Sea fog prediction method and device

Abstract

The invention relates to the field of meteorological prediction, and discloses a sea fog prediction method and device, and the method comprises the following steps: S1, obtaining multi-source heterogeneous data; s2, generating a scene representation and confidence coefficient weight matrix for representing the macroscopic physical background; s3, calculating compound loss; s4, updating parameters of the neural network model based on the composite loss. The device comprises a data acquisition module; a scene representation generation module; a confidence coefficient matrix generation module; a space-time prediction module; a composite loss calculation module; a parameter optimization module; a computer program for executing the scheme is stored in the storage medium. Through a dynamic adjustment mechanism of a macroscopic scene and a mesoscale process, the model can deeply understand how a long-term climate background affects the generation and elimination process of local sea fog, and the accuracy of sea fog prediction and the generalization ability of the model are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather prediction, in particular to a sea fog prediction method and device. BACKGROUND

[0002] Sea fog forecasting is a key link to ensure the safety of marine activities and has important value to improve shipping efficiency and safety. At present, the main prediction technology paths are mainly divided into two categories. One is the numerical weather prediction model based on physical process, which simulates the complete physical process of sea fog by solving fluid mechanics and thermodynamics equation set. The other is a data-driven deep learning method, which uses convolutional neural network or recurrent neural network model to automatically learn the complex spatio-temporal characteristics and evolution law of sea fog from massive historical satellite images and meteorological reanalysis data.

[0003] However, with the increasing demand for prediction refinement and reliability, the existing technical solutions still face some challenges in application. When dealing with time series, the deep learning model is often limited to short-term mesoscale meteorological evolution, which makes it difficult for the model to effectively integrate long-term macro background factors such as persistent sea surface temperature anomalies and large-scale circulation into the prediction logic. At the same time, the learning paradigm purely driven by data reflects the statistical regularity of data to some extent, and lacks the following of the underlying physical mechanism, which may deviate from the physical common sense. In order to make up for this deficiency, some studies try to introduce physical constraints, but usually impose uniform constraint strength on the entire region, without considering the differences in observation data quality at different geographical locations, making it difficult to achieve the best balance between data-driven and physical priori. SUMMARY

[0004] The purpose of the present application is to provide a sea fog prediction method and device, which solves the problem of low prediction accuracy and insufficient physical authenticity caused by the lack of dynamic coupling of multi-scale physical processes and adaptive physical constraints in the existing sea fog prediction.

[0005] To achieve the above purpose, the technical scheme is as follows: A sea fog prediction method, comprising the following steps: S1, acquiring multi-source heterogeneous data for model training; S2, generating a scene representation for representing macro physical background and a confidence weight matrix with spatially variable weights in the prediction area based on the multi-source heterogeneous data through a neural network model; S3, dynamically adjusting the processing process of the multi-source heterogeneous data using the scene representation to generate a sea fog prediction result, and calculating a compound loss based on the sea fog prediction result, sea fog true label data and the confidence weight matrix; S4, updating parameters of the neural network model by a back propagation algorithm based on the compound loss.

[0006] Preferably, in the step S1, the multi-source heterogeneous data comprises sea fog true value label data, macro-scale scenario data, mesoscale spatiotemporal feature data, and data quality prior indicators.

[0007] Preferably, in the step S2, the generating the scenario representation for representing the macro-physical background comprises the following steps: S211, inputting the macro-scale scenario data in the acquired multi-source heterogeneous data into a Transformer encoder in the neural network model; S212, processing the inputted entire time sequence by using a self-attention mechanism inside the Transformer encoder, capturing long-term dependencies, periodic regularities and influences of key events existing in the sequence; S213, taking the final output state result of the sequence as the scenario representation after the Transformer encoder completes the calculation of the sequence; The generating the confidence weight matrix with spatially variable weights in the prediction area comprises the following steps: S221, inputting the data quality prior indicators in the multi-source heterogeneous data into a lightweight convolutional neural network in the neural network model; S222, extracting relevant spatial features from the inputted spatial data graph by a convolutional layer of the lightweight convolutional neural network, and learning a complex nonlinear mapping relationship between the prior indicators and the optimal physical constraint weights; S223, finally ensuring that each pixel value of the output is between 0 and 1 by a Sigmoid activation function built in the lightweight convolutional neural network, and the final output result is the confidence weight matrix.

