Marine heat wave prediction method

Abstract

The application relates to a marine heat wave prediction method, which comprises the following steps: obtaining physical variable historical time series data of each region of a target sea area, generating a physical variable prediction sequence containing physical variable prediction values of future time points through a composite model obtained through training, and mining physical variable relationship data. Subsequently, the physical variable prediction sequence, the physical variable relationship data and a preset task instruction are input into a heat wave analysis module, the input data are accurately analyzed through the heat wave analysis module, and conclusion deviation caused by non-uniform standards during manual interpretation is avoided. The automatic analysis reduces the direct participation of manual work, improves the analysis efficiency and the standardization degree. Finally, marine heat wave information output by the heat wave analysis module is obtained, and a prediction result is obtained.

Description

Technical Field

[0001] This application relates to the field of marine heat wave technology, and in particular to a method for predicting marine heat waves. Background Technology

[0002] Under the general trend of global warming, marine heat waves, as an extreme marine meteorological phenomenon, are occurring more frequently and with increasing severity, causing multifaceted impacts on marine ecosystems and human society. Marine heat waves refer to a short-term sea surface temperature that is significantly higher than historical averages for the same period, lasting from several days to several months, and spatially covering areas extending from nearshore regions to ocean basins. This abnormal warming phenomenon can trigger a series of serious consequences, such as large-scale coral bleaching, disruption of marine biological communities, and consequently, ecological and economic losses including reduced fisheries production; it can also influence air-sea interactions, inducing extreme weather events such as typhoons and torrential rains, seriously threatening the safety of production and daily life in coastal areas.

[0003] Currently, prediction technologies for marine heat waves have certain limitations. Existing prediction methods lack accuracy and reliability. Some schemes rely heavily on human experience for interpretation, which is not only time-consuming and labor-intensive but also prone to bias due to the lack of unified interpretation standards, making it difficult to support the standardized application of marine heat wave prediction. Current marine heat wave prediction technologies are insufficient to meet the need for accurate prediction of marine heat waves, enabling proactive countermeasures to reduce their negative impacts on marine ecosystems and human society. Therefore, developing a marine heat wave prediction method that improves accuracy and reliability while reducing reliance on human intervention is of significant practical importance. Summary of the Invention

[0004] Therefore, the purpose of this application is to provide a method for predicting marine heat waves, so as to achieve accurate prediction of marine heat waves and take countermeasures in advance to reduce their negative impact on marine ecosystems and human society.

[0005] The marine heat wave prediction method described in this application includes the following steps:

[0006] Acquire historical time-series data of physical variables for various areas of the target sea area; the historical time-series data of physical variables includes observed values ​​of physical variables at several historical time points;

[0007] Historical time-series data of physical variables in each region are input into a preset composite model to obtain predicted sequences of physical variables and relational data of physical variables in each region; wherein, the composite model is used to generate predicted sequences of physical variables and relational data of physical variables based on the historical time-series data of physical variables; the predicted sequences of physical variables include predicted values ​​of physical variables at several future time points;

[0008] The predicted sequences of physical variables for each region, the relationship data of the physical variables, and the preset task instructions are input into the preset heat wave analysis module; wherein, the preset task instructions are used to prompt the heat wave analysis module to analyze and obtain marine heat wave information based on the predicted sequences of physical variables for each region and the relationship data of the physical variables.

[0009] Obtain the marine heat wave information output by the heat wave analysis module; based on the marine heat wave information, obtain the marine heat wave prediction result for the target sea area.

[0010] This application embodiment acquires historical time-series data of physical variables from various areas of the target sea area and inputs it into a preset composite model. The model delves into the intrinsic relationships between physical variables, generating a physical variable prediction sequence containing predicted values ​​for future time points and outputting physical variable relationship data. By fully utilizing the rich information in historical data, compared to traditional methods relying on limited human experience, it significantly improves the comprehensiveness and accuracy of data processing, making the prediction results more reliable and precise. Subsequently, the physical variable prediction sequence, physical variable relationship data, and preset task instructions are input into the heatwave analysis module. The preset task instructions provide a clear analytical direction for the heatwave analysis module, enabling it to accurately analyze the input data and avoid conclusion biases caused by inconsistent standards in manual interpretation. This automated analysis method reduces direct human intervention, improving analysis efficiency and standardization. Finally, the prediction results are obtained by acquiring the marine heatwave information output by the heatwave analysis module. The embodiments of this application enable accurate prediction of marine heat waves, providing a strong basis for relevant departments to take countermeasures in advance, thereby effectively reducing the negative impact of marine heat waves on marine ecosystems and human society, and having significant advantages in ensuring marine ecological balance and the safety of production and life in coastal areas.

[0011] To better understand and implement this application, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the marine heat wave prediction method according to an embodiment of this application;

[0013] Figure 2 This is a schematic diagram illustrating the steps of training a composite model in an embodiment of this application;

[0014] Figure 3 This is a schematic diagram illustrating the steps of converting historical observation sequences into high-dimensional feature tensors in an embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings. Wherein, when the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements.

[0016] It should be understood that the embodiments described below do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0017] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application are also intended to include the plural forms unless the context clearly indicates otherwise. Furthermore, in the description of this application, unless otherwise stated, “a plurality” means two or more. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items, for example, A and / or B, which can represent: A alone, A and B together, and B alone; the character “ / ” generally indicates that the preceding and following objects are in an “or” relationship.

[0018] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, this information should not be limited to these terms, and these terms are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances. Depending on the context, the word "if" as used in this application can be interpreted as "when," "when," or "in response to determination."

[0019] Under the general trend of global warming, marine heat waves, as an extreme marine meteorological phenomenon, are occurring more frequently and with increasing severity, causing multifaceted impacts on marine ecosystems and human society. Marine heat waves refer to a short-term sea surface temperature that is significantly higher than historical averages for the same period, lasting from several days to several months, and spatially covering areas extending from nearshore regions to ocean basins. This abnormal warming phenomenon can trigger a series of serious consequences, such as large-scale coral bleaching, disruption of marine biological communities, and consequently, ecological and economic losses including reduced fisheries production; it can also influence air-sea interactions, inducing extreme weather events such as typhoons and torrential rains, seriously threatening the safety of production and daily life in coastal areas.

[0020] Currently, prediction technologies for marine heat waves have certain limitations. Existing prediction methods lack accuracy and reliability. Some schemes rely heavily on human experience for interpretation, which is not only time-consuming and labor-intensive but also prone to bias due to the lack of unified interpretation standards, making it difficult to support the standardized application of marine heat wave prediction. Current marine heat wave prediction technologies are insufficient to meet the need for accurate prediction of marine heat waves, enabling proactive countermeasures to reduce their negative impacts on marine ecosystems and human society. Therefore, developing a marine heat wave prediction method that improves accuracy and reliability while reducing reliance on human intervention is of significant practical importance.

[0021] This application provides a method for predicting marine heat waves, enabling accurate prediction of marine heat waves so that countermeasures can be taken in advance to reduce their negative impact on marine ecosystems and human society.

