Multi-scale physical information fusion large language model sea surface temperature prediction method and system

Abstract

The present application belongs to the technical field of sea surface temperature prediction, and particularly relates to a multi-scale physical information fusion large language model sea surface temperature prediction method and system; the method comprises the following steps: inputting historical sea surface temperature data, preprocessing, generating sea temperature input data fused with spatial geographic features, and performing multi-level and multi-scale spatial and temporal feature extraction to obtain physical features; obtaining a pre-constructed marine field knowledge library, obtaining word vectors from a pre-trained language model, linearly mapping text semantic features, fusing the physical features and the text semantic features in the same feature space to obtain joint semantic features; constructing prompt word vectors related to the prediction task, inputting the joint semantic features into a large language model, guiding the model to perform deep feature interpretation and prediction through the prompt word vectors, and outputting the prediction result of the future sea temperature field; thereby the complex spatio-temporal change information of the sea temperature can be mined, and high-precision spatio-temporal prediction of the sea temperature field can be realized.

Description

Technical Field

[0001] This invention belongs to the field of sea surface temperature prediction technology, specifically involving a large language model method and system for sea surface temperature prediction based on multi-scale physical information fusion. Background Technology

[0002] Sea surface temperature (SST) forecasting is of great significance in many fields, including marine science, marine engineering, climate change research, fisheries resource management, and extreme weather early warning. As an important physical quantity reflecting the interaction between the ocean and the atmosphere, the spatiotemporal evolution of SST not only affects the formation and path of typhoons, but also determines the structure of ocean circulation, the distribution of marine ecosystems, and global climate change trends.

[0003] Most advanced SST prediction methods currently employ data-driven deep learning models. These models extract spatial features of the temperature field using convolutional neural networks (CNNs), then model time-series changes using recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), and finally output the prediction results through fully connected layers. Compared to traditional numerical models, these methods offer advantages in computational efficiency and achieve a certain level of prediction accuracy in some data-intensive regions.

[0004] However, this method still has the following shortcomings in practical applications: Insufficient capture of anomalous changes: When extracting sea surface temperature features, traditional CNN+RNN models are often more sensitive to most areas with gentle spatial distribution, but are not good at capturing anomalous changes at small and medium scales such as warm pools, fronts, and eddies. This can easily lead to underestimation in the prediction of extreme or local events.

[0005] Lack of mechanistic and semantic integration: Existing methods usually rely only on numerical observation data for modeling, failing to fully integrate prior knowledge and dynamic mechanisms in the field of oceanography. This results in insufficient prediction accuracy when the model is dealing with long-term series, cross-regional generalization, or sparse observation areas, making it difficult to achieve high-precision spatiotemporally consistent sea surface temperature prediction. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a large-scale language model-based sea surface temperature prediction method and system that integrates multi-scale physical information fusion. The method primarily employs marine geospatial feature encoding, multi-scale physical feature extraction, multi-modal feature fusion, and task-driven prediction with prompts. It introduces spectral neural operators and a space-time pyramid network to simultaneously enhance feature representation at both the frequency domain and spatial multi-resolution levels. Furthermore, it utilizes an attention mechanism to represent the multi-scale physical features. The domain semantic vectors, i.e., text semantic features, provided by large language models Deep alignment and interaction are performed to enable the model to simultaneously reference "data-driven physical evidence," i.e., physical feature representations, when making predictions. And "the mechanism of domain priors" that is, text semantic features This will improve the characterization of complex spatiotemporal changes and the accuracy of predicting future sea surface temperature fields; To achieve the above objectives, the first aspect of this invention proposes a sea surface temperature prediction method based on a large language model that integrates multi-scale physical information fusion, comprising: S1, Marine Geospatial Information Coding: Collect and input historical sea surface temperature data Spatial geographic location normalization and spatial location coding enhancement are performed to generate sea surface temperature input data. ; S2. Multi-scale physical feature extraction for ocean dynamic processes: Sea surface temperature input data Entering the frequency domain feature modeling branch, the backbone features are extracted using spectral neural operators. Main features Entering the multi-scale spatiotemporal feature extraction branch, a multi-scale physical feature set is extracted through a space-time pyramid network. By integrating through the fusion module, a unified physical feature representation is obtained. ; S3. Multimodal feature fusion based on marine domain knowledge: Obtain a pre-built marine domain knowledge lexicon and a word embedding matrix based on a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon, and perform linear mapping on the word vectors to obtain text semantic features. Physical characteristics representation Textual semantic features By fusing features in the same feature space, joint semantic features are obtained. ; S4. Large language model prediction based on task prompt words: Constructing prompt word vectors The feature of the prompt word is obtained through embedding. cue word features With joint semantic features After concatenation, the data is input into a large language model for deep feature interpretation and prediction, outputting the predicted results of the future sea surface temperature field. .

[0007] Furthermore, the collection and input of historical sea surface temperature data... Spatial geographic location normalization and spatial location coding enhancement are performed to generate sea surface temperature input data. ,include: Collect and input historical sea surface temperature data Based on the actual latitude and longitude coordinates of the observation points in the ocean region, historical sea surface temperature data were analyzed. The latitude and longitude coordinates of each spatial grid point are obtained by normalization. ; Based on latitude and longitude coordinates Geocoded vectors representing the spatial location information of ocean areas are constructed through encoding mapping operations, and these encoded vectors are concatenated to form historical sea surface temperature data. Dimensions, to obtain enhanced input ; The input is enhanced through convolutional projection. Mapping to a uniform feature width yields sea surface temperature input data. .