[0008] Preferably, in the step S3, the compound loss comprises a data loss term and a physical constraint loss term associated with the confidence weight matrix.

[0009] Preferably, in the step S3, the dynamically adjusting the processing procedure of the multi-source heterogeneous data by using the scenario representation to generate the sea fog prediction result comprises the following steps: S311, inputting the mesoscale spatiotemporal feature data in the multi-source heterogeneous data and the scenario representation into a convolutional long short-term memory network in the neural network model; S312, in the convolutional long short-term memory network, the scenario representation will be inputted into one or more small fully connected networks, and the fully connected networks will output a group of dynamically changeable modulation parameters generated according to the current macro background; S313, the convolutional long short-term memory network starts to process the input mesoscale spatiotemporal feature sequence in time steps, wherein in each calculation of the time steps, when the convolutional long short-term memory network updates the input gate and the forget gate, an affine transformation is performed on the standard gate calculation result by using the modulation parameter; S314, after processing all the input sequences of the time steps, the convolutional long short-term memory network outputs the final hidden state thereof; The calculating the composite loss based on the sea fog prediction result, the sea fog true value label data, and the confidence weight matrix comprises the following calculation steps: S321, the sea fog prediction result output by the model is compared with the real sea fog label data pixel by pixel, a total difference value is calculated by using a binary cross-entropy function, and a data loss value is output; S322, according to the saturation condition and the thermal advection physical law, the degree of violation of the model prediction result to the physical law at each pixel point is calculated, a physical residual matrix is formed, and the physical residual matrix is multiplied with the confidence weight matrix representing the data credibility of each position element by element, to obtain a physical loss value; S323, the data loss value and the physical loss value are weighted and summed to obtain a final composite loss value.

[0010] Preferably, in step S313, the affine transformation is realized by the following formula: M ′ = γ ⊙ M + β; In the formula, M is the standard gate calculation result, M ′ is the transformed gate calculation result, γ and β are scale and bias modulation parameters generated according to the scene representation, and ⊙ represents Hadamard product.

[0011] Preferably, in step S322, when the physical law is thermal advection, the residual R adv at any position in the physical residual matrix is calculated by the following discrete form of the advection equation: In the formula, T is the temperature field predicted by the model, u is the wind field vector, is the partial derivative of temperature with respect to time, is the temperature gradient, |. | represents taking an absolute value, and R adv is the residual at any position in the physical residual matrix.

[0012] Preferably, in step S323, the final composite loss value is calculated by the following formula: L total = L data+ lambda * L phys ; wherein L data is the data loss value, L phys is the physical loss value, lambda is a preset weight hyperparameter for balancing the two losses, L total is the final composite loss value.

[0013] A sea fog prediction device comprises: a data acquisition module configured to acquire multi-source heterogeneous data required for model training; a scenario representation generation module connected to the data acquisition module and configured to receive macro-scale scenario data in the multi-source heterogeneous data and generate a scenario representation for representing a macro-physical background; a confidence matrix generation module connected to the data acquisition module and configured to receive data quality prior indicators in the multi-source heterogeneous data and generate a confidence weight matrix having spatially variable weights in a prediction area; a spatio-temporal prediction module connected to the data acquisition module and the scenario representation generation module and configured to receive mesoscale spatio-temporal feature data and dynamically adjust an internal processing process thereof by using the scenario representation received from the scenario representation generation module to generate a final sea fog prediction result; a composite loss calculation module connected to the data acquisition module, the confidence matrix generation module and the spatio-temporal prediction module and configured to receive sea fog true value label data, a confidence weight matrix and a sea fog prediction result and calculate a composite loss therefrom; a parameter optimization module connected to the composite loss calculation module and configured to receive the composite loss and update internal parameters of the scenario representation generation module, the confidence matrix generation module and the spatio-temporal prediction module based on the loss by using a back propagation algorithm.

