feature information containing multiple meteorological factors to the multi-dimensional feature matrix of the encoder-decoder.

[0012] S4. The model is trained using cross-validation, and the trained model is used to predict two-dimensional sea fog field information.

[0013] Furthermore, various multimodal meteorological factors that influence the formation and spatial variation of sea fog include: wind field, surface temperature field, air pressure field, humidity field, and air density field.

[0014] Furthermore, S1 specifically includes:

[0015] S11. Collect and summarize the time series spatial variation information of various multimodal meteorological factors in the selected area, and summarize the low cloud coverage range of the selected area as the target variable for the potential spatial variation distribution of sea fog.

[0016] S12. Use Z-Score standardization to standardize the data for tensors below coverage and various meteorological factor tensors.

[0017] Furthermore, S2 includes the following steps:

[0018] S21. Establish the time series correlation function and partial autocorrelation function;

[0019] S22. Calculate the time and space mean series of each meteorological factor, and use the correlation analysis function to calculate the correlation weight between each factor and the target low cloud cover factor.

[0020] S23. Remove negatively correlated weighted factors and retain positively correlated meteorological factor variables.

[0021] Furthermore, the multi-head attention mechanism module maps high-dimensional feature information containing multiple meteorological factors to a multi-dimensional feature matrix of the encoder and decoder. This includes: for the input tensor that superimposes and fuses multiple meteorological factors, three aggregated attention feature matrices are projected using cubic linear projection. The value, query, and key matrices of the three projection matrices are computed in parallel, and finally, a projection weight matrix containing deep high-level semantics and features is calculated.

[0022] Furthermore, the Huber loss function is used as the loss function of the prediction model to calculate the error between the predicted value and the target value:

[0023]

[0024] Wherein, Ψ represents the two-dimensional field forecast data of sea fog and low clouds created in this invention, and O represents the reanalysis data of sea fog and low clouds.

[0025] Furthermore, S4 specifically includes:

[0026] S41. Use the K-fold cross-validation method to divide the dataset into several equal parts;

[0027] S42. The standardized meteorological factor data and sea fog and low cloud datasets are divided into two parts: a training set and a test set.

[0028] S43. Use the K-fold method to divide the training data into K parts; of which K-1 parts are used to train the model, and the remaining 1 part is used to verify the model's prediction performance.

[0029] S44. Repeat step S43 until the model's performance prediction is validated for each data set.

[0030] S45. Using Huber loss error as the standard, calculate the forecast performance of K models in step S43, and use the average forecast performance of the models as the final forecast reference.

[0031] This invention also provides a two-dimensional distribution field forecasting system for sea fog and low cloud cover, comprising:

[0032] The data acquisition module performs regional interception and collection of two-dimensional field information of various multimodal meteorological factors that affect the formation and spatial variation of sea fog, and selects the two-dimensional field distribution information of fog and low clouds in the corresponding sea area.

[0033] The data processing module uses the correlation function method to calculate the autocorrelation coefficients of each meteorological factor with the two-dimensional field of fog and low clouds, and obtains the weight factors of each meteorological factor; effective meteorological factors are then selected based on the weight factors.

[0034] The model construction module builds a two-dimensional field forecast analysis model for the spatiotemporal variation of sea fog and low clouds based on a U-shaped deep two-dimensional convolutional neural network. The model takes spatial meteorological factors as input and outputs the two-dimensional field forecast results for the spatiotemporal variation of sea fog and low clouds. The model construction includes: using 2D CNN layers to build a deep encoder-decoder and constructing a U-shaped deep two-dimensional convolutional neural network model. The 2D convolutional kernels are used to mine the deep coupling and interaction characteristics between various meteorological factors. The encoder extracts high-dimensional feature maps of tensor sequences with low-order nonlinear information, while the decoder module derives high-dimensional semantic information. A residual learning mechanism is introduced into the U-shaped deep two-dimensional convolutional neural network model. This mechanism controls the feature extraction and fidelity fusion of the model's information flow, extracting and preserving potential physical disturbance information in the original high-dimensional mapping space of each variable. The output information flow of each residual module is the sum of the original information flow and the output information flow after convolutional mapping. A multi-head attention mechanism module is further introduced into the U-shaped deep two-dimensional convolutional neural network model. This multi-head attention mechanism module maps the high-dimensional feature information containing multiple meteorological factors to the multi-dimensional feature matrix of the encoder-decoder.