[0008] Further, sea surface temperature input data Entering the frequency domain feature modeling branch, the backbone features are extracted using spectral neural operators. Main features The process then moves to the multi-scale spatiotemporal feature extraction branch, where a multi-scale physical feature set is extracted using a spatial-temporal pyramid network. This set is then integrated by a fusion module to obtain a unified physical feature representation. ,include: S2.1: Sea surface temperature input data Entering the frequency domain feature modeling branch, the frequency domain feature modeling branch takes sea surface temperature input data as input. Perform Fourier transform Mapping to the frequency domain yields the frequency domain representation. Represented in the frequency domain Based on the frequency index of each frequency component, a low-frequency mask is constructed. and intermediate frequency mask The low-frequency components characterizing large-scale background changes and the mid-frequency components characterizing mesoscale spatial changes were selected to obtain the selected frequency domain features. ; Learnable complex frequency domain weights are applied to the filtered low-frequency and mid-frequency components, and then the data is restored to the physical domain using inverse Fourier transform, and compared with sea surface temperature input data. The residual mapping results are fused and then subjected to nonlinear activation to obtain frequency domain features. .

[0009] S2.2: Frequency Domain Characteristics After amplitude and phase co-reconstruction, the reconstruction results are used to form a gated map score through residual superposition. The gated map score is then compared with the sea surface temperature input data. The core features are obtained by fusing pointwise multiplication and normalization and convolution. ; S2.3: Main Characteristics Entering the multi-scale spatiotemporal feature extraction branch, this branch extracts the main features along the spatial path using a space-time pyramid network. Hierarchical downsampling is used to construct multi-resolution feature maps. Along the time path, temporal convolutional groups with different dilation rates are introduced at each scale. After layer normalization and the Gule activation function, the maps are uniformly projected to the same number of channels, resulting in a multi-scale physical feature set. ,in, The number of scales representing multiple physical features. Indicates scale index; S2.4: Multi-scale physical space collection Entering the fusion module, the fusion module combines the physical features at various scales. Global average pooling is used to generate channel descriptors, which are then passed through convolutional layers and a sigmoid activation function to obtain initial weights. Then, normalized weights are obtained through Softmax regularization. Adaptive weighted fusion is performed to obtain the physical feature representation. .

[0010] Furthermore, a pre-constructed marine domain knowledge lexicon is obtained, based on the word embedding matrix of a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon, and perform linear mapping on the word vectors to obtain text semantic features. Physical characteristics representation Textual semantic features By fusing features in the same feature space, joint semantic features are obtained. ,include: S3.1: Obtain a pre-built marine domain knowledge lexicon, based on the word embedding matrix of a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon. Perform dimensionality alignment and reduction on the word vectors using linear mapping to obtain the text semantic features. ; S3.2: Physical Feature Representation Textual semantic features The mapped text semantic feature representation is obtained by mapping to a unified cross-modal aligned feature space through both physical feature embedding layers and text feature embedding layers. With the mapped physical feature representation ; S3.3: Calculate the query vector Query, key vector Key, and value vector Value using adaptive weights, as shown in the following expression: ; ; ; in, The query vector Query is represented by physical feature embeddings. Mapping generation, The key vectors Key and Key represent The value vector, Value, is represented by embedding textual semantic features. Mapping generation; , , These represent the learnable linear mapping matrices used to generate queries, keys, and values, respectively. S3.4: Utilize a cross-modal attention fusion mechanism to obtain the final fused features, i.e., joint semantic features. The expression is as follows: ; in, Represents the query vector in the cross-modal attention mechanism. With key vector The feature dimensions.

[0011] Furthermore, construct the prompt word vectors. The feature of the prompt word is obtained through embedding. cue word features With joint semantic features After concatenation, the data is input into a large language model for deep feature interpretation and prediction, outputting the predicted results of the future sea surface temperature field. ,include: S4.1: Construct task prompt words based on the statistical characteristics of the prediction task-related parameters and sea surface temperature input data. Then, use a word segmenter and word vector embedding to obtain prompt word vectors. Then, the prompt word features are obtained through linear layer mapping. Among them, the parameters related to the prediction task include: predictor variables, target sea area, historical input time window, prediction step size, and output spatial resolution. The statistical characteristics of the sea surface temperature input data include: mean, variance, maximum value, minimum value, and trend. S4.2: Features of Prompt Words With joint semantic features The input sequence is spliced ​​together to form the fused sequence. ; S4.3: Input Sequence Input a large language model, perform sequence-level inference, and output intermediate representations. ; S4.4: Intermediate Characterization After linear mapping and activation, the original spatial resolution is restored through an upsampling layer to obtain the final temperature prediction field. .

[0012] Furthermore, the original spatial resolution is restored through an upsampling layer to obtain the final temperature prediction field. ,include: The upsampling layer uses a combination of multi-layer transposed convolution and bilinear interpolation to progressively restore spatial resolution. Each upsampling process includes linear mapping and nonlinear activation operations, ultimately resulting in a temperature prediction field. The expansion form based on layer-by-layer upsampling is as follows: ; in, Indicates the number of upsampling layers. Indicates the first The upsampling operator for the layer can be transposed convolution or bilinear interpolation. This represents the corresponding convolution kernel parameters. This represents the convolution operation. This represents the GELU activation function.