[0022] Please refer to Figure 1 The marine heat wave prediction method described in this application includes the following steps:

[0023] S101: Obtain historical time-series data of physical variables for each area of ​​the target sea area; the historical time-series data of physical variables includes observed values ​​of physical variables at several historical time points;

[0024] S102: Input the historical time series data of physical variables of each region into a preset composite model to obtain the predicted sequence of physical variables and the relationship data of physical variables of each region; wherein, the composite model is used to generate the predicted sequence of physical variables and the relationship data of physical variables based on the historical time series data of physical variables; the predicted sequence of physical variables includes predicted values ​​of physical variables at several future time points;

[0025] S103: Input the predicted sequence of physical variables for each region, the relationship data of the physical variables, and the preset task instructions into the preset heat wave analysis module; wherein, the preset task instructions are used to prompt the heat wave analysis module to analyze and obtain marine heat wave information based on the predicted sequence of physical variables for each region and the relationship data of the physical variables.

[0026] S104: Obtain the marine heat wave information output by the heat wave analysis module; based on the marine heat wave information, obtain the marine heat wave prediction result for the target sea area.

[0027] This application embodiment acquires historical time-series data of physical variables from various areas of the target sea area and inputs it into a preset composite model. The model delves into the intrinsic relationships between physical variables, generating a physical variable prediction sequence containing predicted values ​​for future time points and outputting physical variable relationship data. By fully utilizing the rich information in historical data, compared to traditional methods relying on limited human experience, it significantly improves the comprehensiveness and accuracy of data processing, making the prediction results more reliable and precise. Subsequently, the physical variable prediction sequence, physical variable relationship data, and preset task instructions are input into the heatwave analysis module. The preset task instructions provide a clear analytical direction for the heatwave analysis module, enabling it to accurately analyze the input data and avoid conclusion biases caused by inconsistent standards in manual interpretation. This automated analysis method reduces direct human intervention, improving analysis efficiency and standardization. Finally, the prediction results are obtained by acquiring the marine heatwave information output by the heatwave analysis module. The embodiments of this application enable accurate prediction of marine heat waves, providing a strong basis for relevant departments to take countermeasures in advance, thereby effectively reducing the negative impact of marine heat waves on marine ecosystems and human society, and having significant advantages in ensuring marine ecological balance and the safety of production and life in coastal areas.

[0028] The marine heat wave prediction method described in this application uses a computer as the execution subject, and the following provides a detailed description of each step.

[0029] For step S101, obtain historical time-series data of physical variables for each area of ​​the target sea area; the historical time-series data of physical variables includes observed values ​​of physical variables at several historical time points.

[0030] The target sea area refers to a specific marine region that needs to be studied or monitored for marine heat waves. The scope and boundaries of this area can be defined according to actual needs and research objectives; for example, it could be the sea area near a coastal city or a sea area with specific ecological characteristics. Within the target sea area, multiple sub-regions can be further subdivided. These sub-regions can be divided based on various factors such as geographical location and marine environmental characteristics. For example, they can be divided into nearshore, midshore, and offshore areas according to their distance from the coastline; or into shallow and deep-sea areas according to water depth.

[0031] Historical time-series data of physical variables is a collection of observed values ​​of physical variables in various areas of a target sea area at different historical points in time.

[0032] Physical variables refer to various parameters that reflect the state of the marine environment, such as seawater temperature, which reflects the distribution of ocean heat and has an important impact on the survival of marine life and ocean circulation.

[0033] In this embodiment, appropriate physical variables can be selected as needed. These variables include, but are not limited to, sea surface temperature, 2-meter air temperature, total precipitation, significant wave height, sea level pressure, wind speed and direction, relative humidity, atmospheric precipitable water, evaporation, sea surface salinity, ocean currents (velocity and direction), ocean heat content, tidal level / height, and mixed layer depth. Sea surface temperature refers to the temperature of seawater within a certain depth range above the ocean surface. It is a direct reflection of ocean heat and has a crucial impact on the survival and distribution of marine life and ocean circulation. For example, different sea surface temperature zones attract different types of marine life, and changes in sea surface temperature can trigger adjustments in ocean circulation. 2-meter air temperature is the air temperature at a height of 2 meters above the sea surface, reflecting the thermal conditions of the atmosphere near the ocean surface. 2-meter air temperature and sea surface temperature interact and jointly influence the ocean-atmosphere interaction process, thereby affecting the formation and change of weather and climate. For example, an increase in sea surface temperature may cause an increase in 2-meter air temperature, triggering localized weather changes. Total precipitation refers to the total amount of precipitation that falls on the ocean surface within a certain time and space range. Total precipitation affects the ocean's salinity balance and water cycle processes. Excessive precipitation can dilute seawater, lower salinity, and impact the habitat of marine life. Significant wave height refers to the average of one-third of the maximum wave heights over a period of time; it reflects the energy and intensity of the waves. Significant wave height has a significant impact on ocean navigation, offshore operations, and coastal erosion. Higher significant wave heights can make navigation difficult and increase the risks of offshore operations. Sea level pressure refers to the atmospheric pressure at sea level and is a crucial element in meteorology. The distribution and changes in atmospheric pressure are closely related to the formation, movement, and development of weather systems. For example, low-pressure areas are often accompanied by cloud cover and precipitation, while high-pressure areas tend to have clear weather. Wind speed represents the distance air travels per unit time, usually measured in meters per second (m / s) or kilometers per hour (km / h). Wind speed affects the intensity of meteorological phenomena; for example, strong winds can lead to storms, hurricanes, and other severe weather. Wind direction refers to the direction from which the wind blows, generally expressed using 16 azimuths or angles. Wind direction significantly influences the movement of weather systems and the distribution of meteorological elements; for example, monsoons bring specific seasonal climate characteristics. Relative humidity refers to the percentage of water vapor pressure in the air compared to the saturation water vapor pressure at the same temperature, reflecting the degree to which the air is close to saturation. Relative humidity affects human comfort, plant transpiration, and some industrial processes. High relative humidity makes people feel stuffy and hot; low relative humidity makes the air dry, easily leading to skin and respiratory problems. Atmospheric precipitable water refers to the total amount of water vapor contained in a vertical column of air per unit area from the ground to the top of the atmosphere; if all of it condensed and fell to the ground, it would be the amount of precipitation that would be obtained. It is an important indicator for assessing a region's water resource potential and is of great significance for precipitation forecasting and water resource management.Evaporation refers to the total amount of water that evaporates from the Earth's surface (such as water, soil, and plant surfaces) into the atmosphere within a certain period. Evaporation is affected by various factors, including temperature, humidity, wind speed, and solar radiation. In agriculture, understanding evaporation helps in the rational planning of irrigation; in water resource management, evaporation is an important parameter for calculating the water resource balance. Sea surface salinity refers to the salt concentration in the surface layer of seawater, usually expressed as parts per thousand (ppm). Sea surface salinity is affected by factors such as precipitation, evaporation, river runoff, and glacial melting. Changes in salinity affect the density of seawater, thus influencing ocean currents. For example, high-salinity seawater has a higher density and sinks, driving vertical movement in the ocean. Ocean currents are large-scale movements of seawater in the ocean with relatively stable velocity and direction. Velocity represents the distance a current travels per unit time, usually measured in centimeters per second (cm / s) or knots (1 knot = 1.852 km / h). The magnitude of ocean current velocity affects the transport of marine materials and the transfer of energy. Direction refers to the direction of ocean current flow, expressed as azimuth. The direction and speed of ocean currents have a significant impact on global climate, marine ecosystems, and navigation. For example, warm currents make coastal areas warm and humid, while cold currents make them cold and dry. Ocean heat content refers to the total amount of heat energy stored in the ocean, which is related to the temperature, volume, and specific heat capacity of seawater. Changes in ocean heat content play a crucial role in regulating global climate. The ocean can absorb and store large amounts of heat; when ocean heat content changes, it affects atmospheric circulation and climate patterns. For example, the El Niño phenomenon is closely related to abnormal changes in ocean heat content in the Pacific Ocean. Tide level refers to the vertical position of seawater above the ocean surface, usually with a reference level (such as mean sea level). Tide height refers to the height of the tide level relative to a reference level. Changes in tide level and height are caused by the gravitational pull of the moon and sun, as well as other factors (such as meteorological conditions and topography). Tidal phenomena have a significant impact on port navigation, coastal engineering, and fishing activities in coastal areas. For example, ships can easily enter and leave ports at high tide, while low tide may expose some shallows. The mixing layer depth refers to the depth of a relatively homogeneous layer of seawater in the ocean due to turbulent mixing, characterized by similar temperature and salinity. The mixing layer depth is influenced by factors such as wind stress, solar radiation, and ocean heat flux. The magnitude of the mixing layer depth affects the exchange of heat, momentum, and matter between the ocean and the atmosphere, playing a crucial role in the upper ocean ecosystem and climate. For example, a deeper mixing layer can promote the exchange between the upper and lower ocean water layers, influencing nutrient distribution and phytoplankton growth.

