Marine environment short-term prediction method and system based on conditional adversarial network
By constructing a spatio-temporal graph neural network model enhanced by time-sequence variational autoencoder and attention mechanism, combined with adaptive fusion strategy, the shortcomings of traditional marine environment prediction methods in initial condition sensitivity and nonlinear feature capture are solved, and high efficiency is achieved , accurate short-term predictions.
Patent Information
- Application Number
- CN202411259930.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-09-10
AI Technical Summary
Traditional marine environment prediction methods, including numerical model method and statistical model method, respectively, are sensitive to initial conditions and boundary conditions and complex calculations, making it difficult to achieve real-time prediction, and is difficult to capture nonlinear and non-stationary characteristics of the marine environment, resulting in prediction accuracy. limited.
Using a method based on conditional adversarial network, a spatiotemporal graph neural network model with enhanced attention mechanism is constructed. Through long-term memory network and graph convolution operations, the weights are dynamically adjusted to capture the spatial and temporal evolution law of the marine environment. and spatial correlation mode, and weighted aggregation model output through adaptive fusion strategy, generate accurate short-term prediction results.
Effectively modeling of high-dimensional nonlinear features of marine environments improves the accuracy and robustness of predictions, makes the prediction results physically reasonable and interpretable, and enhances prediction performance and robustness.
Smart Images

Figure CN118779582B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to surveying and mapping technology, and in particular to a method and system for short-term prediction of marine environment based on a conditional adversarial network. Background Art
[0002] Traditional marine environment prediction methods mainly include numerical model method and statistical model method.
[0003] The numerical model method simulates the physical processes of the ocean environment by solving fluid dynamics equations and predicts future ocean conditions.
[0004] However, the numerical model method is sensitive to initial and boundary conditions and has high computational complexity, making it difficult to achieve real-time predictions.
[0005] The statistical model method analyzes historical observation data, establishes statistical relationships, and predicts future changes in the marine environment.
[0006] However, the statistical model method has difficulty in capturing the nonlinear and non-stationary characteristics of the marine environment, and its prediction accuracy is limited. Summary of the invention
[0007] The embodiments of the present invention provide a method and system for short-term prediction of marine environment based on conditional adversarial network, which can solve the problems in the prior art.
[0008] According to a first aspect of the embodiments of the present invention,
[0009] Provides a short-term prediction method for the marine environment based on conditional adversarial networks, including:
[0010] Constructing a temporal variational autoencoder model, the temporal variational autoencoder model comprising an encoder and a decoder, the encoder using a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder using a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state;
[0011] Construct an attention-enhanced spatiotemporal graph neural network model, model the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregate the local features of nodes through graph convolution operations, and use a multi-head attention mechanism to dynamically adjust the weights between different nodes to model the spatial correlation pattern of the marine environment;
[0012] The temporal variational autoencoder model generates a potential representation of multiple future time steps based on the ocean environment state at the current moment and future condition settings, and the attention mechanism enhanced spatiotemporal graph neural network model predicts the evolution of the ocean environment map for multiple future time steps based on the ocean environment map at the current moment and future condition settings; an adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weightedly aggregated to obtain the final prediction results of the spatiotemporal distribution of ocean environment physical quantities for multiple future time steps.
[0013] In an optional embodiment,
[0014] Constructing a temporal variational autoencoder model, the temporal variational autoencoder model includes an encoder and a decoder, the encoder uses a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder uses a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state, including:
[0015] Constructing a temporal variational autoencoder model of an encoder-decoder architecture, wherein the encoder uses a multi-layer long short-term memory network to encode the spatiotemporal sequence data of the marine environment in the target area. At each time step, the long short-term memory network unit receives the marine environment state vector at the current moment, and updates the hidden state and memory unit at the current moment in combination with the hidden state and memory unit at the previous moment, thereby capturing the contextual information and long-range dependencies in the time series data;
[0016] The parameters of the encoder and the decoder are optimized by maximizing the variational lower bound, the output of the encoder is passed through two fully connected layers to estimate the mean and logarithmic variance of the latent variables respectively, the latent variables are sampled from the estimated Gaussian distribution and input into the decoder for reconstruction;
[0017] The decoder adopts a long short-term memory network structure that is symmetrical with the encoder, recursively generates a sequence of ocean environment states based on latent variables and decoding results at the previous moment, and uses conditional Gaussian distribution as a generation model in the output layer of the decoder, and optimizes the parameters of the encoder and the decoder by minimizing reconstruction error and KL divergence loss.
[0018] In an optional embodiment,
[0019] Optimizing the parameters of the encoder and the decoder by minimizing the reconstruction error and the KL divergence loss includes:
[0020] The reconstruction error is expressed as follows: ;
[0021] The KL divergence loss is shown in the following formula: ;
[0022] in, L R represents the reconstruction error, L K represents the KL divergence loss, N , T represent the sample size and time period respectively, Indicates i samples at time step t The observation vector of Indicates i The latent variable corresponding to the sample is Indicates i samples at time step t The previous observation sequence, μ θ () represents the mean of the decoder estimate, Σ θ () represents the covariance matrix estimated by the decoder; z represents the latent variable, represents the posterior distribution estimated by the encoder, which represents the conditional probability distribution of the latent variable given a sequence of observations, p(z) represents the prior distribution of the latent variable.
[0023] In an optional embodiment,
[0024] Construct an attention-enhanced spatiotemporal graph neural network model, model the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregate the local features of nodes through graph convolution operations, and use a multi-head attention mechanism to dynamically adjust the weights between different nodes to model the spatial correlation patterns of the marine environment, including:
[0025] The distribution of physical quantities of the marine environment is represented as a space-time graph, wherein the space-time graph is composed of M nodes, wherein the M nodes represent observation points at different locations, wherein the edges between the nodes of the space-time graph represent the spatial associations between the nodes, and the node attribute tensor of the space-time graph represents the changes of the node attributes over time;
[0026] Performing a graph convolution operation on the spatiotemporal graph to update the features of each node by aggregating the features of neighboring nodes of each node in the spatiotemporal graph;
[0027] On the basis of the graph convolution operation, a multi-head attention mechanism is introduced, and the attention weight of each attention head is obtained by calculating the evaluation value of the association strength between nodes by different attention heads. The features of neighboring nodes in the graph convolution operation are weightedly aggregated using the attention weight to model the dynamic association pattern between nodes in the spatiotemporal graph.