[0013] The second aspect of this invention proposes a large-scale language model sea surface temperature prediction system based on multi-scale physical information fusion, comprising: The marine geospatial information coding module is used for geographic location coding: collecting and inputting historical sea surface temperature data. Spatial geographic location normalization and spatial location coding enhancement are performed to generate sea surface temperature input data. ; A multi-scale physical feature extraction module for ocean dynamic processes is used for multi-scale physical feature extraction: sea surface temperature input data. Entering the frequency domain feature modeling branch, the backbone features are extracted using spectral neural operators. Main features Entering the multi-scale spatiotemporal feature extraction branch, a multi-scale physical feature set is extracted through a spatial-temporal pyramid network. Multi-scale physical feature set By integrating through the fusion module, a unified physical feature representation is obtained. ; A multimodal feature fusion module based on marine domain knowledge is used for multimodal feature fusion of marine domain knowledge: acquiring a pre-built marine domain knowledge lexicon, and using the word embedding matrix based on a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon, and perform linear mapping on the word vectors to obtain text semantic features. Physical characteristics representation Textual semantic features By fusing features in the same feature space, joint semantic features are obtained. ; The task-prompt-driven large language model prediction module is used for task-prompt-driven large language model prediction: constructing prompt word vectors. The feature of the prompt word is obtained through embedding. cue word features With joint semantic features After concatenation, the data is input into a large language model for deep feature interpretation and prediction, outputting the predicted results of the future sea surface temperature field. .

[0014] A third aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the large language model sea surface temperature prediction method as described above, which integrates multi-scale physical information.

[0015] Compared with the prior art, the advantages of the present invention are as follows: This invention enhances the ability to capture anomalous changes: Traditional modeling methods based on Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) often rely more on the dominant background field features with large-scale, gently distributed patterns when processing sea surface temperature fields, while lacking sensitivity to small- and medium-scale changes with strong non-uniformity, such as warm pools, fronts, and eddies. Since these regions typically only account for a small portion of the entire sea surface temperature field, their anomalous signals are easily diluted or even ignored during conventional convolution, leading to systematic underestimation of extreme events or local anomalies in the model's predictions. This invention introduces spectral neural operators and a space-time pyramid network to simultaneously enhance feature representation at both the frequency domain and spatial multi-resolution levels, enabling the model to effectively preserve and amplify anomalous change features, thereby significantly improving the prediction accuracy for extreme and local events.

[0016] This invention achieves the fusion of mechanistic information and semantic knowledge: most existing methods rely solely on numerical observation data for fitting and modeling, lacking explicit utilization of oceanographic knowledge and dynamic mechanisms. Therefore, they are significantly inadequate in long-term series prediction, cross-regional generalization, and inference over sparsely observed regions. This invention uses an attention mechanism to deeply align and interact with multi-scale physical features and domain semantic vectors provided by a large language model, enabling the model to simultaneously reference both "data-driven physical evidence" and "domain-prior mechanistic explanations" when making predictions. This dual-constraint mechanism not only improves the stability of long-term series modeling but also enhances the model's generalization ability in different sea areas and complex scenarios, thereby achieving high-precision, spatiotemporally consistent sea surface temperature prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the structure of the large language model sea surface temperature prediction system for multi-scale physical information fusion according to the present invention; Figure 2 This is a schematic diagram of the multi-scale physical feature extraction module for ocean dynamic processes according to the present invention; Figure 3 This is a schematic diagram of the multimodal feature fusion module based on marine domain knowledge of the present invention; Figure 4 This is a schematic diagram of the large language model prediction structure based on task prompt words according to the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0020] Example 1: See Figures 2-4 This embodiment provides a large language model for sea surface temperature prediction based on multi-scale physical information fusion, specifically including the following steps: S1, Marine Geospatial Information Coding: Collect and input historical sea surface temperature data Spatial geographic location normalization and spatial location coding enhancement are performed to generate sea surface temperature input data. ; S2. Multi-scale physical feature extraction for ocean dynamic processes: Sea surface temperature input data Entering the frequency domain feature modeling branch, through the core features of the spectral neural operator... Main features Entering the multi-scale spatiotemporal feature extraction branch, a multi-scale physical feature set is extracted through a space-time pyramid network. By integrating through the fusion module, a unified physical feature representation is obtained. ; S3. Multimodal feature fusion based on marine domain knowledge: Obtain a pre-built marine domain knowledge lexicon and a word embedding matrix based on a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon, and perform linear mapping on the word vectors to obtain text semantic features. Physical characteristics representation Textual semantic features By fusing features in the same feature space, joint semantic features are obtained. ; S4. Large language model prediction based on task prompt words: Constructing prompt word vectors Through embedding processing (i.e. Figure 4 (The prompt word mapping in the middle) obtains prompt word features cue word features With joint semantic features After concatenation, the data is input into a large language model for deep feature interpretation and prediction, outputting the predicted results of the future sea surface temperature field. .

[0021] Specifically, step S1 in this embodiment includes: Collect and input historical sea surface temperature data Based on the actual latitude and longitude coordinates of the observation points in the ocean region, in order to incorporate prior geospatial information, historical sea surface temperature data were analyzed. The latitude and longitude coordinates of each spatial grid point are obtained by normalization. , is represented as: ; Among them, historical sea surface temperature data The dimensions are T×H×W, where T represents the length of the historical time series, and H and W are the dimensions of the spatial grid points in the latitudinal and longitudinal directions, respectively. The first , The latitude and longitude of each grid point , , , These represent the minimum and maximum latitude and longitude of the sea area, respectively. This represents the normalized latitude and longitude coordinates corresponding to each spatial grid point; Based on latitude and longitude coordinates Geocoded vectors representing the spatial location information of ocean areas are constructed through encoding mapping operations, and these geocoded vectors are concatenated to form historical sea surface temperature data. The last dimension yields the enhanced input. , is represented as: ; ; in, Represents a geocoded vector, historical sea surface temperature data. The dimensions are T×H×W, where T represents the length of the historical time series, and H and W represent the spatial dimensions of the grid points in the sea area, respectively. To align with the channel dimensions of the subsequent physical feature extraction backbone network, the enhanced input is projected using a 1×1 convolutional projection. Mapping to a uniform feature width yields sea surface temperature input data. The expression is as follows: ; in, express convolution, Indicates the feature width; The marine geospatial information encoding operation in this embodiment ensures the identification of features from different sources (historical sea surface temperature data). Geographic information coding vector It can be modeled in the same feature space, while helping the model distinguish the differences in dynamic environments in different regions, improving the accuracy of prediction and cross-regional generalization ability.

