Artificial intelligence-based seawater carbon dioxide concentration prediction and early warning method and system

Abstract

The present application relates to the technical field of marine environment monitoring and early warning, in particular to a seawater carbon dioxide concentration prediction and early warning method and system based on artificial intelligence, in the present application, the adjacent weight calculation is performed on the concentration sequence of the upstream node and the downstream node through the graph attention network, the arrival of the sea current front, the passage of the upwelling band and the parking period of the stagnant water mass are uniformly included in the weight distribution process, the propagation influence between different nodes is reflected in the form of numerical weight, the strong propagation path and the delayed propagation path are distinguished, the concentration change is no longer presented in isolation, the continuous time window concentration segment is cut out on the propagation constraint sequence, the isolated forest is used to perform the split level division on the segment, the easy isolation segment and the normal segment are distinguished, the jump segment, the slow rise segment and the reciprocating jump segment are grouped by the amplification order and the interval distribution, the abnormal recognition is changed from single point threshold to joint determination based on structural difference and distribution sparsity, and the abnormal evolution process can be identified in the pre-threshold stage.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring and early warning technology, and in particular to a method and system for predicting and warning seawater carbon dioxide concentration based on artificial intelligence. Background Technology

[0002] Seawater carbon dioxide concentration is a crucial parameter characterizing the state of air-sea exchange, changes in the marine carbon cycle, and the risk of localized ocean acidification. Especially in nearshore waters, semi-enclosed bays, and aquaculture areas, abnormal increases in seawater carbon dioxide concentration can easily lead to water acidification, ecological disturbance, and deterioration of the aquaculture environment. Because seawater carbon dioxide concentration is influenced by ocean currents, upwelling activity, stagnant water masses, air-sea exchange, and changes in temperature, salinity, and pH, it exhibits significant spatiotemporal coupling characteristics. Therefore, it is necessary to establish prediction and early warning methods specifically for seawater carbon dioxide concentration.

[0003] In existing technologies, early warning methods for marine environmental indicators mostly adopt single-point monitoring, fixed threshold discrimination, or ordinary time series prediction. They usually focus on statistical analysis of changes in monitoring values, lacking a dedicated data processing chain based on the formation mechanism of seawater carbon dioxide concentration, and especially lacking comprehensive consideration of seawater carbon dioxide concentration monitoring data acquisition, auxiliary parameter correction processing, and detailed classification of early warning levels.

[0004] Existing technologies focus on identifying, assessing, predicting, and issuing early warnings about risk sources, formation conditions, evolution paths, and impact ranges, emphasizing a shift from post-event response to pre-event warning and in-event intervention. However, in practice, these technologies often rely on comprehensive judgments based on overall trends, abnormal increases, and durations, easily treating changes in seawater carbon dioxide concentration as a single time series problem. They lack detailed constraints on spatial diffusion chains, differences in ocean current direction, upstream and downstream transmission rhythms, and local lag relationships. This makes it difficult to distinguish between common sources and cascading propagation scenarios when multiple sea areas experience simultaneous increases, resulting in a coarse risk source identification. Furthermore, existing technologies emphasize closed-loop control mechanisms and... While quantitative evidence exists, in practice, early warnings are often triggered by uniform thresholds or overall anomaly levels. When faced with different evolutionary patterns such as slow rises, repeated oscillations, and local jumps, early anomalies are easily mistaken for normal fluctuations, and short-term disturbances are misjudged as high-level risks. Existing technologies lack detailed expressions for continuous over-limit periods, contribution ratio ranking, boundary node backtracking, and regional connectivity. Early warning outputs mostly remain at the level of whether there is a risk and the level of risk, making it difficult to directly support zonal control, graded disposal, and sequential intervention, thereby increasing regulatory costs and causing a lag in the response to ecological disturbance risks, ocean acidification risks, and aquaculture environment deterioration risks during the diffusion process. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based method and system for predicting and warning seawater carbon dioxide concentration, in order to solve the problems of insufficient data acquisition and utilization, inadequate characterization of spatiotemporal propagation relationship, and insufficiently refined warning classification in the existing technology for this specific marine environmental parameter of seawater carbon dioxide concentration.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence, comprising the following steps:

[0007] S1: Acquire seawater carbon dioxide concentration monitoring data and auxiliary parameter data related to changes in seawater carbon dioxide concentration in the target sea area. The auxiliary parameter data includes water temperature, salinity, pH value, total alkalinity, ocean current velocity, and ocean current direction. Perform time alignment, spatial registration, missing value repair, and outlier removal on the acquired data to generate a basic dataset of seawater carbon dioxide concentration.