[0014] In summary, the present application has at least one of the following beneficial technical effects: 1. The present application uses a dynamic adjustment mechanism of macro scenario and mesoscale process to enable the model to deeply understand how long-term climate background affects the generation and dissipation process of local sea fog, thereby improving the accuracy of sea fog prediction and the generalization ability of the model. Compared with the prior art which mostly relies on short-term meteorological data for extrapolation prediction, the present application solves the problem of lack of long-term physical background cognition, which leads to the decline of the prediction performance of the model under rare weather patterns.

[0015] 2.The method of the present application introduces a physical constraint loss term into the objective function of model training. It can effectively guide the learning direction of the model and force the prediction results to follow the basic physical laws such as the thermal advection equation. This ensures that the final generated sea fog distribution and related meteorological field have physical authenticity and consistency. Compared with the deep learning prediction model purely driven by data, the present application solves the inherent defect that the prediction results may be physically inconsistent and produce phenomena that violate the laws of nature.

[0016] 3.The method of the present application uses a confidence prediction network to generate a spatially variable confidence weight matrix based on data quality prior indicators, and adaptively adjusts the punishment strength of physical constraints in different spaces. Compared with the existing technology of applying uniform physical constraints to the entire prediction area, the present application solves the limitation that the model cannot be treated differently according to the geographical differences in the quality of observation data, leading to uneven training of the model in data sparse and dense areas. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The method flowchart of the present application is shown in the figure; Figure 2 The device module architecture diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0018] The following will be described in detail in combination with the accompanying Figure 1 - the accompanying Figure 2 , the present application will be further described in detail.

[0019] The present application provides a sea fog prediction method and device.

[0020] As Figure 1 shown, the sea fog prediction method can include the following steps: S1, obtaining multi-source heterogeneous data for model training; Specifically, step S1, obtaining multi-source heterogeneous data for model training. This step aims to provide a standardized, multi-dimensional physical information containing, and strictly aligned in space and time for subsequent model construction and training data basis, so as to support the decoupling and dynamic fusion of multi-scale physical processes in the subsequent steps.

[0021] In one specific embodiment, the multi-source heterogeneous data obtained in step S1 further includes sea fog true value label data, macro-scale scenario data, mesoscale spatiotemporal feature data, and data quality prior indicators. These four types of data together constitute a complete input set for the model to perform supervised learning and physical constraint learning.

[0022] For the acquisition of sea fog ground truth label data, exemplarily, it can be achieved by processing multi-channel remote sensing images taken by geostationary satellites, such as FY-4 satellite. First, a series of satellite cloud images at different time points in history are collected; then, a pre-trained semantic segmentation network, such as U-Net model, is used to analyze and identify these cloud images pixel by pixel to automatically label the spatial range of sea fog occurrence; finally, the labeling results are binarized to generate sea fog ground truth label data corresponding to the prediction area grid. This data is used as a supervision signal, i.e., the "standard answer", in model training.

[0023] For the acquisition of macro-scale scenario data, which is configured to represent the long-term climate and ocean background of sea fog generation. Further, this data can include sea surface temperature anomaly index (SSTA), intensity index of major ocean currents such as Kuroshio, and astronomical tide phase encoding calculated from astronomical data. These data are usually low-frequency time series with a long time span (e.g., several months to several years), which are collectively organized into a sequence data set for subsequent steps to generate scenario representation.

[0024] For the acquisition of mesoscale spatiotemporal feature data, which aims to provide direct physical driving factors for the evolution of sea fog generation and dissipation. In one embodiment, this data can be derived from the ERA5 reanalysis dataset of the European Centre for Medium-Range Weather Forecasts or the output products of the WRF numerical model. Further, this data includes high-resolution meteorological fields at N time points (e.g., 24 time points) aligned with the sea fog ground truth label in time and space. The meteorological field can include temperature field, relative humidity field, three-dimensional wind field (U / V / W component) at different pressure levels, and sea surface temperature (SST) and sea-air temperature difference field. This sequence data constitutes the core input for the model to perform spatiotemporal evolution prediction.

[0025] For the acquisition of data quality prior indicators, the purpose is to provide a basis for subsequent application of spatially variable physical constraints. Further, this indicator can include one or more static or quasi-static data in the form of spatial grid maps. For example, it can include a spatial density distribution map generated based on the location of historical observation stations, an offshore distance map calculated based on coastline data, a bathymetric map describing the shape of the seabed, and a statistical average cloud cover rate map of satellite data in each season.