[0022] For details, see Figure 2 In this embodiment, step S2 includes: Extracting backbone features using spectral neural operators It can directly capture the global coupling relationship between long-distance grid points, making up for the deficiency of convolutional networks that can only expand the receptive field by stacking layers. Specifically: S2.1: Sea surface temperature input data Entering the frequency domain feature modeling branch, the frequency domain feature modeling branch takes sea surface temperature input data as input. Perform a two-dimensional Fourier transform Mapping to the frequency domain yields the frequency domain representation. Represented in the frequency domain Based on the frequency index of each frequency component, a low-frequency mask is constructed. and intermediate frequency mask Based on low-frequency mask and intermediate frequency mask The low-frequency components characterizing large-scale background changes and the mid-frequency components characterizing mesoscale spatial changes are selected separately to obtain the selected frequency domain features. , is represented as: ; in, This represents element-wise multiplication; Learnable complex frequency domain weights are applied to the filtered low-frequency and mid-frequency components, and then the data is restored to the physical domain using inverse Fourier transform, and compared with sea surface temperature input data. The residual mapping results are fused and then subjected to nonlinear activation to obtain frequency domain features. , is represented as: ; in, This represents the sea surface temperature input data after being enhanced with geolocation coding; This indicates that the sea surface temperature input data has undergone Fourier transform. The resulting frequency domain representation; and These represent low-frequency and mid-frequency masks, respectively, used to extract from... Select low-frequency and mid-frequency components; A learnable complex frequency domain kernel is equivalent to a learnable complex frequency domain weight; This represents element-wise multiplication; Indicates channel by channel convolution; Indicates the convolution operation; The GELU nonlinear activation function; It represents the frequency domain characteristics.

[0023] S2.2: Frequency Domain Characteristics After amplitude and phase co-reconstruction, the reconstruction results are used to form a gated map score through residual superposition. The gated map score is then compared with the sea surface temperature input data. The core features are obtained by fusing pointwise multiplication and normalization and convolution. ; Multi-resolution feature maps are constructed through hierarchical downsampling in the spatial path; in the temporal path, temporal convolutional groups with different dilation rates are configured at each scale to simultaneously cover the "short-medium-long" temporal receptive field; it can adaptively select effective scale features in different scenarios, such as emphasizing low-frequency components when the background is evolving at a large scale, and amplifying high-frequency details in fronts or anomalous events. Specifically: S2.3: Main Characteristics Entering the multi-scale spatiotemporal feature extraction branch, to balance local details and large-scale background, this branch uses a spatial-temporal pyramid network to extract core features along the spatial path. Multi-resolution feature maps are constructed through hierarchical downsampling. Along the temporal path, temporal convolutional groups with different dilation rates are introduced at each scale to simultaneously cover the "short-medium-long" temporal receptive fields. After layer normalization and the GULE activation function, the temporal convolutional groups are uniformly projected to the same number of channels, resulting in a multi-scale physical feature set. ,in, The number of scales representing multiple physical features. Indicates scale index; S2.4: Multi-scale physical space collection Enter the fusion module Fusion module Physical characteristics at various scales The fusion weights for each scale are dynamically generated through global statistics and learnable mapping, enabling the fusion result to automatically adjust the contribution ratio of features at each scale according to the physical feature distribution of different samples. Specifically, channel descriptors are generated through global average pooling, and initial weights are obtained through convolutional layers and the Sigmoid activation function. Initial weights Then, normalized weights are obtained through Softmax regularization. Normalized weights Adaptive weighted fusion is performed to obtain physical feature representations. Right now Figure 4 The multi-scale fusion features in the data are represented as follows: ; ; ; in, Represents a collection of physical spaces at multiple scales The Middle Physical feature representation at each scale; The number of scales representing physical characteristics; Indicates scale index; Represents the set of physical spaces at multiple scales. A multi-scale feature fusion module for integration; This represents the number of physical feature scales involved in the fusion. Indicates the first Each scale physical feature represents the corresponding index. This represents the summation index during the Softmax normalization process. This represents the first generation generated by global average pooling, one-dimensional convolution, and the Softmax activation function. Each scale physical feature represents the corresponding initial weight response. Represents the normalized weights after Softmax regularization, used to characterize the first... Each scale physical feature represents the relative importance in the multi-scale fusion process; The fusion process of the multi-scale feature fusion module in this embodiment ensures that the model can adaptively select the optimal scale features in different scenarios, such as highlighting low-frequency components when predicting large-scale events and emphasizing high-frequency details when predicting extreme weather events.

[0024] Step S3 in this embodiment is specifically as follows: S3.1: To enhance the explanation and constraint capabilities of complex dynamic phenomena, a pre-constructed marine domain knowledge lexicon is obtained, based on the word embedding matrix of a pre-trained large language model. Word vectors corresponding to each term in the knowledge lexicon are extracted. The word vectors are then subjected to dimensionality alignment and dimensionality reduction through linear mapping to obtain the text semantic features. See Figure 3 The semantic features of the ocean domain ensure that the semantic space is consistent with the hidden dimensions of the subsequent cross-modal attention fusion mechanism, which reduces computational costs and avoids information bottlenecks caused by dimensional mismatch between modalities. This is represented as: ; ; in, This represents a learnable linear mapping matrix; This indicates the number of lexical units in the marine domain knowledge lexicon. Represents the original word embedding matrix of a pre-trained large language model. The dimension of the word embedding vector; This represents the dimension of the text semantic features after linear mapping, and is related to the physical features in subsequent cross-modal fusion. The projection dimensions remain consistent.