[0034] Historical time-series data emphasizes that these observations are arranged in chronological order, covering multiple historical time points. Through long-term data accumulation, it can reflect the changing patterns of the marine environment over time.

[0035] For step S102, the historical time series data of physical variables in each region are input into a preset composite model to obtain the predicted sequence of physical variables and the relationship data of physical variables in each region; wherein, the composite model is used to generate the predicted sequence of physical variables and the relationship data of physical variables based on the historical time series data of physical variables; the predicted sequence of physical variables includes predicted values ​​of physical variables at several future time points.

[0036] The pre-set composite model is a mathematical model that has been trained and configured beforehand. Through learning and analyzing a large amount of historical data, this model establishes a mapping relationship between historical time-series data of physical variables and their future states. When historical time-series data of physical variables from various regions are input into the composite model, the model uses its internal algorithms to process and predict the data, generating a sequence containing predicted values ​​for physical variables at several future time points. For example, it can predict the daily values ​​of physical variables such as sea surface temperature, air temperature at 2 meters, total precipitation, and significant wave height for the next week. These predicted values ​​can help us understand the changing trends of the marine environment in advance.

[0037] The composite model, while generating predicted sequences, also uncovers data on the relationships between different physical variables. This data reveals the intrinsic connections between these variables; for example, how changes in sea surface temperature might affect changes in 2-meter air temperature, total precipitation, and significant wave height, or the correlation between sea surface temperature and total precipitation. Understanding these relationships allows for a deeper understanding of the dynamic mechanisms of marine environmental change.

[0038] In one embodiment, the composite model includes an input adapter, a pre-trained large-scale model backbone network, a prediction module, and a relationship mining module; the pre-trained large-scale model backbone network includes a trainable low-rank matrix; the input adapter is connected to the pre-trained large-scale model backbone network, and the pre-trained large-scale model backbone network is connected to the prediction module and the relationship mining module;

[0039] The input adapter is used to decompose and embed the historical observation sequence data to obtain high-dimensional features, and then transmit the high-dimensional features to the pre-trained large model backbone network.

[0040] The pre-trained large model backbone network is used to receive the high-dimensional features and process them to obtain the high-dimensional hidden state, and transmit the high-dimensional hidden state to the prediction module and the relationship mining module;

[0041] The prediction module is used to convert the high-dimensional hidden state into a physical variable prediction sequence through a linear mapping layer;

[0042] The relationship mining module is used to convert the high-dimensional hidden state into physical variable relationship data through an attention mechanism.

[0043] This embodiment achieves a comprehensive improvement in the accuracy, efficiency, and interpretability of marine heatwave prediction through innovative composite model architecture and optimized training process. The high-dimensional feature processing capabilities of the input adapter, combined with the low-rank adaptation technique of the pre-trained large model backbone network, significantly reduce the computational load and energy consumption of marine monitoring equipment. The collaborative working mechanism between the prediction module and the relationship mining module enables the model to not only output high-precision physical variable prediction sequences but also reveal hidden correlations between variables.

[0044] In one embodiment, the relationship mining module is used to extract the feature vector corresponding to the prediction start time point from the high-dimensional hidden state as a query vector; to use the feature vectors corresponding to all historical time points in the high-dimensional hidden state as key vectors; to perform transpose processing on the last two dimensions of the key vectors to obtain transposed key vectors; to calculate the dot product similarity between the query vector and the transposed key vectors to obtain a correlation score; to perform normalization processing on the correlation score in the historical time dimension to obtain the influence score corresponding to each historical time point; and to obtain physical variable relationship data based on the influence score.

[0045] In the attention mechanism, the query vector represents the specific features at the current time point that need to be focused on, while the key vector represents the feature library of historical time series. The similarity between the two is quantified through dot product operations, enabling the accurate extraction of information strongly correlated with the prediction target from historical data.

[0046] The dot product operation directly reflects the directional matching degree between vectors. Normalization processing (such as Softmax) transforms the original similarity into an influence score in the 0-1 interval, clarifying the contribution weight of each historical moment to the prediction starting point and avoiding analytical bias caused by numerical differences.

[0047] This embodiment utilizes a query-key vector mechanism combined with dot product similarity calculation to enable the model to accurately pinpoint key time windows and anomalous combinations of environmental variables preceding heat waves. For example, in a heat wave event, the module successfully identified a strong correlation between persistently high sea surface temperatures and anomalous fluctuations in significant wave height over the 30 days prior to the heat wave's onset, providing a scientific basis for issuing heat wave warnings 20 days in advance. Normalized influence scores generate physical variable relationship data, revealing not only hidden coupling patterns between variables but also quantifying the contribution weight of each factor to the heat wave. This interpretable output allows meteorological departments to develop targeted protection strategies, such as adjusting irrigation plans in coastal areas based on the strength of the sea surface temperature-precipitation correlation.

[0048] Please refer to Figure 2 In one embodiment, before step S102, which involves inputting the historical time-series data of the physical variables of each region into a preset composite model to obtain the predicted sequences of the physical variables and the relationship data of the physical variables of each region, the method includes:

[0049] Step S1021: Obtain historical observation data of the target sea area; the historical observation data includes time series data of physical variables in various areas of the target sea area; the time series data of physical variables includes observed values ​​of physical variables sorted by time series.

[0050] Historical observation data records various environmental parameters of a target sea area over a past period. Time-series data of physical variables is a crucial component of this historical observation data, containing chronologically ordered observations of physical variables such as sea surface temperature, air temperature at 2 meters, total precipitation, and significant wave height. Collecting these long-term series data allows us to understand the patterns of change in physical variables over time, providing a foundation for subsequent model training and prediction. For example, analyzing multi-year sea surface temperature time-series data reveals trends in sea surface temperature across different seasons and years.