[0028] On top of the graph convolution operation enhanced by the attention mechanism, a temporal convolutional network is introduced to capture the temporal evolution of the attributes of the nodes in the spatiotemporal graph through the combination of causal convolution and dilated convolution. The temporal convolutional network is combined with the graph convolution operation enhanced by the attention mechanism to construct a spatiotemporal graph neural network model.
[0029] The mean square error loss function is used to measure the difference between the predicted value and the true value of the spatiotemporal graph neural network model. The spatiotemporal graph neural network model is trained by minimizing the mean square error loss function and introducing the fluid dynamics equation as a regularization term to obtain the trained spatiotemporal graph neural network model.
[0030] In an optional embodiment,
[0031] Introducing fluid dynamics equations as regularization terms to train the spatiotemporal graph neural network model includes:
[0032] The fluid dynamics equation is shown in the following formula:
[0033] ;
[0034] in, u represents the fluid velocity vector, t Indicates time, ρ represents the fluid density, p Indicates pressure, ν represents the kinematic viscosity of the fluid, f Represents the external force term.
[0035] In an optional embodiment,
[0036] Adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weighted and aggregated to obtain the final prediction results of the spatiotemporal distribution of marine environmental physical quantities for multiple time steps in the future, including:
[0037] Probabilistic modeling is performed on the output results of the temporal variational autoencoder model and the output results of the spatiotemporal graph neural network model enhanced by the attention mechanism, and they are regarded as a probability distribution. Through the Bayesian inference framework, based on the historical observation data and the model output results, the posterior distribution of each model is calculated using the Bayesian theorem to obtain the posterior estimates and uncertainty measures of the model parameters, and the posterior distribution is approximately calculated through Monte Carlo sampling or variational inference methods to obtain the probability distribution of the model output results.
[0038] According to the posterior probability of each model, its weight in the fusion process is calculated, and the spatiotemporal attention mechanism is introduced. According to the feature representation of different time steps and spatial positions, the fusion weight of each model at different spatiotemporal positions is dynamically adjusted. An adaptive learning rate strategy is adopted to dynamically adjust the update step size of the fusion weight according to the error between the model fusion result and the observed data.
[0039] Using the Bayesian model averaging method, the prediction distributions of different models are weighted averaged according to the posterior probability of each model and the adaptive fusion weight to obtain the fused prediction distribution, whose mean is used as the final prediction result and the variance is used as a measure of prediction uncertainty. Based on the fused prediction distribution, the mean of the prediction distribution is temporally interpolated and extrapolated to obtain continuous spatiotemporal distribution prediction results. The variance of the prediction distribution is used to quantify the uncertainty of the prediction results at different spatiotemporal positions and generate confidence intervals or probability maps.
[0040] According to a second aspect of the embodiments of the present invention,
[0041] Provides a short-term marine environment prediction system based on conditional adversarial networks, including:
[0042] The first unit is used to construct a time series variational autoencoder model, which includes an encoder and a decoder. The encoder uses a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder uses a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state;
[0043] The second unit is used to build a spatiotemporal graph neural network model enhanced by an attention mechanism. The distribution of physical quantities in the marine environment is modeled as a spatiotemporal graph. The local features of the nodes are aggregated through graph convolution operations, and the weights between different nodes are dynamically adjusted using a multi-head attention mechanism to model the spatial correlation pattern of the marine environment.
[0044] The third unit is used for the time series variational autoencoder model to generate potential representations for multiple future time steps based on the ocean environment status at the current moment and future condition settings, and the attention mechanism enhanced spatiotemporal graph neural network model to predict the evolution of the ocean environment map for multiple future time steps based on the ocean environment map at the current moment and future condition settings; an adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weightedly aggregated to obtain the final prediction results of the spatiotemporal distribution of ocean environment physical quantities for multiple future time steps.
[0045] According to a third aspect of the embodiments of the present invention,
[0046] An electronic device is provided, comprising:
[0047] processor;
[0048] a memory for storing processor-executable instructions;
[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0050] According to a fourth aspect of the embodiments of the present invention,
[0051] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0052] By introducing a time series variational autoencoder model and an attention mechanism-enhanced spatiotemporal graph neural network model, the present invention can effectively model the high-dimensional nonlinear characteristics of marine environmental data, capture the spatiotemporal evolution laws of the marine environment, and thus generate accurate and reliable short-term prediction results.
[0053] The attention-enhanced spatiotemporal graph neural network model models the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregates the local features of nodes through graph convolution operations, and uses a multi-head attention mechanism to dynamically adjust the weights between different nodes. It can model the spatial correlation patterns between physical quantities in the marine environment and reveal the interactions and causal relationships between physical quantities.
[0054] The temporal variational autoencoder model compresses the representation of the ocean environment state and reconstructs the original ocean environment state sequence from the compressed representation, ensuring the physical rationality of the prediction results.
[0055] At the same time, the attention mechanism-enhanced spatiotemporal graph neural network model makes the prediction results interpretable by modeling the spatial correlation patterns of the marine environment, which facilitates analysis and decision-making by marine experts.
[0056] The advantages of the two models are combined to improve the prediction performance and robustness.
[0057] By adopting an adaptive fusion strategy to dynamically adjust the fusion weights of the temporal variational autoencoder model and the attention mechanism enhanced spatiotemporal graph neural network model, and weightedly aggregate the output results of the two models, the complementary advantages of the two models can be brought into play and the prediction performance and robustness can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 It is a flow chart of a method for short-term prediction of marine environment based on conditional adversarial network according to an embodiment of the present invention;
[0059] Figure 2 It is a schematic diagram of the structure of a marine environment short-term prediction system based on a conditional adversarial network according to an embodiment of the present invention. DETAILED DESCRIPTION
[0060] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions 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, rather than all the embodiments.
[0061] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.
[0062] The technical solution of the present invention is described in detail below with specific embodiments.