[0008] S2: Based on the aforementioned seawater carbon dioxide concentration dataset and target sea area range data, delineate the boundaries of the ocean basin partition, verify the beginning and end nodes of the ocean current channel, check the adjacent positions of the air-sea exchange interface, identify the transmission direction of carbon sink nodes and carbon source nodes, delete transmission breakpoints, write the edge weight level, and establish a propagation topology grid.

[0009] S3: Based on the propagation topology grid, align the seawater carbon dioxide concentration sequences of upstream and downstream nodes, use a graph attention network to compare the arrival order of ocean current fronts, filter out the upwelling band passage chains, delineate the dwelling time of stagnant water masses, identify strong propagation chains and slow propagation chains, and obtain the propagation constraint sequence.

[0010] S4: Based on the propagation constraint sequence, cut out continuous time window segments of seawater carbon dioxide concentration, use isolated forests, arrange the increase order, separate the falling segment and the rising segment, lock the sparse segment, and group the sudden jump segment, the gradual rise segment and the reciprocating jump segment to obtain the abnormal state cluster.

[0011] S5: Based on the anomalous state cluster and propagation constraint sequence, load the short window sequence and long window sequence, select the high proportion trajectory, expand the seawater carbon dioxide concentration sequence for future periods, backfill the repeated residual segments, correct the peak shift segments, and obtain the seawater carbon dioxide concentration evolution trajectory.

[0012] S6: Based on the evolution trajectory of seawater carbon dioxide concentration and the cluster of abnormal states, lock the threshold nodes, rank the upstream contribution ratios, accumulate the continuous over-limit periods, trace back the diffusion links and boundary nodes, splice the regional connectivity pieces, and generate graded early warning results by combining the seawater carbon dioxide concentration threshold, concentration growth rate, continuous over-limit duration and regional diffusion range.

[0013] As a further aspect of the present invention, the basic dataset of seawater carbon dioxide concentration includes seawater carbon dioxide concentration monitoring values, time stamps, spatial location stamps, and auxiliary parameter stamps; the propagation topology grid includes basin partitioning units, unidirectional transmission links, and edge weight hierarchy stamps; the propagation constraint sequence includes propagation order, propagation strength classification, and dwell time period stamps; the abnormal state cluster includes abrupt jump state segments, gradual rise state segments, and reciprocating jump state segments; the seawater carbon dioxide concentration evolution trajectory includes short window change trajectory, long window change trajectory, and offset correction trajectory; and the graded early warning results include sets of over-threshold nodes, diffusion link sets, and boundary node sets corresponding to different early warning levels.

[0014] As a further aspect of the present invention, the specific steps for generating the propagation topology mesh in step S2 are as follows:

[0015] Based on the aforementioned seawater carbon dioxide concentration dataset and target sea area range data, the boundaries of the ocean basin are delineated, the positions of the first and last nodes of the ocean current channel are checked segment by segment, the correspondence between adjacent positions of the air-sea exchange interface is verified, the transmission orientation of carbon sink nodes and carbon source nodes is identified, the reachable direction, disconnection position and adjacent order are recorded, and the bottom table of the connection edge is obtained.

[0016] Based on the aforementioned edge table, delete the edges corresponding to transmission breakpoints, write high transmission level, medium transmission level and low transmission level, merge continuous links in the same direction, eliminate reverse conflict links, correct the order of adjacent nodes and boundary connection relationship, and establish a propagation topology mesh.

[0017] As a further aspect of the present invention, the specific steps for generating the propagation constraint sequence in step S3 are as follows:

[0018] Based on the propagation topology grid, the seawater carbon dioxide concentration sequences of upstream and downstream nodes are aligned, and nodes are matched and connected according to adjacent sampling times. Node pairs with missing values ​​and reverse misaligned node pairs are removed. The node number, time difference number, connection direction, concentration difference, and number of consecutive occurrences are recorded to obtain the propagation time series table.

[0019] Based on the propagation timeline, a graph attention network is used to compare the arrival order of ocean current fronts, filter out continuous upwelling chains, divide the stationary water mass stationary time period according to the stationary duration threshold, and list continuous passage sections, intermittent transition sections and stationary sections. Write the start and end positions of the sections, the duration, and the correspondence between adjacent nodes to obtain the link segmentation table.