[0051] Step S1022: Based on the historical observation data, several historical observation sequences and a target sequence corresponding to each historical observation sequence are obtained; wherein, the historical observation sequence includes several consecutive first historical time points of physical variable observation values ​​in each region; the target sequence includes several consecutive second historical time points of physical variable observation values ​​in each region, and the second historical time point is a consecutive time point following the first historical time point.

[0052] Among them, the historical observation sequence is a sequence of physical variable observation values ​​over a continuous period of time selected from historical observation data. These physical variable observation values ​​at the first historical time point reflect the environmental state of the target sea area during a certain period. For example, a historical observation sequence can be composed of daily observation values ​​such as sea surface temperature and 2-meter air temperature over a continuous week.

[0053] The target sequence is a sequence of physical variable observations occurring at consecutive time points following the first historical observation point, corresponding to the historical observation sequence. It represents the environmental changes in the target sea area over a period of time after the historical observation sequence. For example, if the historical observation sequence contains data for one week, the target sequence might contain data for the following week. This decomposition method establishes a link between past environmental conditions and future environmental changes, providing data support for models to learn this temporal dependency.

[0054] Step S1023: Generate a training dataset based on each of the historical observation sequences and the target sequence corresponding to each of the historical observation sequences; wherein the training dataset includes several training samples, each training sample corresponds to a set of historical observation sequences, and the training label of each training sample is the target sequence corresponding to the set of historical observation sequences.

[0055] The training dataset is the collection of data used to train the composite model. Each training sample consists of a set of historical observation sequences containing information on physical variables in the target sea area over a past period. The training labels are the target sequences corresponding to that set of historical observation sequences, i.e., the observed values ​​of physical variables over a future period. In this way, the model can learn the ability to predict future environmental changes from historical environmental conditions. For example, a training sample might contain historical observation sequences of sea surface temperature, 2-meter air temperature, etc., from the past week, and its training labels would be the corresponding observed values ​​of physical variables for the following week.

[0056] Step S1024: Input the training dataset into the pre-trained composite model for training to obtain the trained composite model.

[0057] Among them, the pre-trained composite model is a model framework that already has a certain initial structure and parameters. It has a certain ability to process and analyze data, but it still needs to be optimized and adjusted using specific training data.

[0058] This embodiment first acquires historical observation data of the target sea area, including time-series data of key physical variables such as sea surface temperature, 2-meter air temperature, total precipitation, and significant wave height. This rich and detailed data provides a comprehensive information foundation for model training, enabling the model to learn the complex changes in physical variables of the target sea area over time. Next, the historical observation data is split to obtain historical observation sequences and corresponding target sequences, generating a training dataset. This processing method cleverly establishes a connection between past and future environmental states, allowing the model to learn the skill of predicting future changes from historical data. By inputting the training dataset into a pre-trained composite model, the model continuously optimizes its parameters, improving the accuracy of predicting future physical variables.

[0059] In one embodiment, step S1021, which involves acquiring historical observation data of the target sea area, includes:

[0060] Step S201: Obtain the offshore distance and seabed topographic gradient of each area in the target sea area.

[0061] Among them, offshore distance refers to the horizontal distance of a certain area in the target sea area from the coastline. Different offshore distances may result in different marine environmental characteristics. For example, nearshore areas may be more affected by land, while offshore areas have more pronounced marine characteristics.

[0062] The seabed topographic gradient indicates the degree of change in seabed topography in the horizontal direction. A large seabed topographic gradient means that the seabed topography is undulating, such as the presence of trenches and seamounts; a small gradient means that the seabed topography is relatively flat.

[0063] Data on offshore distances and seabed topographic gradients for various areas within the target sea area are obtained through multiple methods. For example, satellite remote sensing technology can be used to measure the distance from the ocean surface to the coastline, thus obtaining the offshore distance. For seabed topographic gradients, marine exploration equipment such as multibeam echo sounders can be used to measure water depths at different locations on the seabed and calculate the changes in seabed topography, thereby obtaining the seabed topographic gradient. This data forms the basis for subsequent determination of area type and geographic resolution.

[0064] Step S202: Determine the region type and corresponding geographic resolution of each region based on the offshore distance and seabed topographic gradient of each region; wherein, the region type includes a first type and a second type; the offshore distance of the first type region is less than a preset distance or the fluctuation degree of the seabed topographic gradient is greater than a preset fluctuation degree; the first type region corresponds to a first geographic resolution, and the second type region corresponds to a second geographic resolution; the first geographic resolution is greater than the second geographic resolution.

[0065] In geographic data, geographic resolution is an indicator used to describe the size of the geographic area represented by the data. The higher the geographic resolution, the smaller the data area represented and the richer the geographic details reflected; the lower the geographic resolution, the larger the data area represented and the more macroscopic the geographic information reflected.

[0066] The classification criteria are based on distance from the shore and seafloor topographic gradient. When the distance from the shore to a given area is less than a predetermined distance, it indicates that the area is close to the coastline and may be significantly influenced by terrestrial factors (such as river runoff and human activities). When the fluctuation of the seafloor topographic gradient is greater than a predetermined fluctuation, it indicates that the seafloor topography in the area is complex and may involve unique ocean dynamic processes. Areas meeting either of these two conditions are classified as Type I areas, while the remaining areas are classified as Type II areas.

[0067] The first type of region corresponds to the first geographic resolution, and the second type of region corresponds to the second geographic resolution, with the first geographic resolution being greater than the second. This is because the environmental changes in the first type of region are relatively more complex and drastic, requiring a higher geographic resolution to accurately describe its physical variable characteristics; while the environment in the second type of region is relatively stable and simple, and a lower geographic resolution is sufficient for the research needs. For example, the first type of region may use a smaller grid size to represent the geographic area in order to capture more detailed changes; while the second type of region uses a larger grid size.

[0068] Step S203: Obtain the observation sequence data of the target sea area at the original resolution; the observation sequence data includes the physical variable values ​​of each area of ​​the target sea area at several historical time points.

[0069] Raw resolution observation sequence data refers to the observed values ​​of physical variables (sea surface temperature, 2-meter air temperature, total precipitation, and significant wave height) in a target sea area over a period of time, obtained through various oceanographic observation methods (such as buoy observation and satellite remote sensing). This data possesses raw geographic resolution and serves as the foundational data source for subsequent data sampling and processing. For example, buoys can continuously measure physical variables such as sea surface temperature at specific locations, while satellite remote sensing can rapidly acquire relevant information about the ocean surface over large areas.

[0070] Step S204: For the observation sequence data at the original resolution, sample the time series data of physical variables of the first type of region according to the first geographic resolution, and sample the time series data of physical variables of the second type of region according to the second geographic resolution; wherein, the first geographic resolution is greater than the second geographic resolution.

[0071] Based on the previously determined region type and corresponding geographic resolution, the observation sequence data at the original resolution are sampled. For the first type of region, due to its complex environment, a more refined data description is required, so sampling is performed according to the first geographic resolution, selecting time-series data of physical variables that meet the requirements of this resolution from the original data; for the second type of region, sampling is performed according to the second geographic resolution. This sampling method allows for the reasonable processing of data from different regions while ensuring data quality, making the data more consistent with the environmental characteristics and research needs of each region. For example, data points are selected at smaller spatial intervals in the first type of region, and at larger spatial intervals in the second type of region.