[0063] The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.
[0064] Figure 1 FIG. 1 is a flow chart of a method for short-term prediction of marine environment based on a conditional adversarial network according to an embodiment of the present invention. Figure 1 As shown, the method includes:
[0065] S101. constructing a temporal variational autoencoder model, the temporal variational autoencoder model comprising an encoder and a decoder, the encoder using a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder using a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state;
[0066] S102. Construct an attention-enhanced spatiotemporal graph neural network model, model the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregate the local features of nodes through graph convolution operations, and use a multi-head attention mechanism to dynamically adjust the weights between different nodes to model the spatial correlation pattern of the marine environment;
[0067] S103. The temporal variational autoencoder model generates a potential representation of multiple future time steps based on the ocean environment state at the current moment and future condition settings, and the attention mechanism enhanced spatiotemporal graph neural network model predicts the evolution of the ocean environment map for multiple future time steps based on the ocean environment map at the current moment and future condition settings; an adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weightedly aggregated to obtain the final prediction results of the spatiotemporal distribution of ocean environment physical quantities for multiple future time steps.
[0068] The technical solution of the present application can be used in the field of surveying and mapping technology. It can not only obtain the spatiotemporal series data of the marine environment of the target area through surveying and mapping technology, but also use the spatiotemporal distribution prediction results of the physical quantities of the marine environment in multiple time steps in the future to guide surveying and mapping tasks, such as surveying time and surveying area.
[0069] Furthermore, in order to more intuitively present the prediction results of the spatiotemporal distribution of marine environmental physical quantities at multiple time steps in the future, the prediction results of the spatiotemporal distribution of marine environmental physical quantities at multiple time steps in the future can be displayed through visualization technology.
[0070] Moreover, in order to obtain more realistic or practical results, the predicted spatiotemporal distribution results of marine environmental physical quantities for multiple future time steps can be simulated first, and the corresponding data or parameters can be adaptively adjusted according to the simulation results.
[0071] By introducing a time series variational autoencoder model and an attention mechanism-enhanced spatiotemporal graph neural network model, the present invention can effectively model the high-dimensional nonlinear characteristics of marine environmental data, capture the spatiotemporal evolution laws of the marine environment, and thus generate accurate and reliable short-term prediction results.
[0072] The attention-enhanced spatiotemporal graph neural network model models the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregates the local features of nodes through graph convolution operations, and uses a multi-head attention mechanism to dynamically adjust the weights between different nodes. It can model the spatial correlation patterns between physical quantities in the marine environment and reveal the interactions and causal relationships between physical quantities.
[0073] The temporal variational autoencoder model compresses the representation of the ocean environment state and reconstructs the original ocean environment state sequence from the compressed representation, ensuring the physical rationality of the prediction results.
[0074] At the same time, the attention mechanism-enhanced spatiotemporal graph neural network model makes the prediction results interpretable by modeling the spatial correlation patterns of the marine environment, which facilitates analysis and decision-making by marine experts.
[0075] The advantages of the two models are combined to improve the prediction performance and robustness.
[0076] By adopting an adaptive fusion strategy to dynamically adjust the fusion weights of the temporal variational autoencoder model and the attention mechanism enhanced spatiotemporal graph neural network model, and weightedly aggregate the output results of the two models, the complementary advantages of the two models can be brought into play and the prediction performance and robustness can be improved.
[0077] In an optional embodiment,
[0078] Constructing a temporal variational autoencoder model, the temporal variational autoencoder model includes an encoder and a decoder, the encoder uses a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder uses a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state, including:
[0079] Constructing a temporal variational autoencoder model of an encoder-decoder architecture, wherein the encoder uses a multi-layer long short-term memory network to encode the spatiotemporal sequence data of the marine environment in the target area. At each time step, the long short-term memory network unit receives the marine environment state vector at the current moment, and updates the hidden state and memory unit at the current moment in combination with the hidden state and memory unit at the previous moment, thereby capturing the contextual information and long-range dependencies in the time series data;
[0080] The parameters of the encoder and the decoder are optimized by maximizing the variational lower bound, the output of the encoder is passed through two fully connected layers to estimate the mean and logarithmic variance of the latent variables respectively, the latent variables are sampled from the estimated Gaussian distribution and input into the decoder for reconstruction;
[0081] The decoder adopts a long short-term memory network structure that is symmetrical with the encoder, recursively generates a sequence of ocean environment states based on latent variables and decoding results at the previous moment, and uses conditional Gaussian distribution as a generation model in the output layer of the decoder, and optimizes the parameters of the encoder and the decoder by minimizing reconstruction error and KL divergence loss.
[0082] In an optional embodiment,
[0083] Optimizing the parameters of the encoder and the decoder by minimizing the reconstruction error and the KL divergence loss includes:
[0084] The reconstruction error is shown in the following formula: In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments.
[0085] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.
[0086] The technical solution of the present invention is described in detail below with specific embodiments.
[0087] The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.
[0088] Figure 1 FIG. 1 is a flow chart of a method for short-term prediction of marine environment based on a conditional adversarial network according to an embodiment of the present invention. Figure 1 As shown, the method includes:
[0089] S101. constructing a temporal variational autoencoder model, the temporal variational autoencoder model comprising an encoder and a decoder, the encoder using a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder using a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state;
[0090] S102. Construct an attention-enhanced spatiotemporal graph neural network model, model the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregate the local features of nodes through graph convolution operations, and use a multi-head attention mechanism to dynamically adjust the weights between different nodes to model the spatial correlation pattern of the marine environment;
[0091] S103. The temporal variational autoencoder model generates a potential representation of multiple future time steps based on the ocean environment state at the current moment and future condition settings, and the attention mechanism enhanced spatiotemporal graph neural network model predicts the evolution of the ocean environment map for multiple future time steps based on the ocean environment map at the current moment and future condition settings; an adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weightedly aggregated to obtain the final prediction results of the spatiotemporal distribution of ocean environment physical quantities for multiple future time steps.