[0035] The forecasting module uses cross-validation to train the model and then uses the trained model to forecast two-dimensional sea fog field information.

[0036] Advantages and positive effects of the present invention:

[0037] Current research on sea fog and low cloud forecasting focuses on judging and identifying their occurrence probability. This invention, however, establishes a two-dimensional field forecasting analysis of the spatiotemporal changes of sea fog and low cloud based on a deep learning model. It dynamically forecasts the spatial distribution of sea fog and low cloud over time, allowing for early prediction and understanding of the formation and dissipation of sea fog and low cloud. This provides a broader perspective for forecasting the distribution of sea fog and low cloud for water, land, and air transportation, further ensuring transportation safety.

[0038] In this invention's forecast model, a U-shaped encoder-decoder maps two-dimensional field information of remote multimodal meteorological factors to a high-dimensional nonlinear space, deeply characterizing the spatiotemporal nonlinear coupling and interaction of multimodal meteorological and hydrological elements affecting the formation and dissipation of sea fog. Based on model parameter weight matching calculations, data source information is filtered and controlled, achieving convenient and efficient end-to-end data fusion calculations. While ensuring the quality of the forecast information field, it significantly saves computational resources and improves forecast accuracy. Since it does not require complex physical and mathematical formula calculations and system simulations, the intelligent computational forecast model overcomes the computational bottleneck of traditional weather forecasting methods. The residual learning module in the model aggregates the processed output by adding the identity information of one neural layer to the output of the previous layers, preserving more original spatial variables. The attention block introduced in the model further fuses and refines low-level and high-level spatiotemporal sequence features, significantly improving the accuracy of the forecast model. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a model for predicting the spatiotemporal variations of sea fog and low clouds in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the forecasting process in an embodiment of the present invention. Detailed Implementation

[0042] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0043] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0044] like Figure 1-2 As shown in the figure, an embodiment of the present invention provides a method for forecasting the two-dimensional distribution field of sea fog and low cloud cover based on residual learning and multi-head attention mechanism, which includes the following steps:

[0045] S1. Collect and summarize the two-dimensional field information of various multimodal meteorological factors that affect the formation and spatial variation of sea fog, and select the two-dimensional field distribution information of fog and low clouds in the corresponding sea area.

[0046] The information includes two-dimensional field data of various multimodal meteorological factors influencing sea fog formation and spatial variation, such as wind field, surface temperature field, air pressure field, humidity field, and air density field. Specifically, S1 includes:

[0047] S11. Using the European Centre for Medium- and Long-Term Forecasting (ECMWF) reanalysis dataset ERA5, we selected a portion of the coastal waters and landmass of China as the overall study area. We collected and summarized the temporal and spatial variation information of the wind field, pressure field, air density field, temperature field, and air humidity field at a 2-meter depth on the ground surface of the selected area. We also summarized the low cloud cover range of the selected study area as the target variable for the potential spatial variation distribution of sea fog.

[0048] S12. Z-Score standardization is used to standardize the data of tensors below coverage and various meteorological factor tensors, further eliminating calculation errors caused by different dimensions.

[0049]

[0050] Where σ is the standard deviation of the sample variable, and μ represents the time mean of the sample variable.

[0051] S2. Using the correlation function method, calculate the autocorrelation coefficients of each meteorological factor with the two-dimensional field of fog and low clouds, and obtain the weight factors of each meteorological factor; select effective meteorological factors based on the weight factors.

[0052] In its specific implementation, S2 includes the following steps:

[0053] S21. Establish the time series correlation function and partial autocorrelation function;

[0054] S22. Calculate the temporal and spatial mean series of meteorological fields such as surface wind field, temperature field, pressure field, humidity field and air density field, and use the correlation analysis function to calculate the correlation weight between each factor and the target low cloud cover factor.

[0055] The formula for calculating the correlation coefficient between multimodal factors is as follows:

[0056]

[0057] S23. Remove negatively correlated weighted factors and retain positively correlated meteorological factor variables.