[0020] Based on the link segmentation table, the link order is determined according to edge weight level, arrival order difference, dwell time ratio, and continuous passage length. Strong propagation links and slow propagation links are identified, and node transmission restrictions, segment passage restrictions, time period restrictions, and direction restrictions are written to obtain the propagation constraint sequence.

[0021] As a further aspect of the present invention, the graph attention network first performs an adjacency node traversal on the seawater carbon dioxide concentration sequence corresponding to each node, reads the upstream node number, downstream node number, edge weight level, time difference sequence number, and concentration difference value, performs time-by-time matching calculations on the seawater carbon dioxide concentration sequence of the upstream node and the seawater carbon dioxide concentration sequence of the current node, obtains multiple sets of node pair influence values, normalizes the multiple sets of influence values ​​to form a weight distribution, performs a weighted summation with the corresponding upstream node concentration value to generate node update values, and then performs multi-channel parallel calculations on the multiple sets of node update values, respectively recording the influence values ​​of nearby nodes, delayed nodes, and low-weight nodes, splicing and rearranging the multi-channel calculation results to form a node propagation influence set, and writes it into the link segmentation calculation process.

[0022] As a further aspect of the present invention, the specific steps for generating the abnormal state cluster in step S4 are as follows:

[0023] Based on the propagation constraint sequence, continuous time window segments of seawater carbon dioxide concentration are cut out, and the segments are rearranged according to the window length and the sliding step size. The starting value, ending value, peak position, valley position, change range, duration and interval between adjacent segments of each segment are recorded to obtain the window segment table.

[0024] Based on the window segment table, an isolated forest is used to arrange the increase order, separate the falling segment and the rising segment, lock the sparse segment according to the interval between adjacent segments, the frequency of recurrence, and the degree of amplitude dispersion, merge the continuous segments in the same direction and segments with similar amplitude, write the segment arrangement relationship and concentration position, and obtain the morphological distribution table.

[0025] Based on the morphological distribution table, the segments are classified according to the increase span, duration, number of round trips, and turning interval. Sudden jump segments, gradual rise segments, and reciprocating jump segments are grouped together, and the segment affiliation, position index, adjacency relationship, and aggregation range are written to obtain the abnormal state cluster.

[0026] As a further aspect of the present invention, the isolated forest first reads the starting point value, ending point value, peak position, valley position, change amplitude, duration, and interval between adjacent segments for each window segment to form a sample vector. Multiple sets of segment records and multiple fields are randomly selected from the sample vector. Splitting conditions are set layer by layer according to the field value range. All sample vectors are repeatedly divided, and the number of layers experienced by each window segment before falling into a leaf node is recorded. The corresponding layers of multiple trees are summarized, and the average level is calculated. Window segments are arranged from smallest to largest according to the average level. Segments with smaller average levels are assigned to the easily isolated segment set, segments with medium average levels are assigned to the undecided segment set, and segments with larger average levels are assigned to the normal segment set. Then, combined with the increase value sorting results, upward increase records in the easily isolated segment set are separated and written into the rising segment, and downward increase records in the easily isolated segment set are separated and written into the falling segment. Sparse segments are marked according to the interval between adjacent segments, frequency of recurrence, and amplitude dispersion, and the segment arrangement relationship and concentration position are written into the table to obtain the morphological distribution table.

[0027] As a further aspect of the present invention, the specific steps for generating the seawater carbon dioxide concentration evolution trajectory in step S5 are as follows:

[0028] Based on the aforementioned abnormal state clusters and propagation constraint sequences, short window sequences and long window sequences are loaded, adjacent seawater carbon dioxide concentration change segments are aligned according to time period positions, the proportions of multiple candidate trajectories are compared, and trajectory segments with high proportions, continuous fluctuations, and concentrated turning points are screened out. The seawater carbon dioxide concentration arrangement for future time periods is then expanded to obtain the trajectory bottom sequence.

[0029] Based on the trajectory base sequence, repeated residual segments are backfilled segment by segment, the amplitude corresponding to the peak offset segment is corrected, the slope changes of adjacent time periods are aligned, the interval between abrupt jump intervals and lag intervals is compressed, and the connection sequence and turning point of seawater carbon dioxide concentration in future time periods are rearranged to obtain the evolution trajectory of seawater carbon dioxide concentration.