[0026] Finally, this step also includes a key data preprocessing process. This process uniformly processes all the data obtained above to the same geographic coordinate system and spatial resolution grid through interpolation, resampling, etc., and ensures that all time series data are strictly aligned in time stamp. This ensures that all data input into the neural network model have uniform format and dimension, which is a prerequisite for the subsequent model to correctly process multi-source heterogeneous data.

[0027] S2, generating a scenario representation for representing a macro-physical context based on the multi-source heterogeneous data via a neural network model, and simultaneously generating a confidence weight matrix with spatially variable weights in the prediction area; Specifically, the S2 step performs two independent sub-tasks in parallel via a unified neural network model: one is to generate a scenario representation for representing a macro-physical context based on macro-scale scenario data; the other is to generate a confidence weight matrix with spatially variable weights in the prediction area based on data quality prior indicators. The two generated intermediate results will be the core basis for dynamic prediction and adaptive loss calculation in the subsequent steps.

[0028] In a specific embodiment, the neural network model internally includes a macro-scale scenario modulation module and a spatio-temporal confidence prediction auxiliary network. The two modules are simultaneously invoked in one forward propagation calculation and are respectively responsible for the execution of the above two sub-tasks.

[0029] For the sub-task of generating a scenario representation, the specific implementation is as follows. First, the macro-scale scenario data obtained in step S1, such as a time series with a length of L, is input into the macro-scale scenario modulation module. Further, the module can be implemented as a Transformer encoder.

[0030] Subsequently, the Transformer encoder processes the input entire time series using its internal self-attention mechanism. This mechanism can capture long-term dependencies, periodic patterns, and the influence of key events (such as a significant sea temperature anomaly) in the sequence data, thereby effectively encoding the macro-context information.

[0031] After completing the calculation of the sequence, the Transformer encoder outputs its final output state, such as a fixed-dimensional vector, as the scenario representation C. The scenario representation C is a highly condensed summary of the long-time-scale physical context, which can provide context information for the subsequent prediction of mesoscale processes.

[0032] In parallel, for the sub-task of generating a confidence weight matrix, the specific implementation is as follows. First, the data quality prior indicators obtained in step S1, such as historical observation station density maps or off-shore distance maps in image form, are input into the spatio-temporal confidence prediction auxiliary network. Further, the auxiliary network can be implemented as a lightweight convolutional neural network (CNN).

[0033] Subsequently, relevant spatial features are extracted from the input spatial data map through multiple convolutional layers of the CNN. The network is trained to learn the complex nonlinear mapping relationship between the prior indicators and the optimal physical constraint weights. It can identify which regions (such as nearshore, observation station dense areas) have more reliable data and which regions (such as open sea) have lower data confidence.

[0034] Finally, the output layer of the CNN is processed by a Sigmoid activation function, which is calculated by the following formula: where x represents the original output value of the output layer of the lightweight convolutional neural network at any spatial position; e is the base of the natural logarithm, which is a mathematical constant; and σ(x) represents the output probability value at the spatial position after Sigmoid function calculation, which is normalized to the range (0, 1). This function can ensure that each pixel value of the final output is within the range of 0 to 1. The final output of the network is a two-dimensional matrix consistent with the grid size of the predicted area, i.e., the confidence weight matrix W. This matrix provides spatially variable weights for the calculation of physical loss in subsequent steps.

[0035] S3, dynamically adjusting the processing process of the multi-source heterogeneous data to generate the sea fog prediction result, and calculating a composite loss based on the sea fog prediction result, the sea fog true value label data, and the confidence weight matrix; Specifically, the S3 step is configured to perform a complete forward propagation calculation and loss evaluation, which includes two consecutive stages: first, using the scenario representation generated in S2, dynamically adjusting the processing process of the mesoscale spatiotemporal feature data to generate a sea fog prediction result; second, based on the prediction result, the sea fog true value label data, and the confidence weight matrix, a composite loss is calculated to guide the model optimization.