[0092] The technical solution of the present application can be used in the field of surveying and mapping technology. It can not only obtain the spatiotemporal series data of the marine environment of the target area through surveying and mapping technology, but also use the spatiotemporal distribution prediction results of the physical quantities of the marine environment in multiple time steps in the future to guide surveying and mapping tasks, such as surveying time and surveying area.
[0093] Furthermore, in order to more intuitively present the prediction results of the spatiotemporal distribution of marine environmental physical quantities at multiple time steps in the future, the prediction results of the spatiotemporal distribution of marine environmental physical quantities at multiple time steps in the future can be displayed through visualization technology.
[0094] Moreover, in order to obtain more realistic or practical results, the predicted spatiotemporal distribution results of marine environmental physical quantities for multiple future time steps can be simulated first, and the corresponding data or parameters can be adaptively adjusted according to the simulation results.
[0095] By introducing a time series variational autoencoder model and an attention mechanism-enhanced spatiotemporal graph neural network model, the present invention can effectively model the high-dimensional nonlinear characteristics of marine environmental data, capture the spatiotemporal evolution laws of the marine environment, and thus generate accurate and reliable short-term prediction results.
[0096] The attention-enhanced spatiotemporal graph neural network model models the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregates the local features of nodes through graph convolution operations, and uses a multi-head attention mechanism to dynamically adjust the weights between different nodes. It can model the spatial correlation patterns between physical quantities in the marine environment and reveal the interactions and causal relationships between physical quantities.
[0097] The temporal variational autoencoder model compresses the ocean environment state and reconstructs the original ocean environment state sequence from the compressed representation, ensuring the physical rationality of the prediction results. At the same time, the attention mechanism-enhanced spatiotemporal graph neural network model models the spatial correlation pattern of the ocean environment, making the prediction results interpretable and facilitating analysis and decision-making by ocean experts.
[0098] The advantages of the two models are integrated to improve the prediction performance and robustness. By using an adaptive fusion strategy to dynamically adjust the fusion weights of the temporal variational autoencoder model and the attention mechanism enhanced spatiotemporal graph neural network model, and weightedly aggregate the output results of the two models, the complementary advantages of the two models can be brought into play to improve the prediction performance and robustness.
[0099] In an optional embodiment,
[0100] Constructing a temporal variational autoencoder model, the temporal variational autoencoder model includes an encoder and a decoder, the encoder uses a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder uses a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state, including:
[0101] Constructing a temporal variational autoencoder model of an encoder-decoder architecture, wherein the encoder uses a multi-layer long short-term memory network to encode the spatiotemporal sequence data of the marine environment in the target area. At each time step, the long short-term memory network unit receives the marine environment state vector at the current moment, and updates the hidden state and memory unit at the current moment in combination with the hidden state and memory unit at the previous moment, thereby capturing the contextual information and long-range dependencies in the time series data;
[0102] The parameters of the encoder and the decoder are optimized by maximizing the variational lower bound, the output of the encoder is passed through two fully connected layers to estimate the mean and logarithmic variance of the latent variables respectively, the latent variables are sampled from the estimated Gaussian distribution and input into the decoder for reconstruction;
[0103] The decoder adopts a long short-term memory network structure that is symmetrical with the encoder, recursively generates a sequence of ocean environment states based on latent variables and decoding results at the previous moment, and uses conditional Gaussian distribution as a generation model in the output layer of the decoder, and optimizes the parameters of the encoder and the decoder by minimizing reconstruction error and KL divergence loss.
[0104] In an optional embodiment,
[0105] Optimizing the parameters of the encoder and the decoder by minimizing the reconstruction error and the KL divergence loss includes:
[0106] The reconstruction error is expressed as follows: ;
[0107] The KL divergence loss is shown in the following formula:
[0108] ;
[0109] in, L R represents the reconstruction error, L K represents the KL divergence loss, N , T represent the sample size and time period respectively, Indicates i samples at time step t The observation vector of Indicates i The latent variable corresponding to the sample is Indicates i samples at time step t The previous observation sequence, μ θ () represents the mean of the decoder estimate, Σ θ () represents the covariance matrix estimated by the decoder; z represents the latent variable, represents the posterior distribution estimated by the encoder, which represents the conditional probability distribution of the latent variable given a sequence of observations, p(z) represents the prior distribution of the latent variable.
[0110] Exemplarily, this application details the key technologies and implementation details for constructing a temporal variational autoencoder (VAE) model in a marine environment rolling prediction scheme.
[0111] The time series VAE model aims to learn the low-dimensional latent representation of the spatiotemporal sequence data of the marine environment, capture the intrinsic laws and dynamic patterns of the evolution of the marine environment, and provide a more compact and informative feature representation for subsequent prediction tasks.
[0112] In the design of the time-series VAE model, an encoder-decoder architecture is adopted, in which the encoder is responsible for mapping the high-dimensional ocean environment spatiotemporal sequence data into a low-dimensional latent space, and the decoder reconstructs the original ocean environment state sequence based on the latent representation.
[0113] Different from the traditional VAE model, the long short-term memory network (LSTM) is introduced in the encoder and decoder to better model the long-range dependencies and dynamic evolution patterns in time series data.
[0114] Specifically, the encoder receives the spatiotemporal sequence data of the ocean environment in the target area and encodes it through a multi-layer LSTM network. At each time step, the LSTM unit receives the ocean environment state vector at the current moment, and combines the hidden state and memory unit at the previous moment to update the hidden state and memory unit at the current moment.
[0115] Through this recurrent encoding approach, the encoder is able to capture contextual information and long-range dependencies in time series data.
[0116] In the last layer of the encoder, an attention mechanism is introduced to allow the model to adaptively focus on time steps that have a greater impact on the potential representation according to the importance of different moments.
[0117] Through the attention mechanism, the encoder can more effectively extract key patterns and dynamic features in time series data.
[0118] In the latent space, the ocean environment state is assumed to obey a Gaussian distribution, and the parameters of the encoder and decoder are optimized by maximizing the variational lower bound (ELBO).
[0119] Specifically, the output of the encoder passes through two fully connected layers to estimate the mean and log-variance of the latent variables, respectively.
[0120] Subsequently, latent variables are sampled from the estimated Gaussian distribution and fed into the decoder for reconstruction.