[0030] As a further aspect of the present invention, the specific steps for generating the graded early warning result in step S6 are as follows:

[0031] Based on the evolution trajectory of seawater carbon dioxide concentration and the cluster of abnormal states, the nodes exceeding the threshold are identified, and the contribution ratios of upstream sources are ranked one by one. The continuous periods of exceeding the limit are accumulated, and the relationship between the starting point, duration, and fallback position of exceeding the limit and the abnormal cluster is recorded. The diffusion conditions and time distribution of the nodes are summarized to obtain a diffusion criterion table.

[0032] Based on the aforementioned diffusion criterion table, the diffusion link is traced back along the high-proportion direction to lock boundary nodes and turning points. Adjacent regions are connected, and diffusion segments from the same source are merged. The warning levels are divided according to the seawater carbon dioxide concentration threshold, concentration growth rate, duration of continuous exceedance, and regional diffusion range. The correspondence between the warning level, trigger time period, and coverage area is written to generate a graded warning result.

[0033] This invention also provides an artificial intelligence-based seawater carbon dioxide concentration prediction and early warning system, the system being used to execute the above-described artificial intelligence-based seawater carbon dioxide concentration prediction and early warning method, the system comprising:

[0034] Data acquisition and processing module: used to acquire seawater carbon dioxide concentration monitoring data of the target sea area and auxiliary parameter data related to changes in seawater carbon dioxide concentration, and perform time alignment, spatial registration, missing value repair and outlier removal processing to generate a basic dataset of seawater carbon dioxide concentration.

[0035] Marine network construction module: Based on the aforementioned seawater carbon dioxide concentration dataset and target sea area range data, completes the delineation of ocean basin zoning boundaries, verification of the beginning and end nodes of ocean current channels, verification of adjacent positions of air-sea exchange interfaces, identification of transmission orientation, deletion of breakpoints and writing of edge weights, and establishes a propagation topology mesh.

[0036] Propagation constraint module: Based on the propagation topology grid, align the seawater carbon dioxide concentration sequences of upstream and downstream nodes, register the time difference, direction, and concentration difference, and combine the graph attention network to allocate adjacency weights to distinguish between passage sections and stopping sections, thereby obtaining the propagation constraint sequence;

[0037] Anomaly Clustering Module: Based on the propagation constraint sequence, it cuts out continuous time window segments of seawater carbon dioxide concentration, registers the start and end values, peak and valley positions, amplitude and duration, divides the segments into hierarchical levels in combination with isolated forest, and clusters abrupt jump segments, gradual rise segments and repetitive jump segments to obtain anomaly state clusters;

[0038] Trajectory extrapolation module: Based on the anomalous state cluster and propagation constraint sequence, short window sequence and long window sequence are loaded, high proportion trajectory is selected, seawater carbon dioxide concentration sequence for future time period is expanded, residual fragments are backfilled and peak shift is corrected to obtain the evolution trajectory of seawater carbon dioxide concentration.

[0039] Early warning classification module: Based on the evolution trajectory of seawater carbon dioxide concentration and abnormal state clusters, it locks the threshold-exceeding nodes, accumulates the continuous over-limit period, traces back the diffusion link and boundary nodes, splices the regional connectivity pieces, and generates classified early warning results by combining the seawater carbon dioxide concentration threshold, concentration growth rate, continuous over-limit duration and regional diffusion range.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0041] This invention first acquires monitoring data on seawater carbon dioxide concentration in the target sea area, along with related parameters such as water temperature, salinity, pH, total alkalinity, and ocean current parameters. It then performs time alignment, spatial registration, missing value repair, and outlier removal on the acquired data, thereby improving the effectiveness and reliability of the basic seawater carbon dioxide concentration data. Furthermore, this invention transforms the spatial node relationships from discrete points into a connected structure with directional constraints and transmission intensity distinctions by delineating the boundaries of the ocean basin around the target sea area, verifying the beginning and end nodes of ocean current channels, checking the adjacent positions of the air-sea exchange interface, and writing them into the edge weight hierarchy.

[0042] This invention uses a graph attention network to perform adjacency weight calculation on the seawater carbon dioxide concentration sequence of upstream and downstream nodes. It incorporates the arrival order of ocean current fronts, the upwelling band passage chain, and the stationary period of stagnant water masses into the weight allocation process, so that the propagation influence between different nodes is reflected in the form of numerical weights. Strong propagation paths and delayed propagation paths are distinguished, and the changes in seawater carbon dioxide concentration are no longer presented in isolation, but form an orderly transmission relationship along the spatial link.

