CN-121980177-A - Typhoon wave prediction method and system based on field adaptation large language model

CN121980177ACN 121980177 ACN121980177 ACN 121980177ACN-121980177-A

Abstract

The invention discloses a typhoon wave prediction method and a typhoon wave prediction system based on a field-adaptive large language model. In order to solve the problems of misexpression, lack of physical consistency due to semantic deficiency and alignment of multivariable nonstationary typhoon prediction due to the continuous Hsig time sequence data and discrete token modal gap, the invention provides a typhoon wave prediction method and a typhoon wave prediction system based on a domain adaptation large language model. The invention realizes the alignment of numerical patches and prompt embedding by reprogramming the fusion layer based on the Time-LLM framework, injects physical constraints into a large language model for marine field Time sequence prediction, ensures output physical consistency, integrates an interpretation evaluation module and realizes efficient Hsig prediction.

Inventors

HE ZHIGUO
ZHAO YITING
HAN DONGRUI
ZHU YE
LI LI
SHEN HUI
Samuel Upon Okun

Assignees

浙江大学
浙江大学苏州工业技术研究院

Dates

Publication Date: 20260505
Application Date: 20260115

Claims (10)

1. A typhoon wave prediction method based on a domain-adaptive large language model is characterized by comprising the following steps: firstly, patch extraction, numerical embedding and discrete token processing are carried out on effective wave height Hsig time sequence data so as to bridge continuous wave data and discrete token modes; step two, generating a prompt sequence containing physical constraint and typhoon wave history information, and providing a space-time background context for the model by introducing domain knowledge to guide a attention mechanism; Thirdly, performing multi-mode alignment on the discrete token obtained in the first step and the prompt sequence generated in the second step, and generating unified fusion embedding through reprogramming layer fusion patch embedding and prompt embedding; step four, based on the fusion embedding generated in the step three, adopting sparse cross-period sampling processing to decouple the trend and the periodicity, and realizing the characteristic compression of the input sequence; step five, dimension mapping is carried out on the fusion characteristics processed in the step four, and the space expression capacity of the characteristics is enhanced by adjusting the dimension of the sampling characteristics to match with a preset prediction window; step six, extracting a prediction window part based on the fusion characteristics mapped in the step five, generating continuous Hsig predicted values and supporting long-period extrapolation; And seventhly, executing Hsig prediction and extreme event evaluation based on the predicted sequence deduced in the step six, reversely recovering the predicted sequence to the original scale, and carrying out error analysis and risk evaluation of the extreme event based on a preset threshold value.
2. The method of claim 1 wherein step one said patch extraction uses a continuous overlap division of length 16 and step size 8 and normalized by a reversible instance to zero data mean and standard deviation in units.
3. The method according to claim 1 or 2, wherein step two said hint sequence is constructed by extracting Hsig minimum, maximum, median, hysteresis, trend and quantile of input statistics and coupling shallow water equation continuity, momentum equation and historical typhoon path with peak event patterns.
4. A method according to claim 3, wherein the hysteresis value is obtained by calculating a top_k value by a fast fourier transform, the historical typhoon data being organized in the form of a knowledge graph, nodes being paths and events, edges being peaks.
5. The method of claim 1, wherein step three utilizes a reprogramming layer to map patch embedding into a multi-headed key-value space by linear query projection and calculates attention scores by a multi-headed attention mechanism to achieve cross-modal alignment of tokens and hints.
6. The method of claim 5, wherein the multi-headed attention uses 8-headed parallelism, an attention dropout rate of 0.1, and query, key, value projection dimensions are consistent and implemented in a linear layer.
7. The method of claim 1, wherein step four said sparse cross-period sampling identifies a dominant pattern of daily and half-daily periods by fourier transform and sub-samples with a 0.8 quantile threshold retaining 20% data points, thereby decoupling trend from periodicity and compressing the sequence by 70%.
8. The method of claim 1, wherein the dimension mapping in step five employs a flattening process starting from the penultimate dimension and mapping features to the prediction window dimension through a linear layer, the mapping process integrating regularization of a dropout rate of 0.1.
9. The method of claim 1, wherein step six the predicted sequence is pre-zeroed by zero padding the tag sequence and supporting 720 steps of zero sample extrapolation, the extraction process employing inverse normalization to recover the original scale and ensure prediction consistency.
10. Typhoon wave prediction system based on field adaptation large language model, characterized by comprising: the input embedding module is used for carrying out patch extraction, numerical value embedding and discrete token processing on the effective wave height Hsig time sequence data so as to bridge continuous wave data and discrete token modes; the prompt generation module is used for generating a prompt sequence containing physical constraint and typhoon wave history information, and providing a space-time background context for the model by introducing domain knowledge to guide an attention mechanism; The reprogramming layer is used for carrying out multi-mode alignment on the discrete token output by the input embedding module and the prompt sequence output by the prompt generation module, fusing patch embedding and prompt embedding, and generating unified fusion embedding; The large language model processing module is used for fusion embedding of the reprogramming layer output, adopts sparse cross-period sampling processing to decouple trend and periodicity, and realizes characteristic compression of an input sequence; The feature enhancement module is used for performing dimension mapping on the fusion features compressed by the large language model processing module, and enhancing the spatial expression capacity of the features by adjusting the dimension of the sampling features to match with a preset prediction window; the time sequence model processing module is used for extracting a prediction window part from the fusion characteristics mapped by the characteristic enhancement module, generating continuous Hsig predicted values and supporting long-period extrapolation; And the task execution module is used for executing Hsig prediction and extreme event evaluation on the prediction sequence output by the time sequence model processing module, reversely recovering the prediction sequence to the original scale, and carrying out error analysis and risk evaluation of the extreme event based on a preset threshold value.