[0025] In this embodiment, the pre-trained large language model is a large language model trained on authoritative marine scientific literature, observation station descriptions, marine event reports, and other texts.

[0026] S3.2: See Figure 3 Physical characteristics representation Textual semantic features The mapped text semantic feature representation is obtained by mapping to a unified cross-modal aligned feature space through both physical feature embedding layers and text feature embedding layers. Embedded representation of mapped physical features , is represented as: ; ; in, and These represent the physical feature embedding layer and the text feature embedding layer, respectively, used to map features from different modalities to a unified cross-modal aligned feature space. This represents the physical feature embedding representation after mapping through the physical feature embedding layer. This represents the semantic feature representation of the text after mapping through the text feature embedding layer; S3.3: Calculate the query vector Query, key vector Key, and value vector Value using adaptive weights, represented as: ; ; ; in, The query vector Query is represented by physical feature embeddings. Mapping generation, The key vectors Key and Key represent The value vector, Value, is represented by embedding textual semantic features. Mapping generation; , , These represent the learnable linear mapping matrices used to generate queries, keys, and values, respectively. S3.4: Utilize a cross-modal attention fusion mechanism to obtain the final fused features, i.e., joint semantic features. Represented as: ; in, Represents the query vector in the cross-modal attention fusion mechanism. With key vector Feature dimensions, Indicates attention score, This represents the normalization exponential function, used to normalize the obtained attention score and generate an attention weight matrix, so that the sum of the contribution weights of each semantic token to the current physical token is 1.

[0027] Among them, the cross-modal attention fusion mechanism uses physical tokens as query vectors. Using semantic tokens as key vectors K and value vectors V, the parts of the physics that need interpretation retrieve interpretable fragments from the domain knowledge, thus transforming the originally black-box feature transformation into a "semantically readable" alignment process. In this embodiment, attention is performed using multi-head parallel computation (number of heads...). (Configurable), and performs splicing and linear mapping at the output to restore the dimension. This approach balances the complementarity of global and local semantics. Through this cross-modal coupling, the model can leverage semantic priors to make stable inferences when observing sparse or anomalous samples, significantly improving extrapolation robustness.

[0028] See Figure 4 In step S4 of this embodiment, in order to enable the large language model to fully understand and utilize the joint semantic features This invention introduces a task-prompt-driven large language model prediction component into its overall architecture. This component integrates task-prompt-word information synchronously during the feature input stage, thereby constructing prompt-word features closely related to the prediction target. This enhances the model's ability to utilize domain knowledge during the reasoning process and enables the application of joint semantic features. The deep interpretation and prediction generation ensure that the prediction results are consistent with the real observation field at different scales, and are suitable for inference of large-scale and long-term series. See Figure 4 As an example, the task prompts in this embodiment can be functionally divided into several sub-parts and organized in a structured manner at the input end to form clear semantic constraints: First, the task description, used to clarify the "target variable, target step size, and output form," such as "given the past T-step sea surface temperature sequence, predict the sea surface temperature at time Δt in the future"; Second, the time and space constraints, used to indicate the latitude and longitude boundaries, grid step size, and time window involved in this prediction, so that the model follows the specified sea area and time period during the inference phase; Third, the observation statistics, including but not limited to the mean, variance, extreme values, and quantiles within the input window, used to provide the model with stable background scale information and mitigate scale shift; Fourth, the phased change trend, used to characterize the warming / cooling tendency and low-frequency background (e.g., weekly / monthly / seasonal cycles) in the historical window, so as to enhance the grasp of the slow evolution process during the generation phase, and ultimately achieve joint constraint inference of "task semantics - physical evidence."

[0029] Specifically, in this embodiment S4, the following is included: S4.1: Based on the relevant parameters of the prediction task and the sea surface temperature input data The statistical features are used to construct task prompt words, which are then processed by a word segmenter and word vector embedding to obtain prompt word vectors. Then, the prompt word features are obtained through linear layer mapping. The parameters related to the prediction task include: predictor variables, target sea area, historical input time window, prediction step size, output spatial resolution, and sea surface temperature input data. The statistical characteristics include: mean, variance, maximum value, minimum value, and trend; expressed as: ; ; ; in, Indicates task prompts; For prompt word constructor; This represents the predictor variable, namely sea surface temperature; Indicates the latitude and longitude range of the target sea area; Indicates the historical input time window; Indicates the prediction step size; Indicates the output grid or spatial resolution; Indicates sea surface temperature input data Statistical characteristics; Represents word vector embedding, Indicates a word segmenter. and Indicates the parameters of the linear layer; This represents the features of the mapped prompt words.

[0030] S4.2: Features of Prompt Words With joint semantic features The input sequence is spliced ​​together to form the fused sequence. , is represented as: ; in, Indicates a splicing operation; S4.3: Input Sequence Input a large language model, perform sequence-level inference, and output intermediate representations for high-resolution temperature prediction. , is represented as: ; in, Representing a large language model; S4.4: Intermediate Characterization After linear mapping and activation, the original spatial resolution is restored through an upsampling layer to obtain the final temperature prediction field. ; ; in, For GELU activation function, Indicates the upsampling layer. This represents a linear mapping.