[0072] Step S205: Based on the time series data of physical variables of the first type of region and the time series data of physical variables of the second type of region obtained by sampling, the historical observation data of the target sea area is obtained.

[0073] The time-series data of physical variables from the first and second type regions, obtained through sampling processing, are integrated to form complete historical observation data for the target sea area. This data contains physical variable information for each region of the target sea area at different historical time points, and the geographic resolution of the data matches the environmental characteristics of each region, providing an accurate and suitable data foundation for subsequent model training and marine heatwave prediction.

[0074] This embodiment first obtains the offshore distance and seabed topographic gradient of each region in the target sea area. These two key parameters accurately reflect the marine environmental characteristics of different regions. Based on the offshore distance and seabed topographic gradient, the region type and corresponding geographic resolution are determined. This differentiated processing fully considers the environmental differences between different regions. Regions of the first type, due to their proximity to the coastline or complex seabed topography, are assigned a higher geographic resolution to more accurately capture subtle changes in their physical variables. Regions of the second type have relatively stable environments and are assigned a lower geographic resolution, ensuring data validity while reducing data processing workload. This reasonable geographic resolution setting ensures that the data accurately reflects the actual situation of each region while improving data processing efficiency. After obtaining the observation sequence data at the original resolution, sampling is performed according to the corresponding geographic resolution for different region types, further optimizing data quality. The sampled data better matches the environmental characteristics of each region, avoiding errors caused by mismatched data resolution. The finally integrated historical observation data of the target sea area contains rich and accurate temporal information on physical variables, providing a solid foundation for subsequent training of the composite model.

[0075] In one embodiment, step S204, which involves sampling time-series physical variable data of a first type of region according to a first geographic resolution and sampling time-series physical variable data of a second type of region according to a second geographic resolution, further includes:

[0076] Step S2041: Obtain meteorological information of the target sea area at several historical time points.

[0077] Meteorological information encompasses a variety of factors that affect the marine environment, such as wind speed, wind direction, air pressure, and precipitation type (e.g., whether it is severe convective precipitation). These meteorological factors can cause varying degrees of disturbance to the marine environment, thereby affecting physical variables such as sea surface temperature, air temperature at 2 meters, total precipitation, and significant wave height.

[0078] This step can obtain meteorological information of the target sea area at several historical points in time through various means. Meteorological satellites can monitor the weather conditions over the ocean in real time over a large area, providing information such as wind speed, wind direction, and cloud distribution; in addition to measuring physical variables, some ocean buoys are also equipped with meteorological sensors to acquire local meteorological data; in addition, coastal meteorological observation stations can also provide relevant meteorological information. Through data integration and processing, a relatively comprehensive historical meteorological information of the target sea area can be obtained.

[0079] Step S2042: Determine the time resolution corresponding to each historical time point based on the meteorological information at each historical time point; wherein, the first type of meteorological information corresponds to the first time resolution, and the second type of meteorological information corresponds to the second time resolution; the first type of meteorological information has a greater disturbance effect on the sea area than the second type of meteorological information; the first time resolution is greater than the second time resolution.

[0080] The first category of meteorological information refers to combinations of meteorological factors or extreme weather conditions that significantly disturb the sea area, such as strong winds (high wind speeds and long durations), low-pressure systems (potentially causing storm surges), and severe convective precipitation (large amounts of rainfall in a short period, possibly accompanied by thunderstorms and other severe weather phenomena). These meteorological conditions cause rapid and drastic changes in the physical variables of the sea area. The second category of meteorological information refers to relatively stable meteorological conditions that cause less disturbance to the sea area, such as light winds, gentle pressure changes, and light to moderate rain.

[0081] Because the first type of meteorological information causes drastic changes in marine physical variables, a higher temporal resolution is needed to accurately capture these rapid changes; that is, a larger first temporal resolution means multiple data samplings within a short period of time. For example, in strong winds, sea surface temperature may change significantly within minutes, thus requiring data collection every few minutes. In contrast, under the second type of meteorological information, marine physical variables change relatively slowly, and a lower temporal resolution (a smaller second temporal resolution) is sufficient, such as collecting data every hour.

[0082] Step S2043: Based on the first geographic resolution and the time resolution, sample the time series data of physical variables of the first type of region; based on the second geographic resolution and the time resolution, sample the time series data of physical variables of the second type of region.

[0083] For the first type of region, physical variable time-series data sampling is performed using a combination of first geographic resolution and first temporal resolution. First geographic resolution ensures a detailed spatial description of the region's characteristics, while first temporal resolution ensures the capture of rapid changes in physical variables caused by first-type meteorological information. For example, in a first-type region near the coast with complex seabed topography, during strong winds, data on sea surface temperature, significant wave height, and other physical variables are collected every few minutes in small areas defined by the first geographic resolution. For the second type of region, sampling is performed using second geographic resolution and second temporal resolution. Second geographic resolution adapts to the relatively stable environmental characteristics of the region, while second temporal resolution matches the slow changes in physical variables under second-type meteorological information. For example, in a second-type region far from the coast with flat seabed topography, under conditions of light winds and stable air pressure, relevant physical variable data are collected hourly in large areas defined by the second geographic resolution.

[0084] This embodiment, by combining meteorological information to determine the temporal resolution and perform sampling, can more accurately reflect the actual changes of physical variables in different regions under different meteorological conditions. Under conditions of significant meteorological disturbance, high temporal resolution sampling can capture rapid fluctuations in physical variables; when the weather is relatively stable, low temporal resolution sampling avoids data redundancy while ensuring data representativeness. This makes the sampled historical observation data more accurately reflect the actual situation of the target sea area.

[0085] In one embodiment, before step S1023, which involves generating a training dataset based on each of the historical observation sequences and the target sequence corresponding to each historical observation sequence, the following steps are included:

[0086] Step S301: Convert each of the historical observation sequences into a historical observation sequence in the form of a high-dimensional feature tensor;

[0087] Before step S102, which involves inputting the historical time-series data of physical variables from each region into a preset composite model to obtain the predicted sequences of physical variables and the relationship data of physical variables for each region, the following steps are included:

[0088] Step S302: Convert the historical time series data of the physical variables into historical time series data of the physical variables in the form of high-dimensional feature tensors.

[0089] High-dimensional feature tensors refer to data structures that transform raw sequence data into multi-dimensional arrays through feature engineering or deep learning encoders. These tensors contain multi-dimensional feature information such as time, space, and physical variable types, enabling a more comprehensive depiction of the spatiotemporal evolution of the marine environment. For example, time-series data of physical variables such as sea surface temperature and 2-meter air temperature can be mapped onto a three-dimensional tensor, where one dimension represents the time step, one dimension represents the spatial grid location, and one dimension represents the different physical variable types.