[0121] The decoder adopts an LSTM network structure that is symmetrical with the encoder, and recursively generates a sequence of ocean environment states based on the latent variables and the decoding results of the previous moment.
[0122] In the output layer of the decoder, conditional Gaussian distribution is used as the generation model, that is, it is assumed that given the latent variables and the decoding results of the previous moment, the state of the ocean environment at the current moment obeys the Gaussian distribution.
[0123] By minimizing the reconstruction error and KL divergence loss, the parameters of the encoder and decoder are optimized so that the decoder can generate a sequence of ocean environment states from the latent representation that is as similar as possible to the original data.
[0124] In order to further improve the representation ability and generalization performance of the time series VAE model, a number of innovative technologies were introduced in model training.
[0125] First, the KL divergence loss annealing strategy was adopted, that is, a smaller weight was assigned to the KL divergence loss term in the early stage of training, allowing the model to focus on the reconstruction loss first and learn the overall structure and distribution characteristics of the data; as the training progressed, the weight of the KL divergence loss was gradually increased, prompting the model to learn more refined and decoupled potential representations.
[0126] This annealing strategy can balance the reconstruction quality and the regularity of the potential representation at different stages of training, effectively avoiding the problem that the KL divergence loss dominates the training process too early, resulting in a decrease in the model's generation ability.
[0127] Secondly, a regularization method based on contrastive learning is introduced to maximize the mutual information between positive sample pairs and minimize the mutual information between negative sample pairs, so as to encourage the model to learn more compact and discriminative latent representations.
[0128] Specifically, for each training sample, a positive sample pair (i.e., two different data augmented versions of the same time series data) and multiple negative sample pairs (i.e., pairings between different time series data) are constructed.
[0129] Then, the mutual information between positive and negative sample pairs is calculated in the latent space and added to the training objective as an additional regularization term.
[0130] By minimizing the mutual information between positive sample pairs and maximizing the mutual information between negative sample pairs, the model can learn more robust and discriminative latent representations, improving the generalization ability of the time series VAE model.
[0131] Finally, in order to further enhance the adaptability of the time series VAE model to complex marine environments, adaptive noise injection technology is used.
[0132] Specifically, the intensity of the noise is adaptively adjusted according to the complexity and prediction difficulty of each time series data.
[0133] For more complex and difficult to predict time series data, inject more noise to increase the robustness of the model; for simpler and easier to predict time series data, inject less noise to avoid over-regularization.
[0134] Through this adaptive noise injection technology, the time series VAE model can dynamically adapt to ocean environment data of different complexities, improving the robustness and generalization performance of the model.
[0135] In summary, this application introduces in detail the construction process and key technologies of the time series VAE model in the marine environment rolling prediction scheme.
[0136] By introducing LSTM network, attention mechanism, KL divergence loss annealing strategy, contrastive learning regularization and adaptive noise injection technology, the time series VAE model can effectively learn the low-dimensional potential representation of the spatiotemporal sequence data of the marine environment and capture the inherent laws and dynamic patterns of the evolution of the marine environment.
[0137] This provides a more compact and informative feature representation for subsequent prediction tasks, significantly improving the accuracy and reliability of marine environment predictions.
[0138] In an optional embodiment,
[0139] Construct an attention-enhanced spatiotemporal graph neural network model, model the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregate the local features of nodes through graph convolution operations, and use a multi-head attention mechanism to dynamically adjust the weights between different nodes to model the spatial correlation patterns of the marine environment, including:
[0140] The distribution of physical quantities of the marine environment is represented as a space-time graph, wherein the space-time graph is composed of M nodes, wherein the M nodes represent observation points at different locations, wherein the edges between the nodes of the space-time graph represent the spatial associations between the nodes, and the node attribute tensor of the space-time graph represents the changes of the node attributes over time;
[0141] Performing a graph convolution operation on the spatiotemporal graph to update the features of each node by aggregating the features of neighboring nodes of each node in the spatiotemporal graph;
[0142] On the basis of the graph convolution operation, a multi-head attention mechanism is introduced, and the attention weight of each attention head is obtained by calculating the evaluation value of the association strength between nodes by different attention heads. The features of neighboring nodes in the graph convolution operation are weightedly aggregated using the attention weight to model the dynamic association pattern between nodes in the spatiotemporal graph.
[0143] On top of the graph convolution operation enhanced by the attention mechanism, a temporal convolutional network is introduced to capture the temporal evolution of the attributes of the nodes in the spatiotemporal graph through the combination of causal convolution and dilated convolution. The temporal convolutional network is combined with the graph convolution operation enhanced by the attention mechanism to construct a spatiotemporal graph neural network model.
[0144] The mean square error loss function is used to measure the difference between the predicted value and the true value of the spatiotemporal graph neural network model. The spatiotemporal graph neural network model is trained by minimizing the mean square error loss function and introducing the fluid dynamics equation as a regularization term to obtain the trained spatiotemporal graph neural network model.
[0145] In an optional embodiment,
[0146] By minimizing the mean square error loss function and introducing the fluid dynamics equation as a regularization term, the spatiotemporal graph neural network model is trained including:
[0147] The fluid dynamics equation is shown in the following formula:
[0148] ;
[0149] in, u represents the fluid velocity vector, t Indicates time, ρ represents the fluid density, p Indicates pressure, ν represents the kinematic viscosity of the fluid, f Represents the external force term.
[0150] Exemplarily, a self-supervised learning paradigm is introduced, time series reconstruction tasks and graph structure reconstruction tasks are designed, and unlabeled data are fully utilized for pre-training.
[0151] Time series reconstruction task: Randomly block certain time steps in the time series fragment, input the blocked time series fragment into the spatiotemporal graph neural network, and require the network to reconstruct the complete time series fragment.
[0152] By minimizing the reconstruction error, the network learns the intrinsic patterns and dynamic characteristics of time series data.
[0153] Graph structure reconstruction task: Randomly remove some edges in the space-time graph, input the incomplete graph structure into the space-time graph neural network, and require the network to predict the existence probability of the removed edges.
[0154] By minimizing the prediction error, the network learns the topological features of graph structured data and the association patterns between nodes.