[0031] In this embodiment, the upsampling layer uses a combination of multi-layer transposed convolution and bilinear interpolation to progressively restore spatial resolution. Each upsampling process includes linear mapping and nonlinear activation operations, ultimately resulting in a temperature prediction field. The expansion form based on layer-by-layer upsampling is as follows: ; in, Indicates the number of upsampling layers. Indicates the first The upsampling operator for the layer can be transposed convolution or bilinear interpolation. This represents the corresponding convolution kernel parameters. This represents the convolution operation. This represents the GELU activation function.

[0032] Experimental verification: To verify the effectiveness of the method of this invention, this embodiment selects sea surface temperature prediction as the application scenario. The experimental data comes from the ERA5 sea surface temperature reanalysis dataset, which has global coverage and high spatiotemporal resolution and is widely used in climate change assessment, extreme event simulation, and ocean dynamic process research. Considering that this method needs to take into account both large-scale circulation background and small-to-medium-scale anomalous processes, this embodiment selects a certain sea area as the study area.

[0033] In the data preprocessing stage, the daily sea surface temperature (SST) fields are uniformly interpolated onto a fixed grid to construct a continuous spatiotemporal tensor sequence. To form training samples, a sliding window approach is used to truncate the input and output: the model input is the SST field of the past 7 consecutive days, and the output targets are the SST distributions for the next 1, 3, and 7 days. In other words, each sample corresponds to an input tensor and three target tensors with prediction time intervals, and the three are jointly modeled in the channel dimension to simultaneously achieve short-term, medium-term, and long-term sea surface temperature predictions.

[0034] In terms of sample partitioning, the dataset was divided chronologically into a training set (approximately 80%), a validation set (approximately 10%), and a test set (approximately 10%) to ensure no overlap between data from different stages. Model training was performed in parallel on eight NVIDIA RTX 4090 GPUs using the Adam optimizer, with a maximum training epoch of 200 epochs and a batch size of 8. The initial learning rate was 1×10⁻⁶. -4 The learning rate is dynamically adjusted during training using a learning rate scheduling mechanism to avoid overfitting and convergence instability. The mean squared error (MSE) is used as the loss function to measure the difference between predicted and actual observations.

[0035] To objectively evaluate prediction performance, this embodiment employs multiple evaluation metrics, including root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination. Of these, RMSE focuses more on reflecting the overall prediction bias, while MAE describes the mean absolute error level. This characterizes the model's ability to explain real trends. By comparing the performance of various indicators under three scenarios with prediction time intervals of 1 day, 3 days, and 7 days, the stability and accuracy of this method under different prediction scales can be comprehensively evaluated.

[0036] To demonstrate the effectiveness of this invention, U-Net, Conv-LSTM, and LSM were selected for comparative experiments with the method proposed in this invention. Detailed information on the experimental results is shown in Tables 1-3 below, which respectively pertain to RMSE, MAE, and... The comparison.

[0037] Table 1: Comparison of RMSE between the present invention and existing methods

[0038] Table 2: Comparison of MAE between the present invention and existing methods

[0039] Table 3: Comparison of the present invention and existing methods Comparison

[0040] Across all prediction timescales, the method of this invention consistently exhibits the lowest RMSE values, particularly in 1-day, 3-day, and 7-day predictions, with RMSE values ​​of 0.1695, 0.3995, and 0.5901, respectively, significantly outperforming U-Net (0.2093, 0.4318, and 0.5993), Conv-LSTM (0.2974, 0.5119, and 0.7852), and LSM (0.2075, 0.4221, and 0.6407). This demonstrates the significant advantages of this invention in short- to medium-term forecasting, maintaining its leading position even as the forecast period lengthens. Regarding the MAE index, the prediction results of this method for 1 day, 3 days, and 7 days are 0.1079, 0.2622, and 0.4014, respectively, representing reductions of 0.0239, 0.0449, and 0.0125 compared to U-Net, 0.0731, 0.0698, and 0.1265 compared to Conv-LSTM, and 0.0275, 0.0109, and 0.0445 compared to LSM. This indicates that the method also has a stable advantage in reducing absolute bias. In terms of metrics, 0.0275, 0.0109, and 0.0445 are all close to 1 and are generally higher than the comparison methods such as U-Net, Conv-LSTM, and LSM, showing a stronger ability to explain data variance and maintaining better robustness over a longer prediction period.

[0041] To verify the model's predictive ability for the spatiotemporal evolution of sea surface temperature, the sea surface temperature thermograms of this invention were compared with those of Conv-LSTM and LSM at three prediction scales: 1 day, 3 days, and 7 days. In the 1-day prediction, the color band distribution and nearshore-outshore temperature transition of this invention were more consistent with the actual field, and the orientation and range of the main isotherms matched better, resulting in a better overall fit to observations. As the prediction period was extended to 3 days and 7 days, all methods showed some degradation, but this invention maintained a more consistent gradient and a more reasonable spatial structure, with less color blocking and boundary blurring. The consistency of the predicted field across different time points was better, demonstrating better stability and reliability. This invention more closely approximates the actual distribution in the spatiotemporal reconstruction of sea surface temperature on short-term to weekly scales and exhibits stronger generalization ability and robustness under different scenarios.