[0043] This invention extracts continuous time window segments of seawater carbon dioxide concentration from propagation constraint sequences and performs hierarchical division of the segments through isolated forests to distinguish easily isolated segments from normal segments. Combined with the increase order and interval distribution, it completes the aggregation of abrupt jump segments, gradual increase segments and repetitive jump segments, so that the anomaly identification is transformed from a single point exceeding the threshold to a joint judgment based on structural differences and distribution sparsity. The anomaly evolution process can be identified in the stage before exceeding the threshold.

[0044] This invention unfolds the trajectory of seawater carbon dioxide concentration in future periods by superimposing short-window and long-window sequences, and corrects for repetition residuals and peak shifts, thus taking into account both short-term fluctuations and mid-term upward trends. Subsequently, it combines threshold node locking, upstream contribution ratio ranking, and diffusion link backtracking, and introduces seawater carbon dioxide concentration thresholds, concentration growth rates, duration of continuous exceedance, and regional diffusion range to achieve early warning level classification, so that the early warning results have spatial directionality, temporal continuity, and the ability to support graded response. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0046] Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are only for explaining the invention and are not intended to limit the scope of protection of the invention. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the invention, and all such modifications and improvements should fall within the scope of protection of the invention.

[0048] It should be noted that the "seawater carbon dioxide concentration" in this invention can be seawater carbon dioxide concentration data directly obtained by monitoring equipment, or it can be data characterizing the seawater carbon dioxide state obtained by conversion or inversion based on parameters such as seawater carbon dioxide partial pressure, pH value, total alkalinity, and dissolved inorganic carbon. The "target sea area" in this invention can be nearshore waters, bays, estuaries, semi-enclosed sea areas, aquaculture areas, or other sea areas where seawater carbon dioxide concentration monitoring and early warning are required. The "node" in this invention can be a fixed monitoring point, buoy location, navigation trajectory sampling point, grid center point, or the center point of a sea area calculation unit generated after spatial discretization. To ensure the reliability of subsequent propagation topology construction, anomaly identification, and early warning determination, when implementing this invention, it is preferable to perform unified time reference correction, unified spatial coordinate mapping, and unified dimensional conversion processing on data from different sources.

[0049] Example 1

[0050] In this embodiment, the seawater carbon dioxide concentration data of the target sea area can be obtained from one or more of the following: in-situ monitoring buoys, fixed monitoring stations, mobile observation platforms, shore-based automatic monitoring equipment, or remote sensing inversion results. Simultaneously, auxiliary parameter data related to changes in seawater carbon dioxide concentration are collected. These auxiliary parameters include at least water temperature, salinity, pH value, total alkalinity, ocean current velocity, and ocean current direction. In some embodiments, they may further include one or more of the following: dissolved oxygen, chlorophyll concentration, tide level, wind speed and direction, sea surface temperature, sea surface salinity, and air pressure. These data collectively characterize the formation background, spatiotemporal variation, and propagation driving conditions of seawater carbon dioxide concentration in the target sea area.

[0051] Before proceeding to the subsequent propagation topology construction step, it is preferable to first perform time alignment, spatial registration, missing value repair, and outlier removal on seawater carbon dioxide concentration data and auxiliary parameter data from different sources to form the basic dataset of seawater carbon dioxide concentration used in subsequent propagation analysis.

[0052] Please see Figure 1 This invention provides a technical solution: a method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence, comprising the following steps:

[0053] S1: Acquire seawater carbon dioxide concentration monitoring data and auxiliary parameter data related to changes in seawater carbon dioxide concentration in the target sea area, and perform time alignment, spatial registration, missing value repair and outlier removal on the acquired data to generate a basic dataset of seawater carbon dioxide concentration.

[0054] Specifically, the target sea area can be selected as a nearshore semi-enclosed bay, estuary, or aquaculture area. Fixed buoy monitoring points, shore-based monitoring stations, and mobile observation sections will be deployed within this target sea area. The fixed buoy monitoring points and shore-based monitoring stations will continuously collect data on seawater carbon dioxide concentration, water temperature, salinity, pH, and total alkalinity. The mobile observation sections will supplement the spatial distribution information of the sea area. For remote sensing inversion data, sea surface temperature, sea surface salinity, and chlorophyll inversion products corresponding to the monitoring period can be selected to supplement the areal distribution of the sea area.