[0155] The time series reconstruction task and the graph structure reconstruction task are jointly used for self-supervised pre-training to learn more general and robust feature representations, providing better initialization for subsequent supervised learning tasks.
[0156] The graph attention network (GAT) is used to replace the multi-head attention mechanism to further enhance the spatial modeling ability of the spatiotemporal graph neural network.
[0157] By using the graph attention layer in GAT, the weights between nodes are adaptively calculated by introducing the attention mechanism to better capture the association patterns and long-range dependencies between nodes.
[0158] The attention weights in GAT are dynamically calculated based on node features, which can better adapt to changes in node features and capture dynamic association patterns between nodes.
[0159] The attention mechanism in GAT is calculated based on the features of node pairs, rather than on the features of a single node like the multi-head attention mechanism, so it can more directly model the interactions and dependencies between nodes.
[0160] The spatiotemporal attention mechanism is introduced to simultaneously consider the attention weights of the time dimension and the spatial dimension, thereby improving the model's ability to model spatiotemporal dependencies.
[0161] In the time dimension, according to the correlation between the current time step and the historical time step, the attention weights of different time steps are calculated to highlight the historical information that is strongly related to the current time step.
[0162] In the spatial dimension, according to the correlation between the current node and its neighbor nodes, the attention weights of different neighbor nodes are calculated to highlight the neighbor information that is strongly correlated with the current node.
[0163] The attention weights of the time dimension and the space dimension are combined to obtain the spatiotemporal attention weights, which are used to weight the features of nodes in the spatiotemporal graph to better model spatiotemporal dependencies.
[0164] A physical information fusion module is designed to introduce the physical mechanism and prior knowledge of the marine environment into the spatiotemporal graph neural network to improve the physical rationality of the prediction results.
[0165] Construct partial differential equations that describe the physical processes of the marine environment, such as fluid dynamics equations, convection-diffusion equations, etc., discretize them and embed them into the calculation process of the spatiotemporal graph neural network.
[0166] A physical information fusion module is introduced into the middle layer of the spatiotemporal graph neural network. Through residual connection, the calculation results of the physical equations are fused with the features learned by the network, so that the prediction results of the network conform to the constraints of physical laws.
[0167] Physical regularization terms are introduced into the loss function, such as the residual term of the fluid dynamics equation. By minimizing the physical regularization terms, the prediction results of the spatiotemporal graph neural network can meet the constraints of the physical equations and improve the physical rationality of the prediction results.
[0168] Transfer learning and domain adaptation techniques are used to improve the generalization ability of the spatiotemporal graph neural network model in different ocean regions and environmental conditions.
[0169] The spatiotemporal graph neural network model pre-trained on the source domain (such as a specific ocean area) is used to quickly adapt to the data distribution and characteristics of the target domain (such as other ocean areas) through fine-tuning.
[0170] Domain adversarial training technology is introduced, and adversarial loss is used to enable the spatiotemporal graph neural network to learn domain-invariant feature representations, reduce the distribution differences between the source domain and the target domain, and improve the cross-domain generalization ability of the model.
[0171] Design adaptive loss functions and regularization terms, and dynamically adjust the weights of loss functions and regularization terms according to the data characteristics and physical properties of the target domain, so that the spatiotemporal graph neural network model can better adapt to the environmental conditions of the target domain.
[0172] By introducing the above-mentioned optimization and improvement techniques, the performance and generalization ability of the spatiotemporal graph neural network model in the task of predicting the distribution of physical quantities in the marine environment can be further improved, the spatiotemporal evolution laws of the marine environment can be better modeled, and more accurate and reliable decision-making support can be provided for the monitoring and early warning of the marine environment.
[0173] In an optional embodiment,
[0174] Adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weighted and aggregated to obtain the final prediction results of the spatiotemporal distribution of marine environmental physical quantities for multiple time steps in the future, including:
[0175] Probabilistic modeling is performed on the output results of the temporal variational autoencoder model and the output results of the spatiotemporal graph neural network model enhanced by the attention mechanism, and they are regarded as a probability distribution. Through the Bayesian inference framework, based on the historical observation data and the model output results, the posterior distribution of each model is calculated using the Bayesian theorem to obtain the posterior estimates and uncertainty measures of the model parameters, and the posterior distribution is approximately calculated through Monte Carlo sampling or variational inference methods to obtain the probability distribution of the model output results.
[0176] According to the posterior probability of each model, its weight in the fusion process is calculated, and the spatiotemporal attention mechanism is introduced. According to the feature representation of different time steps and spatial positions, the fusion weight of each model at different spatiotemporal positions is dynamically adjusted. An adaptive learning rate strategy is adopted to dynamically adjust the update step size of the fusion weight according to the error between the model fusion result and the observed data.
[0177] Using the Bayesian model averaging method, the prediction distributions of different models are weighted averaged according to the posterior probability of each model and the adaptive fusion weight to obtain the fused prediction distribution, whose mean is used as the final prediction result and the variance is used as a measure of prediction uncertainty. Based on the fused prediction distribution, the mean of the prediction distribution is temporally interpolated and extrapolated to obtain continuous spatiotemporal distribution prediction results. The variance of the prediction distribution is used to quantify the uncertainty of the prediction results at different spatiotemporal positions and generate confidence intervals or probability maps.
[0178] Exemplarily, multiple basic prediction models are trained, such as spatiotemporal graph neural network models, physics-driven numerical models, etc., to predict the spatiotemporal distribution of physical quantities in the marine environment.
[0179] The output of each basic prediction model is probabilistically modeled and regarded as a probability distribution.
[0180] Common probability distribution forms such as Gaussian distribution and mixed Gaussian distribution can be used.
[0181] The probability distribution parameters and uncertainties of each underlying prediction model are estimated through a Bayesian inference framework.
[0182] Based on historical observation data and model prediction results, the likelihood function and prior distribution of each model are constructed.
[0183] Using Bayes' theorem, the posterior distribution of each model is calculated to obtain the posterior estimates and uncertainty measures of the model parameters.
[0184] Through methods such as Monte Carlo sampling or variational inference, the posterior distribution is approximately calculated to obtain the probability distribution of the model's prediction results.