[0058] S3. Construct an encoding and decoding framework based on a deep two-dimensional convolutional neural network. In this framework, a residual learning mechanism is introduced as a basic module for automatic feature mining and automatic filtering of the encoder and decoder. A multi-head attention mechanism is introduced into the bottleneck layer of the encoder and decoder and the last layer of the decoding module to further mine and fuse the spatiotemporal sequence features of various meteorological factor fields, sea fog and low cloud fields in low dimension and high latitude, as well as their potential coupling and interaction effects.

[0059] In its specific implementation, S3 includes the following steps:

[0060] S31. A U-shaped deep two-dimensional convolutional neural network model is constructed using 2D CNN layers and a deep encoder-decoder architecture. The 2D convolutional kernels in the model are used to mine the deep coupling and interaction characteristics between various meteorological factors. The encoder extracts high-dimensional feature maps of tensor sequences with low-order nonlinear information, while the decoder module derives high-dimensional semantic information. The U-shaped deep two-dimensional convolutional neural network model is as follows: Figure 1 As shown.

[0061] S32. The residual learning mechanism is introduced into the U-shaped deep two-dimensional convolutional neural network model. The residual learning mechanism controls the feature extraction and fidelity fusion of the model information flow, and further extracts and retains the potential physical disturbance information in the original high-dimensional mapping space of each variable.

[0062] f(x)′=f(x)+x (3)

[0063] The output information stream of each residual module is the sum of the original information stream and the output information stream after convolution mapping.

[0064] S33. A multi-head attention mechanism module is further introduced into the U-shaped degree encoding and decoding model, and a schematic diagram of the multi-head attention mechanism module is attached. Figure 1 The attention module is shown in the figure.

[0065] The number of features of the time-series sequential tensor of the input multimodal meteorological factors in the encoder-decoder model is C, which is the number of two-dimensional fields of the meteorological factors. H and W are the height and width of the model input tensor. The input tensor of the model represents the key K, the query Q, and the value V. The attention mechanism module uses dot-product to further compute the three matrices of the input tensor, obtaining three... The corresponding mapping weight feature matrix values.

[0066]

[0067]

[0068]

[0069] The dimensions of the K and Q feature matrices should be consistent. The i-th query feature value in the iterative operation... and the j-th key feature value The similarity between them is determined by the regularization function. Calculated and obtained. This involves calculating the similarity between different convolutional network layers, such as at different locations. and The similarity calculations are different, therefore these similarity values ​​cannot maintain symmetry. The attention weights at different positions in the feature vector are calculated using the following formula:

[0070]

[0071] Where d k This represents the dimension of the query and the key matrix. A high-dimensional input tensor can cause the gradient of the softmax regularization function to drop sharply, leading the model into local optima. The d introduced in the attention weight calculation formula above... k Factors can alleviate and further reduce gradient vanishing.

[0072] The weight matrix of the i-th row in the input tensor is calculated according to formula (7):

[0073]

[0074] To further simplify:

[0075]

[0076] Furthermore, depending on the different regularization functions, formula (9) can be unified as follows:

[0077]

[0078] Wherein, the regularization function φ(q) i ,k i Calculate q i and k i Similarity of related features between them.

[0079] In the calculation of attention feature weights, it is usually necessary to introduce a specific constraint factor ker(x, y) into the above regularization function: This is used to ensure the non-negativity of the attention computation operation. Based on the characteristic indicator function that introduces the constraint factor, equation (10) can be expressed as:

[0080]

[0081] When the numerator in the formula is represented in vector form:

[0082]

[0083] The linear attention mechanism in the high-dimensional feature extraction calculation of multimodal meteorological factors can reduce computational resource consumption and improve model computational efficiency. In this invention, for the input tensor of multiple superimposed and fused meteorological factors, three aggregated attention feature matrices are projected using cubic linear projection. The value, query, and key matrices of the three projection matrices are calculated in parallel. Finally, a projection weight matrix containing deep high-level semantics and features is obtained. The high-dimensional feature information containing multiple meteorological factors is mapped to the multi-dimensional feature matrix of the encoder and decoder through a multi-head attention mechanism.

[0084] S34. Using the Huber loss function as the loss function for the prediction model, calculate the error between the predicted value and the target value:

[0085]

[0086] Wherein, Ψ represents the two-dimensional field forecast data of sea fog and low clouds created in this invention, and O represents the reanalysis data of sea fog and low clouds, which is "observation data".