[0055] When preprocessing the acquired data, a uniform sampling step of 1 hour is preferred, and data within the same sampling window are time-merged. Data from different sampling locations are uniformly mapped to the latitude and longitude coordinate system of the target sea area, and spatially merged according to a preset rule grid or discrete node set. For data segments with no more than 3 consecutive missing sampling times for a single node, linear interpolation, spline interpolation, or weighted interpolation of adjacent nodes is used to fill in the missing data. For data segments with more than 3 consecutive missing sampling times, the missing marker is retained, and the corresponding node pairs are removed during the subsequent propagation constraint construction.

[0056] When removing outliers, a combined approach using range constraints, rate of change constraints, and statistical outlier rules is preferred. If the change in seawater carbon dioxide concentration at a node exceeds a preset threshold between adjacent sampling times, and no synchronous disturbance is observed in the corresponding water temperature, salinity, and ocean current parameters, this node is marked as a suspected outlier and further filtered using the 3σ criterion, median absolute deviation method, or box plot outlier rules. After this processing, a basic dataset of seawater carbon dioxide concentration is generated, containing monitoring values, time stamps, spatial location markers, and auxiliary parameter markers.

[0057] S2: Based on the basic dataset of seawater carbon dioxide concentration and the target sea area range data, delineate the boundary of the ocean basin, verify the first and last nodes of the ocean current channel, check the adjacent positions of the air-sea exchange interface, identify the transmission direction of carbon sink nodes and carbon source nodes, delete transmission breakpoints, write the edge weight level, and establish the propagation topology grid.

[0058] S3: Based on the propagation topology grid, align the concentration sequences of upstream and downstream nodes, use a graph attention network, compare the arrival order of ocean current fronts, screen out the upwelling band passage chains, delineate the residence time of stagnant water masses, identify strong propagation chains and slow propagation chains, and obtain the propagation constraint sequence.

[0059] S4: Based on the propagation constraint sequence, cut out the continuous time window concentration segment, use isolated forest, arrange the increase order, separate the fall segment and the rise segment, lock the sparse segment, and group the sudden jump segment, the gradual rise segment and the reciprocating jump segment to obtain the abnormal state cluster;

[0060] S5: Based on the abnormal state cluster and propagation constraint sequence, load the short window sequence and long window sequence, select the high proportion trajectory, expand the concentration sequence of future time period, backfill the repeated residual segment, correct the peak shift segment, and obtain the concentration evolution trajectory.

[0061] S6: Based on the evolution trajectory of seawater carbon dioxide concentration and anomalous state clusters, identify the threshold-exceeding nodes, rank the upstream contribution proportions, accumulate continuous over-limit periods, trace back the diffusion links and boundary nodes, splice the regional connectivity pieces, and generate graded early warning results by combining seawater carbon dioxide concentration threshold, concentration growth rate, continuous over-limit duration and regional diffusion range.

[0062] Based on the evolution trajectory of seawater carbon dioxide concentration and anomalous state clusters, threshold-exceeding nodes are identified, the upstream contribution ratios are ranked, the duration of continuous exceedances is accumulated, diffusion links and boundary nodes are traced back, and regional connectivity is pieced together. A tiered early warning result is generated by combining seawater carbon dioxide concentration thresholds, concentration growth rates, duration of continuous exceedances, and regional diffusion range. The propagation topology grid includes basin partitioning units, unidirectional transmission edges, and edge weight hierarchy markers. The propagation constraint sequence includes the propagation order, propagation strength grading, and dwell time markers. Anomalous state clusters include abrupt jump state segments, gradual increase state segments, and reciprocating jump state segments. The concentration evolution trajectory includes short-window change trajectories, long-window change trajectories, and offset correction trajectories. The regional early warning unit includes a set of threshold-exceeding nodes, a set of diffusion links, and a set of boundary nodes.

[0063] After constructing the basic dataset of seawater carbon dioxide concentration, a propagation topology grid is generated based on the dataset, target sea area data, and ocean current transport conditions. The specific steps for generating the propagation topology grid are as follows:

[0064] Based on the basic dataset of seawater carbon dioxide concentration and the target sea area range data, the boundary of the sea basin is delineated, the position of the first and last nodes of the ocean current channel is checked segment by segment, the correspondence of adjacent positions of the air-sea exchange interface is verified, the transmission direction of carbon sink nodes and carbon source nodes is identified, the reachable direction, disconnection position and adjacent order are recorded, and the bottom table of the connection edge is obtained.