[0185] According to the posterior probability of each model, its weight in the fusion process is calculated.
[0186] The higher the posterior probability, the more credible the model's prediction results are, and the greater its weight should be in the fusion.
[0187] The spatiotemporal attention mechanism is introduced to dynamically adjust the fusion weights of each model at different spatiotemporal positions according to the feature representations of different time steps and spatial positions.
[0188] An adaptive learning rate strategy is adopted to dynamically adjust the update step size of the fusion weight according to the error between the model fusion result and the observed data to achieve adaptive optimization of the fusion strategy.
[0189] The Bayesian model averaging method is used to perform weighted averaging on the output probability distributions of different basic prediction models to obtain the fused prediction distribution.
[0190] The prediction distributions of different models are weighted averaged according to the posterior probability of each model and the adaptive fusion weights.
[0191] The fused prediction distribution is obtained, with its mean as the final prediction result and the variance as a measure of prediction uncertainty.
[0192] Based on the fused prediction distribution, the prediction results of the spatiotemporal distribution of marine environmental physical quantities for multiple future time steps are output.
[0193] The mean of the predicted distribution is interpolated and extrapolated in time and space to obtain continuous time and space distribution prediction results.
[0194] The variance of the prediction distribution is used to quantify the uncertainty of the prediction results at different spatial and temporal locations, and to generate confidence intervals or probability maps.
[0195] Introduce online learning mechanism to achieve continuous optimization and adaptive updating of fusion strategies.
[0196] When new observations arrive, the probability distribution parameters and uncertainty estimates of each underlying forecast model are updated.
[0197] According to the updated model posterior probability and fusion weight, the fusion strategy is dynamically adjusted to adapt to the latest data distribution and model performance.
[0198] By adopting the adaptive fusion strategy of Bayesian model averaging, the advantages of multiple basic prediction models can be effectively combined to improve the accuracy and reliability of the prediction of the spatiotemporal distribution of marine environmental physical quantities.
[0199] At the same time, through probabilistic modeling and Bayesian inference, the uncertainty of the prediction results can be quantified, providing more comprehensive information support for decision-making.
[0200] In addition, the introduction of adaptive fusion strategies and online learning mechanisms enables the fusion model to continuously optimize and adapt to the changing ocean environment, maintaining long-term prediction performance and practical value.
[0201] Figure 2 FIG. 1 is a schematic diagram of the structure of a short-term marine environment prediction system based on a conditional adversarial network according to an embodiment of the present invention. Figure 2 As shown, the system comprises:
[0202] The first unit is used to construct a time series variational autoencoder model, which includes an encoder and a decoder. The encoder uses a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder uses a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state;
[0203] The second unit is used to build a spatiotemporal graph neural network model enhanced by an attention mechanism. The distribution of physical quantities in the marine environment is modeled as a spatiotemporal graph. The local features of the nodes are aggregated through graph convolution operations, and the weights between different nodes are dynamically adjusted using a multi-head attention mechanism to model the spatial correlation pattern of the marine environment.
[0204] The third unit is used for the time series variational autoencoder model to generate potential representations for multiple future time steps based on the ocean environment status at the current moment and future condition settings, and the attention mechanism enhanced spatiotemporal graph neural network model to predict the evolution of the ocean environment map for multiple future time steps based on the ocean environment map at the current moment and future condition settings; an adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weightedly aggregated to obtain the final prediction results of the spatiotemporal distribution of ocean environment physical quantities for multiple future time steps.
[0205] According to a third aspect of the embodiments of the present invention,
[0206] An electronic device is provided, comprising:
[0207] processor;
[0208] a memory for storing processor-executable instructions;
[0209] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0210] According to a fourth aspect of the embodiments of the present invention,
[0211] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0212] The present invention may be a method, an apparatus, a system and / or a computer program product.
[0213] The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.
[0214] 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 by 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 short-term prediction method for marine environment based on conditional adversarial network, characterized in that: include: Constructing a temporal variational autoencoder model, the temporal variational autoencoder model comprising an encoder and a decoder, the encoder using a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder using a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state; Construct an attention-enhanced spatiotemporal graph neural network model, model the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregate the local features of nodes through graph convolution operations, and use a multi-head attention mechanism to dynamically adjust the weights between different nodes to model the spatial correlation pattern of the marine environment; The temporal variational autoencoder model generates a potential representation of multiple future time steps based on the ocean environment state at the current moment and future condition settings, and the attention mechanism enhanced spatiotemporal graph neural network model predicts the evolution of the ocean environment map at multiple future time steps based on the ocean environment map at the current moment and future condition settings; an adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weighted and aggregated to obtain the final prediction results of the spatiotemporal distribution of ocean environment physical quantities at multiple future time steps; Constructing a temporal variational autoencoder model, the temporal variational autoencoder model includes an encoder and a decoder, the encoder uses a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder uses a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state, including: Constructing a temporal variational autoencoder model of an encoder-decoder architecture, wherein the encoder uses a multi-layer long short-term memory network to encode the spatiotemporal sequence data of the marine environment in the target area. At each time step, the long short-term memory network unit receives the marine environment state vector at the current moment, and updates the hidden state and memory unit at the current moment in combination with the hidden state and memory unit at the previous moment, thereby capturing the contextual information and long-range dependencies in the time series data; The parameters of the encoder and the decoder are optimized by maximizing the variational lower bound, the output of the encoder is passed through two fully connected layers to estimate the mean and logarithmic variance of the latent variables respectively, the latent variables are sampled from the estimated Gaussian distribution and input into the decoder for reconstruction; The decoder adopts a long short-term memory network structure that is symmetrical with the encoder, recursively generates a sequence of ocean environment states based on latent variables and decoding results at the previous moment, and uses conditional Gaussian distribution as a generation model in the output layer of the decoder, and optimizes the parameters of the encoder and the decoder by minimizing reconstruction error and KL divergence loss.