Description

Typhoon wave prediction method and system based on field adaptation large language model Technical Field The invention belongs to the technical field of marine disaster prediction and artificial intelligence, and particularly relates to a typhoon wave prediction method and system based on a field-adaptive large language model. Background Typhoons pose a threat to coastal areas. As climate changes exacerbate typhoon frequency and intensity, accurate and timely prediction of typhoon induction Hsig (SIGNIFICANT WAVE HEIGHT, effective wave height) is critical for disaster relief, ocean engineering, and sustainable coastal management. Traditional numerical models such as the offshore wave simulation model are used for wave propagation and advanced circulation models are used for storm surge simulation, and the physical process is simulated by solving complex partial differential equations. These models provide interpretable wave dynamics and surge propagation hole findings, but face significant limitations. The inherent nonlinearity of wave mechanisms combined with multi-scale interactions of wind, ocean currents and terrain results in computational demands exceeding real-time operational requirements. For example, coastal grid high resolution simulation in Zhejiang province takes more than 58 hours to run on a high performance computing cluster. Furthermore, these models are sensitive to input uncertainties, such as wind field errors of weather forecasts, produce data distribution shifts in extreme events and reduce long-term predictability. Recent developments in artificial intelligence, particularly large language models, have revolutionized timing predictions by capturing nonlinear patterns through end-to-end learning of massive historical datasets. The large language model is pre-trained in a discrete text corpus, shows emerging reasoning capability and can bridge semantic understanding and sequence data. However, its application to continuous value marine time series data is hampered by fundamental modal gaps, where the discrete token embedding of the large language model does not match the continuous numerical nature of variables such as Hsig. This discrepancy presents three key challenges, representing that mismatch results in loss of numerical continuity in tokenization, that the semantic miss lacks a clear domain context to ensure physical consistency, and that task adaptation requires alignment of multivariate non-stationary predictions to cope with typhoon scenarios. Disclosure of Invention In order to solve the problems of misexpression, lack of physical consistency due to semantic deficiency and alignment of multivariable nonstationary typhoon prediction due to the continuous Hsig time sequence data and discrete token modal gap, the invention provides a typhoon wave prediction method and a typhoon wave prediction system based on a domain adaptation large language model. The invention discloses a typhoon wave prediction method based on a domain-adaptive large language model, which comprises the following steps of: firstly, patch extraction, numerical embedding and discrete token processing are carried out on effective wave height Hsig time sequence data so as to bridge continuous wave data and discrete token modes; step two, generating a prompt sequence containing physical constraint and typhoon wave history information, and providing a space-time background context for the model by introducing domain knowledge to guide a attention mechanism; Thirdly, performing multi-mode alignment on the discrete token obtained in the first step and the prompt sequence generated in the second step, and generating unified fusion embedding through reprogramming layer fusion patch embedding and prompt embedding; step four, based on the fusion embedding generated in the step three, adopting sparse cross-period sampling processing to decouple the trend and the periodicity, and realizing the characteristic compression of the input sequence; step five, dimension mapping is carried out on the fusion characteristics processed in the step four, and the space expression capacity of the characteristics is enhanced by adjusting the dimension of the sampling characteristics to match with a preset prediction window; step six, extracting a prediction window part based on the fusion characteristics mapped in the step five, generating continuous Hsig predicted values and supporting long-period extrapolation; And seventhly, executing Hsig prediction and extreme event evaluation based on the predicted sequence deduced in the step six, reversely recovering the predicted sequence to the original scale, and carrying out error analysis and risk evaluation of the extreme event based on a preset threshold value. The invention relates to a typhoon wave prediction system based on a field-adaptive large language model, which comprises: the input embedding module is used for carrying out patch extraction, numerical value embedding and discrete