[0087] S4. Cross-validation is used to train the encoder-decoder model, and the trained model is used to predict two-dimensional sea fog field information.

[0088] S41. Use the K-fold cross-validation method to divide the dataset into several equal parts.

[0089] S42. The standardized meteorological factor data and sea fog and low cloud datasets are divided into two parts: a training set and a test set.

[0090] S43. Use the K-fold method to divide the training data into K parts; of which K-1 parts are used to train the model, and the remaining 1 part is used to verify the model's prediction performance.

[0091] S44. Repeat step S43 until the model's performance prediction is validated for each data set.

[0092] S45. Using Huber loss error as the standard, calculate the forecast performance of K models in step S43, and use the average forecast performance of the models as the final forecast reference.

[0093] In the above embodiments, this invention establishes a two-dimensional field forecasting and analysis model for the spatiotemporal variation of sea fog and low clouds based on a deep learning model. This model dynamically forecasts the spatial distribution of sea fog and low clouds over time, allowing for early prediction and understanding of their formation and dissipation. This provides a broad perspective for forecasting sea fog and low cloud distribution in water, land, and air transportation, further ensuring transportation safety. The forecasting model uses a U-shaped encoder-decoder to map the two-dimensional field information of multimodal meteorological factors into a high-dimensional nonlinear space, deeply characterizing the spatiotemporal nonlinear coupling and interaction of multimodal meteorological and hydrological elements affecting the formation and dissipation of sea fog. Based on model parameter weight matching calculations, data source information is filtered and controlled, achieving convenient and efficient end-to-end data fusion calculations. While ensuring the quality of the forecast information field, it fully conserves computational resources and improves forecast accuracy. Since it does not require complex physical and mathematical formula calculations and system simulations, the intelligent computational forecasting model overcomes the computational bottleneck of traditional weather forecasting methods. The residual learning module in the model aggregates the processed output by adding the identity information of one neural layer to the output of the previous layers, preserving more original spatial variables. The attention blocks introduced in the model further integrate and refine low-level and high-level spatiotemporal sequence features, significantly improving the accuracy of the forecast model.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 forecasting the two-dimensional distribution field of sea fog and low cloud cover, characterized in that, Includes the following steps: S1. Collect and summarize the two-dimensional field information of various multimodal meteorological factors that affect the formation and spatial variation of sea fog, and select the two-dimensional field distribution information of fog and low clouds in the corresponding sea area. S2. Using the correlation function method, calculate the autocorrelation coefficients of each meteorological factor with the two-dimensional field of fog and low clouds, and obtain the weight factors of each meteorological factor; select effective meteorological factors based on the weight factors. S3. Construct a two-dimensional field forecast analysis model for the spatiotemporal variation of sea fog and low clouds based on a U-shaped deep two-dimensional convolutional neural network. The model takes space meteorological factors as input and outputs the two-dimensional field forecast results for the spatiotemporal variation of sea fog and low clouds. The model construction includes: employing 2D CNN layers, building a deep encoder-decoder, and constructing a U-shaped deep two-dimensional convolutional neural network model. The 2D convolutional kernels are used to mine the deep coupling and interaction characteristics between various meteorological factors. The encoder extracts high-dimensional feature maps of tensor sequences with low-order nonlinear information, while the decoder module derives high-dimensional semantic information. A residual learning mechanism is introduced into the U-shaped deep two-dimensional convolutional neural network model. This mechanism controls the feature extraction and fidelity fusion of the model's information flow, extracting and preserving potential physical disturbance information in the original high-dimensional mapping space of each variable. The output information flow of each residual module is the sum of the original information flow and the output information flow after convolutional mapping. A multi-head attention mechanism module is further introduced into the U-shaped deep two-dimensional convolutional neural network model. Through this multi-head attention mechanism module, high-dimensional feature information containing multiple meteorological factors is mapped to the multi-dimensional feature matrix of the encoder-decoder. S4. The model is trained using cross-validation, and the trained model is used to predict two-dimensional sea fog field information.

2. The method for forecasting the two-dimensional distribution field of sea fog and low cloud cover according to claim 1, characterized in that, Various multimodal meteorological factors that influence the formation and spatial variation of sea fog include: wind field, surface temperature field, air pressure field, humidity field, and air density field.