[0036] In one specific embodiment, the neural network model of the method further includes a mesoscale physical process evolution module. For the stage of generating a sea fog prediction result, the mesoscale spatiotemporal feature data obtained in step S1 and the scenario representation C generated in step S2 are jointly input into the evolution module. In one specific implementation, the evolution module can be implemented as a convolutional long short-term memory network (ConvLSTM).

[0037] Within the ConvLSTM module, the scenario representation C is input into one or more small fully connected networks. These networks are configured to output a set of dynamically variable modulation parameters, such as a scale vector γ and a bias vector β. The values of these parameters are uniquely determined by the current macroscopic background and can achieve dynamic modulation of subsequent spatiotemporal evolution calculations.

[0038] Subsequently, the ConvLSTM module iteratively processes the input mesoscale spatiotemporal feature sequence in time steps. In the calculation of each time step, when the network updates its internal gating units, such as the input gate or the forget gate, an affine transformation is performed on the standard gating calculation result using the modulation parameters (γ, β) generated in the previous step.

[0039] Further, the affine transformation is implemented by the following formula: M ′ = γ 0 M + β; In the formula, M is the standard gating calculation result, M ′ is the transformed gating calculation result, γ and β are scale and bias modulation parameters generated according to the scene representation, and 0 0 represents Hadamard product.

[0040] After processing all the input sequences in the time steps, the ConvLSTM module outputs its final hidden state. The hidden state is then passed through an output layer (for example, a 1x1 convolutional layer and a Sigmoid activation function) to finally generate the sea fog prediction result.

[0041] After generating the prediction result, the calculation of the composite loss is then performed. In one embodiment, the composite loss includes a data loss term and a physical constraint loss term associated with a confidence weight matrix.

[0042] First, the data loss term is calculated. The sea fog prediction result output by the model is compared with the sea fog true value label data obtained in step S1 pixel by pixel, and a total difference value is calculated by a binary cross-entropy function. The value is the data loss term L data .

[0043] Subsequently, the physical constraint loss term is calculated. The calculation first needs to calculate the degree of violation of the model prediction result to one or more physical laws at each pixel point according to the laws, forming a physical residual matrix. For example, when the physical law is thermal advection, the physical residual R adv at any position can be calculated by the discretization form of the advection equation as follows: In the formula, T is the temperature field predicted by the model, u is the wind field vector, is the partial derivative of temperature with respect to time, is the temperature gradient, |. | represents the absolute value, and R adv is the residual at any position in the physical residual matrix.

[0044] After obtaining the physical residual matrix, it is element-wise multiplied with the confidence weight matrix W generated in step S2, which represents the confidence of each position data. This operation can make stronger physical constraints in areas with low data confidence, while relying more on observation data in areas with high data confidence. All elements of the multiplied result matrix are summed or averaged to obtain the final physical loss value L total .

[0045] Finally, the data loss value L data and the physical loss value L total are weighted and summed to obtain the final composite loss value L total . Further, the composite loss value can be calculated by the following formula: L total = L data + λ·L phys ; In the formula, L data is the data loss value, L phys is the physical loss value, λ is a preset weight hyperparameter for balancing the two losses, and L total is the final composite loss value.

[0046] The composite loss value L_total is the final output of step S3, which will be passed to step S4 for parameter update of the model.

[0047] S4, update the parameters of the neural network model based on the composite loss through the backpropagation algorithm; Specifically, the S4 step aims to update all trainable parameters in the neural network model through the backpropagation algorithm based on the composite loss value calculated in step S3. This step is the fundamental link for the model to learn the rules from the data and gradually optimize its prediction ability, and it constitutes the closed loop of the entire construction method.

[0048] In a specific embodiment, this step first starts the backpropagation algorithm. This algorithm starts from the scalar composite loss value L_total output by step S3, and applies the chain rule of calculus from back to front along the computation graph established by the neural network during forward propagation (steps S2 and S3).

[0049] This process can efficiently calculate the partial derivative of the composite loss value L_total with respect to each trainable parameter (e.g., weights within the convolution kernel or biases of the fully connected layer) in the neural network model. This partial derivative is the gradient, which quantifies the contribution of each parameter to the final prediction total error and indicates the adjustment direction that can make the loss decrease the fastest.