2. The method according to claim 1, characterized in that Optimizing the parameters of the encoder and the decoder by minimizing the reconstruction error and the KL divergence loss includes: The reconstruction error is expressed as follows: ; The KL divergence loss is shown in the following formula: ; in, L R represents the reconstruction error, L K represents the KL divergence loss, N , T represent the sample size and time period respectively, Indicates i samples at time step t The observation vector of Indicates i The latent variable corresponding to the sample is Indicates i samples at time step t The previous observation sequence, μ θ () represents the mean of the decoder estimate, Σ θ () represents the covariance matrix estimated by the decoder; z represents the latent variable, represents the posterior distribution estimated by the encoder, which represents the conditional probability distribution of the latent variable given a sequence of observations, p(z) represents the prior distribution of the latent variable.
3. The method according to claim 1, characterized in that Construct an attention-enhanced spatiotemporal graph neural network model, model the distribution of physical quantities in the marine environment as a spatiotemporal graph, aggregate the local features of nodes through graph convolution operations, and use a multi-head attention mechanism to dynamically adjust the weights between different nodes to model the spatial correlation patterns of the marine environment, including: The distribution of physical quantities of the marine environment is represented as a space-time graph, wherein the space-time graph is composed of M nodes, wherein the M nodes represent observation points at different locations, wherein the edges between the nodes of the space-time graph represent the spatial associations between the nodes, and the node attribute tensor of the space-time graph represents the changes of the node attributes over time; Performing a graph convolution operation on the spatiotemporal graph to update the features of each node by aggregating the features of neighboring nodes of each node in the spatiotemporal graph; On the basis of the graph convolution operation, a multi-head attention mechanism is introduced, and the attention weight of each attention head is obtained by calculating the evaluation value of the association strength between nodes by different attention heads. The features of neighboring nodes in the graph convolution operation are weightedly aggregated using the attention weight to model the dynamic association pattern between nodes in the spatiotemporal graph. On top of the graph convolution operation enhanced by the attention mechanism, a temporal convolutional network is introduced to capture the temporal evolution of the attributes of the nodes in the spatiotemporal graph through the combination of causal convolution and dilated convolution. The temporal convolutional network is combined with the graph convolution operation enhanced by the attention mechanism to construct a spatiotemporal graph neural network model. The mean square error loss function is used to measure the difference between the predicted value and the true value of the spatiotemporal graph neural network model. The spatiotemporal graph neural network model is trained by minimizing the mean square error loss function and introducing the fluid dynamics equation as a regularization term to obtain the trained spatiotemporal graph neural network model.
4. The method according to claim 3, characterized in that: Introducing fluid dynamics equations as regularization terms to train the spatiotemporal graph neural network model includes: The fluid dynamics equation is shown in the following formula: ; in, u represents the fluid velocity vector, t Indicates time, ρ represents the fluid density, p Indicates pressure, ν represents the kinematic viscosity of the fluid, f Represents the external force term.
5. The method according to claim 1, characterized in that Adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weighted and aggregated to obtain the final prediction results of the spatiotemporal distribution of marine environmental physical quantities for multiple time steps in the future, including: Probabilistic modeling is performed on the output results of the temporal variational autoencoder model and the output results of the spatiotemporal graph neural network model enhanced by the attention mechanism, and they are regarded as a probability distribution. Through the Bayesian inference framework, based on the historical observation data and the model output results, the posterior distribution of each model is calculated using the Bayesian theorem to obtain the posterior estimates and uncertainty measures of the model parameters, and the posterior distribution is approximately calculated through Monte Carlo sampling or variational inference methods to obtain the probability distribution of the model output results. According to the posterior probability of each model, its weight in the fusion process is calculated, and the spatiotemporal attention mechanism is introduced. According to the feature representation of different time steps and spatial positions, the fusion weight of each model at different spatiotemporal positions is dynamically adjusted. An adaptive learning rate strategy is adopted to dynamically adjust the update step size of the fusion weight according to the error between the model fusion result and the observed data. Using the Bayesian model averaging method, the prediction distributions of different models are weighted averaged according to the posterior probability of each model and the adaptive fusion weight to obtain the fused prediction distribution, whose mean is used as the final prediction result and the variance is used as a measure of prediction uncertainty. Based on the fused prediction distribution, the mean of the prediction distribution is temporally interpolated and extrapolated to obtain continuous spatiotemporal distribution prediction results. The variance of the prediction distribution is used to quantify the uncertainty of the prediction results at different spatiotemporal positions and generate confidence intervals or probability maps.
6. A marine environment short-term prediction system based on a conditional adversarial network, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to construct a time series variational autoencoder model, which includes an encoder and a decoder. The encoder uses a long short-term memory network to map the spatiotemporal sequence data of the marine environment of the target area to a low-dimensional latent space to obtain a compressed representation of the marine environment state, and the decoder uses a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state; The second unit is used to build a spatiotemporal graph neural network model enhanced by an attention mechanism. The distribution of physical quantities in the marine environment is modeled as a spatiotemporal graph. The local features of the nodes are aggregated through graph convolution operations, and the weights between different nodes are dynamically adjusted using a multi-head attention mechanism to model the spatial correlation pattern of the marine environment. The third unit is used for the time series variational autoencoder model to generate potential representations for multiple future time steps based on the ocean environment status at the current moment and future condition settings, and the attention mechanism enhanced spatiotemporal graph neural network model to predict the evolution of the ocean environment map for multiple future time steps based on the ocean environment map at the current moment and future condition settings; an adaptive fusion strategy is used to dynamically adjust the fusion weights of the two models, and the output results of the two models are weightedly aggregated to obtain the final prediction results of the spatiotemporal distribution of ocean environment physical quantities for multiple future time steps.
7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.
8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.
Citation Information
Patent Citations
Intention trajectory prediction method combined with social generative adversarial network
CN111930110A
Pedestrian trajectory prediction method based on conditional variation generative adversarial network
CN114580718A
Sea surface temperature prediction method fusing deep learning and physical mechanism
CN117252290A
Unsupervised generator bearing fault detection method based on variational auto-encoder
CN118310746A
Predictive channel modeling method based on generative adversarial network and long short-term memory artificial neural network
US20240259121A1
Cited By
Offshore wind power operation environment scene construction method considering typhoon forecast error
CN120409210A