[0090] This embodiment significantly enhances the feature extraction capability and pattern recognition accuracy of the marine heatwave prediction model by converting historical observation sequences and historical time-series data of physical variables into high-dimensional feature tensors. The high-dimensional tensor representation comprehensively captures the complex spatiotemporal relationships of physical variables such as sea surface temperature, 2-meter air temperature, total precipitation, and significant wave height, enabling the composite model to learn deeper patterns of marine environmental evolution during training. For example, by simultaneously representing temporal evolution, spatial distribution, and variable type information through a three-dimensional tensor structure, the model can accurately identify multivariate anomalous coupling signals preceding marine heatwaves. During the prediction phase, the high-dimensional tensor input allows the composite model to output more interpretable physical variable relationship data, providing a more reliable decision-making basis for the heatwave analysis module.

[0091] Please refer to Figure 3 In one embodiment, step S301, which involves converting each of the historical observation sequences into a high-dimensional feature tensor form, includes:

[0092] Step S3011: Perform a moving average process on the historical observation sequence to obtain the trend component; perform frequency domain transformation and autocorrelation analysis on the historical observation sequence to obtain the seasonal component; subtract the trend component and the seasonal component from the historical observation sequence to obtain the residual component.

[0093] The moving average process calculates a moving average of the data using a sliding window. This smooths the data, extracts long-term trend components, and reduces short-term fluctuations. For example, a 30-day moving average can reflect the monthly variation trend of sea surface temperature.

[0094] Frequency domain transformation (such as Fourier transform) converts time-series data from the time domain to the frequency domain to identify periodic features; autocorrelation analysis detects the time interval of periodic patterns by calculating the correlation between a sequence and its lagged versions. Combining the two can accurately extract the seasonal fluctuation patterns of marine environmental variables.

[0095] This step applies a sliding window average to the historical observation sequence. For example, a 60-day window is used to calculate the long-term trend of sea surface temperature, filtering out diurnal fluctuations to highlight trends beyond the monthly scale. The main period frequency is identified through Fourier transform, and the seasonal cycle is determined by combining this with the autocorrelation function. For instance, the total precipitation sequence may exhibit annual seasonality, while significant wave height may show tidal-related fluctuations with a bi-monthly cycle. After subtracting the trend and seasonal components from the original sequence, the residual components retain high-frequency random fluctuations and anomalous event signals, providing crucial input for subsequent anomaly detection.

[0096] In step S3012, the trend component is input into the first feature embedding unit, and numerical embedding, location embedding, and time stamp embedding operations are performed sequentially to obtain a first high-dimensional embedding sequence; the seasonal component is input into the second feature embedding unit, and numerical embedding, location embedding, and time stamp embedding operations are performed sequentially to obtain a second high-dimensional embedding sequence; the residual component is input into the third feature embedding unit, and numerical embedding, location embedding, and time stamp embedding operations are performed sequentially to obtain a third high-dimensional embedding sequence.

[0097] The feature embedding unit contains modular processing units for numerical embedding (mapping continuous physical variable values ​​into high-dimensional vectors), location embedding (encoding spatial grid coordinate information), and time stamp embedding (annotating time features at time points), realizing the fusion representation of multimodal features.

[0098] Step S3013: The first high-dimensional embedding sequence, the second high-dimensional embedding sequence, and the third high-dimensional embedding sequence are fused to obtain the historical observation sequence in the form of a high-dimensional feature tensor.

[0099] Three high-dimensional embedding sequences are fused using a 3D stitching or attention mechanism to form a high-dimensional feature tensor with channel dimension including trend, seasonal, and residual features, spatial dimension corresponding to regional grid, and time dimension labeled with timestamps.

[0100] This embodiment significantly improves the spatiotemporal feature extraction capability and prediction reliability of the marine heat wave prediction model through the synergistic optimization of trend-seasonal-residual decomposition and multi-dimensional feature embedding. Moving average and frequency domain analysis accurately extract the multi-scale variation characteristics of physical variables such as sea surface temperature and 2-meter air temperature, enabling the composite model to learn the evolutionary patterns of long-term trends, periodic fluctuations, and random anomalies. The numerical-location-timestamp three-dimensional embedding mechanism achieves deep fusion of spatial distribution, temporal evolution, and numerical features. For example, in nearshore complex topographic areas, high-resolution location embedding can accurately capture local sea surface temperature anomalies, while timestamp embedding enhances the heat wave prediction capability during seasonal transitions. The high-dimensional tensor generated by feature fusion breaks through the linear expression limitations of traditional sequence models, enabling the composite model to simultaneously handle spatiotemporal dependencies and multivariate coupling relationships through a convolutional-recurrent hybrid architecture. In the heat wave analysis module, this high-dimensional feature representation can accurately locate the heat wave occurrence area, quantify the intensity level, and predict the duration. Similarly, step S302 converts the historical time-series data of the physical variables into historical time-series data of the physical variables in the form of high-dimensional feature tensors, as detailed in this embodiment.

[0101] In one embodiment, step S1024, which involves inputting the training dataset into a pre-trained composite model for training to obtain a trained composite model, includes:

[0102] Step S10241: Input the training dataset into the pre-trained composite model to obtain the physical variable prediction sequence corresponding to each training sample generated by the prediction module.

[0103] Training samples containing historical observation sequences and corresponding target sequences are input into the composite model. The input adapter first performs trend-seasonal-residual decomposition and 3D embedding on the historical observation sequences, generating high-dimensional feature tensors that are input into the backbone network. After processing the high-dimensional features, the backbone network generates hidden states, which are then converted into physical variable prediction sequences through a linear mapping layer. For example, daily predictions of sea surface temperature and significant wave height for the next 7 days are made, resulting in an output whose dimension matches that of the target sequence.

[0104] Step S10242: Based on the loss values ​​of the predicted sequence and the corresponding target sequence corresponding to each training sample, update the parameters of the input adapter and the trainable low-rank matrix, and update the parameters of the linear mapping layer of the prediction module.

[0105] The MSE loss values ​​between the predicted and target sequences are calculated, taking into account factors such as sea surface temperature prediction error and 2-meter temperature deviation, to form a multivariate joint loss value. Based on backpropagation of the loss value, the feature embedding parameters of the input adapter, the trainable low-rank matrix of the backbone network, and the linear layer weights of the prediction module are updated preferentially. For example, the learning rate is dynamically adjusted using the Adam optimizer to ensure that parameter updates are both efficient and stable.

[0106] Step S10243: Repeat the steps of inputting training data, calculating loss value and updating parameters until the loss value is less than the preset loss threshold, then determine that the training of the composite model is complete.

[0107] The training data input, loss calculation, and parameter update steps are repeated. In each iteration, the model learns from batch data and gradually adjusts the parameters to reduce prediction error. When the loss value of several consecutive iterations is lower than a preset threshold (e.g., 0.05), the model is considered to have completed training. At this point, the model has fully learned the spatiotemporal evolution patterns in historical data and has high-precision prediction capabilities.

[0108] This embodiment achieves a breakthrough in both prediction accuracy and computational efficiency through a refined training process. The multi-dimensional feature processing of the input adapter, combined with low-rank parameter optimization of the backbone network, effectively solves the overfitting problem caused by parameter redundancy in traditional models. In the iterative training mechanism, the loss-value-driven parameter update strategy ensures the model accurately captures the coupling relationships between multiple physical variables.