3. The method for forecasting the two-dimensional distribution field of sea fog and low cloud cover according to claim 2, characterized in that, S1 specifically includes: S11. Collect and summarize the time series spatial variation information of various multimodal meteorological factors in the selected area, and summarize the low cloud coverage range of the selected area as the target variable for the potential spatial variation distribution of sea fog. S12. Use Z-Score standardization to standardize the data for tensors below coverage and various meteorological factor tensors.

4. The method for forecasting the two-dimensional distribution field of sea fog and low cloud cover according to claim 3, characterized in that, S2 includes the following steps: S21. Establish the time series correlation function and partial autocorrelation function; S22. Calculate the time and space mean series of each meteorological factor, and use the correlation analysis function to calculate the correlation weight between each factor and the target low cloud cover factor. S23. Remove negatively correlated weighted factors and retain positively correlated meteorological factor variables.

5. The method for forecasting the two-dimensional distribution field of sea fog and low cloud cover according to claim 4, characterized in that, The multi-head attention mechanism module maps high-dimensional feature information containing multiple meteorological factors to a multi-dimensional feature matrix of the encoder and decoder. This includes: for the input tensor that superimposes and fuses multiple meteorological factors, using cubic linear projection to generate three aggregated attention feature matrices, and performing parallel computation on the value, query, and key matrices of the three projection matrices, finally calculating a projection weight matrix containing deep high-level semantics and features.

6. A method for forecasting the two-dimensional distribution field of sea fog and low cloud cover according to claim 5, characterized in that, The Huber loss function is used as the loss function for the prediction model to calculate the error between the predicted value and the target value: ； in, This represents two-dimensional field forecast data for sea fog and low clouds, where O represents reanalysis data for sea fog and low clouds.

7. The method for forecasting the two-dimensional distribution field of sea fog and low cloud cover according to claim 6, characterized in that, S4 specifically includes: S41. Use the K-fold cross-validation method to divide the dataset into several equal parts; S42. The standardized meteorological factor data and sea fog and low cloud datasets are divided into two parts: a training set and a test set. S43. Use the K-fold method to divide the training data into K parts; of which K-1 parts are used to train the model, and the remaining 1 part is used to verify the model's prediction performance. S44. Repeat step S43 until the model's performance prediction is validated for each data set. S45. Using Huber loss error as the standard, calculate the forecast performance of K models in step S43, and use the average forecast performance of the models as the final forecast reference.

8. A two-dimensional distribution field forecasting system for sea fog and low cloud cover, characterized in that, include: The data acquisition module performs regional interception and collection of two-dimensional field information of various multimodal meteorological factors that affect the formation and spatial variation of sea fog, and selects the two-dimensional field distribution information of fog and low clouds in the corresponding sea area. The data processing module uses the correlation function method to calculate the autocorrelation coefficients of each meteorological factor with the two-dimensional field of fog and low clouds, and obtains the weight factors of each meteorological factor; effective meteorological factors are then selected based on the weight factors. The model building module constructs a two-dimensional field forecast analysis model for the spatiotemporal variation of sea fog and low clouds based on a U-shaped deep two-dimensional convolutional neural network. The model takes space meteorological factors as input and outputs the two-dimensional field forecast results for the spatiotemporal variation of sea fog and low clouds. The model construction includes: employing 2D CNN layers, building a deep encoder-decoder, and constructing a U-shaped deep two-dimensional convolutional neural network model. The 2D convolutional kernels are used to mine the deep coupling and interaction characteristics between various meteorological factors. The encoder extracts high-dimensional feature maps of tensor sequences with low-order nonlinear information, while the decoder module derives high-dimensional semantic information. A residual learning mechanism is introduced into the U-shaped deep two-dimensional convolutional neural network model. This mechanism controls the feature extraction and fidelity fusion of the model's information flow, extracting and preserving potential physical disturbance information in the original high-dimensional mapping space of each variable. The output information flow of each residual module is the sum of the original information flow and the output information flow after convolutional mapping. A multi-head attention mechanism module is further introduced into the U-shaped deep two-dimensional convolutional neural network model. Through this multi-head attention mechanism module, high-dimensional feature information containing multiple meteorological factors is mapped to the multi-dimensional feature matrix of the encoder-decoder. The forecasting module uses cross-validation to train the model and then uses the trained model to forecast two-dimensional sea fog field information.

Images