[0042] Example 2: As another embodiment of the present invention, a large-scale language model sea surface temperature prediction system based on multi-scale physical information fusion is also proposed to implement the aforementioned large-scale language model sea surface temperature prediction method based on multi-scale physical information fusion, comprising: The marine geospatial information coding module is used for geographic location coding: collecting and inputting historical sea surface temperature data. Spatial geographic location normalization and spatial location coding enhancement are performed to generate sea surface temperature input data. ; A multi-scale physical feature extraction module for ocean dynamic processes is used for multi-scale physical feature extraction: sea surface temperature input data. Entering the frequency domain feature modeling branch, the backbone features are extracted using spectral neural operators. Main features Entering the multi-scale spatiotemporal feature extraction branch, a multi-scale physical feature set is extracted through a spatial-temporal pyramid network. Multi-scale physical feature set By integrating through the fusion module, a unified physical feature representation is obtained. ; A multimodal feature fusion module based on marine domain knowledge is used for multimodal feature fusion of marine domain knowledge: acquiring a pre-built marine domain knowledge lexicon, and using the word embedding matrix based on a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon, and perform linear mapping on the word vectors to obtain text semantic features. Physical characteristics representation Textual semantic features By fusing features in the same feature space, joint semantic features are obtained. ; The task-prompt-driven large language model prediction module is used for task-prompt-driven large language model prediction: constructing prompt word vectors. The feature of the prompt word is obtained through embedding. cue word features With joint semantic features After concatenation, the data is input into a large language model for deep feature interpretation and prediction, outputting the predicted results of the future sea surface temperature field. .

[0043] Example 3: As another embodiment of the present invention, a computer-readable storage medium is proposed, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned large language model sea surface temperature prediction method based on multi-scale physical information fusion. The computer-readable storage medium can be a disk, optical disk, read-only memory, flash memory, solid-state drive, or other forms of non-transitory storage device, all of which can support the operation and reproduction of this method in different computing platforms and hardware environments.

[0044] During operation, the computer program in this embodiment can sequentially complete steps such as ocean geospatial feature encoding, multi-scale physical feature extraction for dynamic processes, cross-modal semantic fusion based on domain knowledge, and large language model prediction driven by task prompt words, thereby outputting the prediction results of the sea surface temperature field for future periods. This invention realizes an end-to-end automated modeling and prediction process through modular design, ensuring the portability of the method and the feasibility of engineering implementation.

[0045] Compared with the prior art, the computer-readable storage medium of the present invention enables the above method to be effectively executed under different software and hardware conditions, thus possessing good universality and scalability. At the same time, by introducing geospatial information, multi-scale physical features, and semantic cue word-driven mechanisms, the present invention can overcome the problems of insufficient capture of anomalous sea surface temperature events and insufficient utilization of domain knowledge in traditional models, and achieve higher accuracy and stronger robustness in predicting sea surface temperature.

[0046] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should be protected by the present invention.

Claims

1. A large-scale language model for sea surface temperature prediction based on multi-scale physical information fusion, characterized in that, Includes the following steps: S1, Marine Geospatial Information Coding: Collect and input historical sea surface temperature data Spatial geographic location normalization and spatial location coding enhancement are performed to generate sea surface temperature input data. ; S2. Multi-scale physical feature extraction for ocean dynamic processes: Sea surface temperature input data Entering the frequency domain feature modeling branch, the backbone features are extracted using spectral neural operators. Main features The process then moves to the multi-scale spatiotemporal feature extraction branch, where a multi-scale physical feature set is extracted using a spatial-temporal pyramid network. This set is then integrated by a fusion module to obtain a unified physical feature representation. ;include: S2.1: Sea surface temperature input data Entering the frequency domain feature modeling branch, the frequency domain feature modeling branch takes sea surface temperature input data as input. Perform Fourier transform Mapping to the frequency domain yields the frequency domain representation. Represented in the frequency domain Based on the frequency index of each frequency component, a low-frequency mask is constructed. and intermediate frequency mask The low-frequency components characterizing large-scale background changes and the mid-frequency components characterizing mesoscale spatial changes were selected to obtain the selected frequency domain features. ; Learnable complex frequency domain weights are applied to the filtered low-frequency and mid-frequency components, and then the data is restored to the physical domain using inverse Fourier transform, and compared with sea surface temperature input data. The residual mapping results are fused and then subjected to nonlinear activation to obtain frequency domain features. ; S2.2: Frequency Domain Characteristics After amplitude and phase co-reconstruction, the reconstruction results are used to form a gated map score through residual superposition. The gated map score is then compared with the sea surface temperature input data. The core features are obtained by fusing pointwise multiplication and normalization and convolution. ; S2.3: Main Characteristics Entering the multi-scale spatiotemporal feature extraction branch, this branch extracts the main features along the spatial path using a space-time pyramid network. Hierarchical downsampling is used to construct multi-resolution feature maps. Along the time path, temporal convolutional groups with different dilation rates are introduced at each scale. After layer normalization and the Gule activation function, the maps are uniformly projected to the same number of channels, resulting in a multi-scale physical feature set. ,in, The number of scales representing multiple physical features. Indicates scale index; S2.4: Multi-scale physical space collection Entering the fusion module, the fusion module combines the physical features at various scales. Global average pooling is used to generate channel descriptors, which are then passed through convolutional layers and a sigmoid activation function to obtain initial weights. Then, normalized weights are obtained through Softmax regularization. Adaptive weighted fusion is performed to obtain the physical feature representation. ; S3. Multimodal feature fusion based on marine domain knowledge: Obtain a pre-built marine domain knowledge lexicon and a word embedding matrix based on a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon, and perform linear mapping on the word vectors to obtain text semantic features. Physical characteristics representation Textual semantic features By fusing features in the same feature space, joint semantic features are obtained. ; S4. Large language model prediction based on task prompt words: Constructing prompt word vectors The feature of the prompt word is obtained through embedding. cue word features With joint semantic features After concatenation, the data is input into a large language model for deep feature interpretation and prediction, outputting the predicted results of the future sea surface temperature field. .