[0050] After the gradients of all parameters are calculated, a preset optimizer algorithm, exemplarily, Adam optimizer, will perform the actual parameter update. The optimizer updates any trainable parameter θ in the model according to the following formula: where θ new denotes the new value of the parameter after the update; θ old denotes the current value of the parameter before the update; η denotes the learning rate, which is a preset hyperparameter for controlling the step size of each parameter update; denotes the composite loss value L total calculated by backpropagation in the last step.

[0051] The above gradient calculation and parameter update together constitute a complete training iteration. In specific implementation, the flow of the entire steps S1 to S4 will be repeatedly executed on multiple data batches of the training dataset. Through iterations of several training cycles, the model parameters are continuously fine-tuned, so that the difference between the prediction results output by the model and the true observation and the degree of violation of the physical law are jointly and systematically reduced. When the performance of the model converges on an independent validation set, the training process terminates, thereby obtaining a constructed sea fog prediction model.

[0052] A sea fog prediction device described below can be correspondingly referred to the sea fog prediction method described above.

[0053] Please refer to the accompanying drawings Figure 2 The present application also provides a sea fog prediction device, comprising: a data acquisition module for acquiring multi-source heterogeneous data required for model training; a scenario representation generation module connected with the data acquisition module, for receiving macro-scale scenario data in the multi-source heterogeneous data and generating a scenario representation for representing the macro physical background; a confidence matrix generation module connected with the data acquisition module, for receiving data quality prior indicators in the multi-source heterogeneous data and generating a confidence weight matrix with spatially variable weights in the prediction area; a spatio-temporal prediction module connected with the data acquisition module and the scenario representation generation module, for receiving mesoscale spatio-temporal feature data and dynamically adjusting its internal processing process using the scenario representation received from the scenario representation generation module to generate the final sea fog prediction result; a composite loss calculation module connected with the data acquisition module, the confidence matrix generation module and the spatio-temporal prediction module, for receiving sea fog true label data, confidence weight matrix and sea fog prediction result, and calculating the composite loss accordingly; A parameter optimization module connected with the compound loss calculation module, configured to receive the compound loss, and update internal parameters of the scene representation generation module, the confidence matrix generation module and the spatio-temporal prediction module based on the loss through a back propagation algorithm.

[0054] The device of the embodiment can be used to execute the method embodiments described above, and has similar principles and technical effects, which will not be described here again.

[0055] Although the embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A sea fog prediction method characterized by, The method comprises the following steps: S1, acquiring multi-source heterogeneous data for model training; S2, generating a scene representation for representing a macro physical background and a confidence weight matrix with spatial variable weights in a prediction area based on the multi-source heterogeneous data through a neural network model; S3, dynamically adjusting a processing procedure of the multi-source heterogeneous data by using the scene representation to generate a sea fog prediction result, and calculating a compound loss based on the sea fog prediction result, a sea fog true value label data and the confidence weight matrix; S4, updating parameters of the neural network model through a back propagation algorithm based on the compound loss.

2. The sea fog prediction method according to claim 1, wherein In the S1 step, the multi-source heterogeneous data comprises a sea fog true value label data, a macro scale scene data, a mesoscale spatiotemporal feature data and a data quality priori index.

3. The sea fog prediction method according to claim 1, wherein In the S2 step, the generation of the scene representation for representing the macro physical background comprises the following steps: S211, inputting the macro scale scene data in the acquired multi-source heterogeneous data into a Transformer encoder in the neural network model; S212, processing the inputted entire time sequence by using a self-attention mechanism inside the Transformer encoder to capture long-term dependence, periodicity and influence of key events existing in the sequence; S213, taking a final output state result of the Transformer encoder after completing the calculation of the sequence as the scene representation; The generation of the confidence weight matrix with spatial variable weights in the prediction area comprises the following steps: S221, inputting the data quality priori index in the multi-source heterogeneous data into a lightweight convolutional neural network in the neural network model; S222, extracting relevant spatial features from the inputted spatial data graph through a convolutional layer of the lightweight convolutional neural network, and learning a complex nonlinear mapping relationship between the priori index and an optimal physical constraint weight; S223, finally ensuring that each pixel value of the output is between 0 and 1 through a Sigmoid activation function inside the lightweight convolutional neural network, and taking a final output result as the confidence weight matrix.