[0109] For step S103, the predicted sequence of physical variables for each region, the relationship data of the physical variables, and the preset task instructions are input into the preset heat wave analysis module; wherein, the preset task instructions are used to prompt the heat wave analysis module to analyze and obtain marine heat wave information based on the predicted sequence of physical variables for each region and the relationship data of the physical variables.

[0110] The preset heat wave analysis module is a program module or neural network model specifically designed for analyzing marine heat wave information. It has specific analysis algorithms and logic and can make comprehensive judgments and processes based on the input data.

[0111] The preset task instructions are used to instruct the heatwave analysis module to analyze and derive marine heatwave information based on the input physical variable prediction sequences and physical variable relationship data for each region. Specifically, the preset task instructions can instruct the heatwave analysis module to determine whether a marine heatwave will occur in the target sea area within a certain future time period, and if so, to provide specific information such as its intensity, duration, and affected area.

[0112] In one embodiment, the preset heat wave analysis module in step S103 is obtained through the following steps:

[0113] Step S1031: Obtain the physical variable prediction sequence and physical variable relationship data corresponding to each training sample output by the completed composite model.

[0114] The trained composite model, upon inputting historical observation sequences, generates predicted sequences of physical variables and data on their relationships (such as the correlation weight matrix between sea surface temperature and significant wave height). This output data serves as the input source for the heatwave analysis module, ensuring that the analysis is based on high-precision, multi-dimensional environmental predictions.

[0115] Step S1032: Based on the predicted sequence of physical variables and the relationship data of physical variables corresponding to each training sample, analyze and obtain the first marine heat wave information corresponding to each training sample.

[0116] Based on the predicted sequences and relational data output by the composite model, first-order marine heatwave information is generated using rule-based algorithms (such as threshold judgment and trend analysis). For example, if the predicted sea surface temperature in a certain area exceeds 28°C for three consecutive days and is positively correlated with the significant wave height, the area is determined to be at risk of a heatwave. This step serves as a benchmark comparison for the heatwave analysis module and is used to evaluate the subsequent training effect.

[0117] Step S1033: Input the predicted sequence of physical variables and the relationship data of physical variables corresponding to each training sample, as well as the preset task instruction, into the pre-trained heat wave analysis module.

[0118] The predicted sequences, relational data, and preset task instructions output by the composite model are input into the pre-trained heat wave analysis module. This module performs in-depth analysis of the input data through an internal multilayer perceptron or decision tree structure to generate secondary marine heat wave information, such as quantitative indicators like the probability of heat wave occurrence and the extent of its impact.

[0119] Step S1034: Obtain the second marine heat wave information corresponding to each training sample output by the heat wave analysis module; update the training parameters of the heat wave analysis module according to the second marine heat wave information corresponding to each training sample and the loss value of the corresponding first marine heat wave information; re-input the physical variable prediction sequence and physical variable relationship data corresponding to each training sample and the preset task instruction into the pre-trained heat wave analysis module; until the loss value is less than the preset threshold, the trained heat wave analysis module is obtained.

[0120] By comparing the differences between the first and second ocean heatwave information, a loss value (such as mean squared error) is calculated. The parameters of the heatwave analysis module (such as the weight matrix and decision threshold) are updated using the backpropagation algorithm, and this input-calculation-update process is repeated until the loss value falls below a preset threshold, completing module training. This process ensures a high degree of matching between the heatwave analysis module and the composite model output, improving prediction consistency.

[0121] This embodiment achieves end-to-end optimization of marine heat wave prediction, from data generation to analysis and decision-making, through a collaborative training mechanism between the composite model and the heat wave analysis module. The high-precision physical variable prediction sequences and relational data output by the composite model provide a multi-dimensional input foundation for heat wave analysis, enabling the analysis module to capture key heat wave precursor signals such as abnormal sea surface temperature increases and coordinated changes in significant wave height. The training process of the heat wave analysis module, through loss-value-driven parameter optimization, ensures a high degree of consistency between its output and the composite model's prediction results, avoiding the subjective bias of manual analysis in traditional methods.

[0122] In one embodiment, the preset task instruction is further used to prompt the heat wave analysis module to output the marine heat wave information according to a preset structured analysis report template.

[0123] The preset task command adds a structured report generation function to the traditional heatwave analysis. This command uses natural language processing technology to parse user requirements and automatically fills the marine heatwave information (such as occurrence time, intensity, and impact range) output by the heatwave analysis module into a preset template, generating a standardized analysis report containing text, tables, and trend graphs. For example, the command can specify that the report should include sections such as "Overview of Heatwave Event," "Correlation Analysis of Physical Variables," and "Early Warning Recommendations," and requires output in PDF or HTML format.

[0124] This embodiment implements an automatic structured report generation function, transforming heat wave analysis results into directly usable decision-making basis and reducing manual processing time. By using preset templates, a unified report structure is enforced, avoiding omissions of key information or inconsistencies in expression caused by manual writing.

[0125] For step S104, the marine heat wave information output by the heat wave analysis module is obtained; based on the marine heat wave information, the marine heat wave prediction result for the target sea area is obtained.

[0126] Among them, marine heat wave information is the specific content about marine heat waves output by the heat wave analysis module after analysis. Specifically, it can include information such as whether a marine heat wave will occur, the specific time of occurrence, the duration of duration, the area affected, and the intensity level of the heat wave.

[0127] Marine heatwave forecasts are overall assessments and summaries of future marine heatwave conditions in target sea areas, based on acquired marine heatwave information. These forecasts can provide decision-making support for relevant departments and personnel. For example, fisheries departments can adjust their fishing plans based on the forecasts to avoid operations during marine heatwaves and reduce economic losses. Disaster prevention and mitigation departments in coastal areas can prepare in advance, such as reinforcing coastal protection facilities and issuing early warnings, to ensure the safety of people's lives and property.

[0128] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and this application also intends to include these modifications and variations.