2. The method according to claim 1, characterized in that, The collection and input of historical sea surface temperature data Spatial geographic location normalization and spatial location coding enhancement are performed to generate sea surface temperature input data. ,include: Collect and input historical sea surface temperature data Based on the actual latitude and longitude coordinates of the observation points in the ocean region, historical sea surface temperature data were analyzed. The latitude and longitude coordinates of each spatial grid point are obtained by normalization. ; Based on latitude and longitude coordinates Geocoded vectors representing the spatial location information of ocean areas are constructed through encoding mapping operations, and these encoded vectors are concatenated to form historical sea surface temperature data. Dimensions, to obtain enhanced input ; The input is enhanced through convolutional projection. Mapping to a uniform feature width yields sea surface temperature input data. .

3. The method according to claim 1, characterized in that, Obtain a pre-built marine domain knowledge lexicon and a word embedding matrix based on a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon, and perform linear mapping on the word vectors to obtain text semantic features. Physical characteristics representation Textual semantic features By fusing features in the same feature space, joint semantic features are obtained. ,include: S3.1: Obtain a pre-built marine domain knowledge lexicon, based on the word embedding matrix of a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon. Perform dimensionality alignment and reduction on the word vectors using linear mapping to obtain the text semantic features. ; S3.2: Physical Feature Representation Textual semantic features The mapped text semantic feature representation is obtained by mapping to a unified cross-modal aligned feature space through both physical feature embedding layers and text feature embedding layers. With the mapped physical feature representation ; S3.3: Calculate the query vector Query, key vector Key, and value vector Value using adaptive weights, as shown in the following expression: ； ； ； in, The query vector Query is represented by physical feature embeddings. Mapping generation, The key vectors Key and Key represent The value vector, Value, is represented by embedding textual semantic features. Mapping generation; , , These represent the learnable linear mapping matrices used to generate queries, keys, and values, respectively. S3.4: Utilize a cross-modal attention fusion mechanism to obtain the final fused features, i.e., joint semantic features. The expression is as follows: ； in, Represents the query vector in the cross-modal attention mechanism. With key vector The feature dimensions.

4. The method according to claim 1, characterized in that, Constructing prompt word vectors The feature of the prompt word is obtained through embedding. cue word features With joint semantic features After concatenation, the data is input into a large language model for deep feature interpretation and prediction, outputting the predicted results of the future sea surface temperature field. ,include: S4.1: Construct task prompt words based on the statistical characteristics of the prediction task-related parameters and sea surface temperature input data. Then, use a word segmenter and word vector embedding to obtain prompt word vectors. Then, prompt word features are obtained through linear layer mapping. Among them, the parameters related to the prediction task include: predictor variables, target sea area, historical input time window, prediction step size, and output spatial resolution. The statistical characteristics of the sea surface temperature input data include: mean, variance, maximum value, minimum value, and trend. S4.2: Features of Prompt Words With joint semantic features The input sequence is spliced ​​together to form the fused sequence. ; S4.3: Input Sequence Input a large language model, perform sequence-level inference, and output intermediate representations. ; S4.4: Intermediate Characterization After linear mapping and activation, the original spatial resolution is restored through an upsampling layer to obtain the final temperature prediction field. .

5. The method according to claim 4, characterized in that, The final temperature prediction field is obtained by restoring the original spatial resolution through an upsampling layer. ,include: The upsampling layer uses a combination of multi-layer transposed convolution and bilinear interpolation to progressively restore spatial resolution. Each upsampling process includes linear mapping and nonlinear activation operations, ultimately resulting in a temperature prediction field. The expansion form based on layer-by-layer upsampling is as follows: ； in, Indicates the number of upsampling layers. Indicates the first The upsampling operator for the layer is either transposed convolution or bilinear interpolation. This represents the corresponding convolution kernel parameters. This represents the convolution operation. This represents the GELU activation function.

6. A large-scale linguistic model sea surface temperature prediction system based on multi-scale physical information fusion, characterized in that it is used to implement the large-scale linguistic model sea surface temperature prediction method based on multi-scale physical information fusion as described in any one of claims 1-5, wherein the large-scale linguistic model sea surface temperature prediction system comprises: Marine geospatial information coding module: Collects and inputs historical sea surface temperature data. Spatial geographic location normalization and spatial location coding enhancement are performed to generate sea surface temperature input data. ; A multi-scale physical feature extraction module for ocean dynamic processes is used for multi-scale physical feature extraction: sea surface temperature input data. Entering the frequency domain feature modeling branch, the backbone features are extracted using spectral neural operators. Main features Entering the multi-scale spatiotemporal feature extraction branch, a multi-scale physical feature set is extracted through a spatial-temporal pyramid network. Then, through integration by the fusion module, a unified physical feature representation is obtained. ; A multimodal feature fusion module based on marine domain knowledge is used for multimodal feature fusion of marine domain knowledge: acquiring a pre-built marine domain knowledge lexicon, and using the word embedding matrix based on a pre-trained large language model. Extract word vectors corresponding to each term in the knowledge lexicon, and perform linear mapping on the word vectors to obtain text semantic features. Physical characteristics representation Textual semantic features By fusing features in the same feature space, joint semantic features are obtained. ; The task-prompt-driven large language model prediction module is used for task-prompt-driven large language model prediction: constructing prompt word vectors. The feature of the prompt word is obtained through embedding. cue word features With joint semantic features After concatenation, the data is input into a large language model for deep feature interpretation and prediction, outputting the predicted results of the future sea surface temperature field. .

Images