4. The sea fog prediction method according to claim 1, wherein In the S3 step, the compound loss comprises a data loss term and a physical constraint loss term associated with the confidence weight matrix.

5. The sea fog prediction method according to claim 1, wherein In the S3 step, the dynamic adjustment of the processing procedure of the multi-source heterogeneous data by using the scene representation to generate the sea fog prediction result comprises the following steps: S311, inputting the mesoscale spatiotemporal feature data in the multi-source heterogeneous data and the scene representation into a convolutional long short-term memory network in the neural network model; S312, in the convolutional long short-term memory network, the scene representation will be inputted into one or more small-sized fully connected networks, and the fully connected networks will output a group of dynamically changeable modulation parameters generated according to a current macro background; S313, the convolutional long short-term memory network starts to process the input mesoscale spatiotemporal feature sequence in time steps, wherein in each time step calculation, when the convolutional long short-term memory network updates the input gate and the forget gate, an affine transformation is performed on the standard gate calculation result using the modulation parameter; S314, after processing all the input sequences of the time steps, the convolutional long short-term memory network outputs the final hidden state thereof; The calculation of the composite loss based on the sea fog prediction result, the sea fog true value label data and the confidence weight matrix comprises the following calculation steps: S321, the sea fog prediction result output by the model is compared with the real sea fog label data pixel by pixel, and a total difference value is calculated by a binary cross-entropy function, and the output is a data loss value; S322, according to the saturation condition and the thermal advection physical law, the degree of violation of the model prediction result to these physical laws at each pixel point is calculated to form a physical residual matrix, and then the physical residual matrix is multiplied element by element with the confidence weight matrix representing the data reliability at each position to obtain a physical loss value; S323, the data loss value and the physical loss value are weighted and summed to obtain a final composite loss value.

6. The sea fog prediction method according to claim 5, wherein, In step S313, the affine transformation is realized by the following formula: M ′ = γ o M + β; where M is the standard gating computation result, M ′ is the transformed gating computation result, γ and β are scale and bias modulation parameters generated according to the context representation, and ⊙⊙ denotes Hadamard product.

7. The sea fog prediction method according to claim 5, wherein In step S322, when the physical law is thermal advection, the residual R at any position in the physical residual matrix is calculated as follows: adv The advection equation is calculated by the following discretized form: where T is the model predicted temperature field, u is the wind field vector, is the partial derivative of temperature with respect to time, is the temperature gradient, |. | denotes taking the absolute value, R adv is the residual at any position in the physical residual matrix.

8. The sea fog prediction method according to claim 5, wherein, In step S323, the final composite loss value is calculated by the following formula: L total = L data + λ · L phys ; In the formula, L data is the data loss value, L phys is the physical loss value, λ is a preset weight hyperparameter for balancing the two losses, L total is the final composite loss value.

9. A sea fog prediction device according to the sea fog prediction method of any one of claims 1 to 8, characterized by, It comprises: A data acquisition module configured to acquire multi-source heterogeneous data required for model training; A scenario representation generation module connected to the data acquisition module and configured to receive macro-scale scenario data in the multi-source heterogeneous data and generate a scenario representation for representing a macro physical background; A confidence matrix generation module connected to the data acquisition module and configured to receive data quality prior indicators in the multi-source heterogeneous data and generate a confidence weight matrix with spatially variable weights in a prediction area; A spatiotemporal prediction module connected to the data acquisition module and the scenario representation generation module and configured to receive mesoscale spatiotemporal feature data and dynamically adjust its internal processing process using the scenario representation received from the scenario representation generation module to generate a final sea fog prediction result; A composite loss calculation module connected to the data acquisition module, the confidence matrix generation module and the spatiotemporal prediction module and configured to receive sea fog true value label data, a confidence weight matrix and a sea fog prediction result and calculate a composite loss therefrom; A parameter optimization module connected to the composite loss calculation module and configured to receive the composite loss and update internal parameters of the scenario representation generation module, the confidence matrix generation module and the spatiotemporal prediction module based on the loss by a back propagation algorithm.

Images