Claims

1. A method of marine heat wave prediction, characterized by, The method comprises the following steps: obtaining physical variable historical time series data of each region in a target sea area; the physical variable historical time series data comprises physical variable observation values at a plurality of historical time points; inputting the physical variable historical time series data of each region into a preset composite model to obtain physical variable prediction sequences of each region and physical variable relationship data; wherein the composite model is used to generate physical variable prediction sequences and physical variable relationship data according to the physical variable historical time series data; the physical variable prediction sequences comprise physical variable prediction values at a plurality of future time points; inputting the physical variable prediction sequences of each region, the physical variable relationship data and a preset task instruction into a preset heat wave analysis module; wherein the preset task instruction is used to prompt the heat wave analysis module to analyze and obtain marine heat wave information according to the physical variable prediction sequences of each region and the physical variable relationship data; obtaining marine heat wave information output by the heat wave analysis module; and obtaining marine heat wave prediction results of the target sea area according to the marine heat wave information; Before the step of inputting the physical variable historical time series data of each region into a preset composite model to obtain physical variable prediction sequences of each region and physical variable relationship data, the method comprises the following steps: obtaining historical observation data of a target sea area; the historical observation data comprises physical variable time series data of each region in the target sea area; the physical variable time series data comprises physical variable observation values sorted in time sequence; splitting to obtain a plurality of historical observation sequences and a target sequence corresponding to each historical observation sequence according to the historical observation data; wherein the historical observation sequence comprises physical variable observation values of a plurality of continuous first historical time points of each region; the target sequence comprises physical variable observation values of a plurality of continuous second historical time points of each region, and the second historical time points are continuous time points subsequent to the first historical time points; generating a training data set according to each historical observation sequence and a target sequence corresponding to each historical observation sequence; wherein the training data set comprises a plurality of training samples, each training sample corresponds to a group of historical observation sequences, and a training label of each training sample is a target sequence corresponding to the group of historical observation sequences; inputting the training data set into a pre-trained composite model for training to obtain a trained composite model; the composite model comprises an input adapter, a pre-trained large model backbone network, a prediction module and a relationship mining module; the pre-trained large model backbone network comprises a trainable low-rank matrix; the input adapter is connected to the pre-trained large model backbone network, and the pre-trained large model backbone network is connected to the prediction module and the relationship mining module; the input adapter is used to decompose and embed the historical observation sequence data to obtain high-dimensional features, and transmit the high-dimensional features to the pre-trained large model backbone network; The pre-trained large model backbone network is configured to receive the high-dimensional features and process the high-dimensional features to obtain a high-dimensional hidden state, and transmit the high-dimensional hidden state to the prediction module and the relationship mining module; The prediction module is configured to convert the high-dimensional hidden state into a physical variable prediction sequence through a linear mapping layer; The relationship mining module is configured to convert the high-dimensional hidden state into physical variable relationship data through an attention mechanism; The relationship mining module is configured to extract a feature vector corresponding to a prediction start time point from the high-dimensional hidden state as a query vector, and extract feature vectors corresponding to all historical time points in the high-dimensional hidden state as key vectors; Transposing the last two dimensions of the key vectors to obtain transposed key vectors, and calculating the dot product similarity between the query vector and the transposed key vectors to obtain an association score; Performing normalization processing on the association score in the historical time dimension to obtain an influence score corresponding to each historical time point, and obtaining physical variable relationship data according to the influence score.

2. The marine heat wave prediction method according to claim 1, wherein the step of obtaining historical observation data of the target sea area comprises: obtaining the off-shore distance and the seafloor topographic gradient of each region of the target sea area; determining the region type and the corresponding geographical resolution of each region according to the off-shore distance and the seafloor topographic gradient of each region; wherein the region type includes a first type and a second type; the off-shore distance of the first type region is less than a preset distance or the fluctuation degree of the seafloor topographic gradient is greater than a preset fluctuation degree; the first type region corresponds to a first geographical resolution, and the second type region corresponds to a second geographical resolution; the first geographical resolution is greater than the second geographical resolution; obtaining observation sequence data of the target sea area at an original resolution; the observation sequence data includes physical variable values of each region of the target sea area at a plurality of historical time points; for the observation sequence data at the original resolution, sampling physical variable time series data of the first type region according to the first geographical resolution, and sampling physical variable time series data of the second type region according to the second geographical resolution; wherein the first geographical resolution is greater than the second geographical resolution; obtaining the historical observation data of the target sea area according to the sampled physical variable time series data of the first type region and the physical variable time series data of the second type region.

3. The marine heat wave prediction method according to claim 2, wherein the step of sampling physical variable time series data of the first type region according to the first geographical resolution and sampling physical variable time series data of the second type region according to the second geographical resolution further comprises: obtaining meteorological information of the target sea area at a plurality of historical time points; determining the time resolution corresponding to each historical time point according to the meteorological information of each historical time point; wherein the first type of meteorological information corresponds to a first time resolution, and the second type of meteorological information corresponds to a second time resolution; the disturbance degree of the first type of meteorological information to the sea area is greater than that of the second type of meteorological information; the first time resolution is greater than the second time resolution; ​ ​ According to the first geographical resolution and the time resolution, the physical variable time series data of the first type of region is sampled; and according to the second geographical resolution and the time resolution, the physical variable time series data of the second type of region is sampled.

4. The marine heat wave prediction method according to claim 1, characterized in that, Before the step of generating the training data set according to each of the historical observation sequences and the target sequence corresponding to each of the historical observation sequences, the method comprises: Converting each of the historical observation sequences into a historical observation sequence in the form of a high-dimensional feature tensor; Before the step of inputting the physical variable historical time series data of each region into a preset composite model to obtain a physical variable prediction sequence of each region and physical variable relationship data, the method comprises: Converting the physical variable historical time series data into physical variable historical time series data in the form of a high-dimensional feature tensor.

5. The marine heat wave prediction method according to claim 4, characterized in that, The step of converting each of the historical observation sequences into a historical observation sequence in the form of a high-dimensional feature tensor comprises: Performing a moving average process on the historical observation sequence to obtain a trend component, performing a frequency domain conversion on the historical observation sequence and performing an autocorrelation analysis to obtain a seasonal component, and subtracting the trend component and the seasonal component from the historical observation sequence to obtain a residual component; Inputting the trend component into a first feature embedding unit to sequentially perform a numerical embedding, a position embedding and a time stamp embedding operation to obtain a first high-dimensional embedded sequence, inputting the seasonal component into a second feature embedding unit to sequentially perform a numerical embedding, a position embedding and a time stamp embedding operation to obtain a second high-dimensional embedded sequence, and inputting the residual component into a third feature embedding unit to sequentially perform a numerical embedding, a position embedding and a time stamp embedding operation to obtain a third high-dimensional embedded sequence; Fusing the first high-dimensional embedded sequence, the second high-dimensional embedded sequence and the third high-dimensional embedded sequence to obtain the historical observation sequence in the form of a high-dimensional feature tensor.

6. The marine heat wave prediction method according to claim 1, characterized in that, The step of inputting the training data set into a pre-trained composite model for training to obtain a trained composite model comprises: Inputting the training data set into the pre-trained composite model to obtain a physical variable prediction sequence corresponding to each training sample generated by the prediction module; Updating parameters of the input adapter and the trainable low-rank matrix and updating linear mapping layer parameters of the prediction module according to a loss value of the physical variable prediction sequence corresponding to each training sample and the target sequence corresponding to each training sample; Repeating the steps of training data input, loss value calculation and parameter updating until the loss value is less than a preset loss threshold, and determining that the composite model is trained.

7. The marine heat wave prediction method according to claim 1, characterized in that, Before the step of inputting the physical variable prediction sequence of each region, the physical variable relationship data and a preset task instruction into a preset heat wave analysis module, the method further comprises: Obtaining the physical variable prediction sequence corresponding to each training sample output by the trained composite model and the physical variable relationship data; According to the physical variable prediction sequence corresponding to each training sample and the physical variable relationship data, the first marine heat wave information corresponding to each training sample is obtained by analysis; The physical variable prediction sequence corresponding to each training sample and the physical variable relationship data and the preset task instruction are input into the pre-trained heat wave analysis module; Obtaining the second marine heat wave information corresponding to each training sample output by the heat wave analysis module; updating the training parameters of the heat wave analysis module according to the second marine heat wave information corresponding to each training sample and the loss value of the corresponding first marine heat wave information; re-inputting the physical variable prediction sequence corresponding to each training sample and the physical variable relationship data and the preset task instruction into the pre-trained heat wave analysis module; until the loss value is less than the preset threshold, the trained heat wave analysis module is obtained.

Images