[0065] Based on the edge base table, delete the edges corresponding to the transmission breakpoints, write the high transmission level, medium transmission level and low transmission level, merge the continuous links in the same direction, eliminate the reverse conflict links, correct the order of adjacent nodes and the boundary connection relationship, and establish the propagation topology mesh.

[0066] Based on the seawater carbon dioxide concentration dataset and target sea area data, the ray-mapping method was used to determine the closed domain of the basin zoning boundaries. A meridional step size of 0.05 degrees and a zonal step size of 0.05 degrees were set, and the boundary inflection point coordinates were written in clockwise order. Odd / even crossing counts were performed on each grid center point; points with an odd number of crossings were marked as in-domain points, and those with an even number were marked as out-of-domain points. Haversine distance was used to verify the positions of the first and last nodes of the current channel segment by segment. An 8-kilometer threshold was set for the distance between adjacent nodes. The latitude and longitude of the first and last nodes, the latitude and longitude of the last node, and the angle of the main current direction were read to calculate the spherical distance between nodes. The azimuth difference is marked as a directional deviation segment when the azimuth difference is greater than 35 degrees. The correspondence between adjacent positions at the air-sea exchange interface is verified by vector angle calculation. The interface normal vector, node edge vector, and adjacency number are read. The cosine value of the angle is calculated. When the angle is greater than 60 degrees, the adjacency relationship is removed. The transmission orientation of carbon sink nodes and carbon source nodes is identified by the sign direction discrimination rule. The node concentration gradient value, ocean current velocity value, and flow direction code value are read. When the gradient product is less than 0, a reciprocal transmission mark is written. When the gradient product is greater than 0, a same-direction transmission mark is written. The reachable direction, disconnection position, and adjacency order are registered to obtain the bottom edge table.

[0067] Based on the edge table, a disjoint-lookup set merge rule is used to delete and merge the edges corresponding to transmission breakpoints. The edge number, start and end node numbers, breakpoint flag value, and direction flag value are read. When the breakpoint flag value is 1, the corresponding edge is deleted; when the breakpoint flag value is 0, the corresponding edge is retained. A hierarchical weighting rule is used to write the edges to high, medium, and low transmission layers. Flow velocity thresholds are set to 0.8 m / s and 0.3 m / s, respectively. Edges with a flow velocity greater than or equal to 0.8 m / s are written with level 3 weights; edges with a flow velocity greater than or equal to 0.3 m / s but less than 0.8 m / s are written with level 2 weights; and edges with a flow velocity less than 0.3 m / s are written with level 2 weights. Write level 1 edge weights, merge consecutive links in the same direction using topological sorting rules, read the preceding node, succeeding node, direction code, and edge weight level, and perform link splicing when the direction codes are consistent and share node numbers, remove conflicting links using conflict edge removal rules, read mutually inverse edge pairs, direction difference, and edge weight difference, delete inverse edges when the direction difference is 180 degrees and the edge weight difference is less than 1, correct the order of adjacent nodes and boundary connection relationship using breadth-first traversal rules, expand the hierarchical traversal in ascending order of the starting node number, rewrite the access sequence number, parent node number, and boundary access number, and establish the propagation topology mesh.

[0068] In this step, based on the propagation topology grid and the basic dataset of seawater carbon dioxide concentration, the seawater carbon dioxide concentration sequences of upstream and downstream nodes are aligned, and a propagation constraint sequence is generated by combining the arrival order of ocean current fronts, the upwelling band pathway, and the residence time of stagnant water masses. The specific steps for generating the propagation constraint sequence are as follows:

[0069] Based on the propagation topology grid, the concentration sequences of upstream and downstream nodes are aligned, and nodes are matched and connected according to adjacent sampling times. Pairs of nodes with missing values ​​and reverse misaligned nodes are removed. The node number, time difference number, connection direction, concentration difference, and number of consecutive occurrences are recorded to obtain the propagation time series table.

[0070] Based on the propagation timeline, a graph attention network is used to compare the arrival order of ocean current fronts, filter out continuous upwelling chains, divide the stationary water mass into stationary periods according to the stationary duration threshold, and list continuous passage sections, intermittent transition sections and stationary sections. The start and end positions of the sections, the duration, and the correspondence between adjacent nodes are written to obtain the link segmentation table.

[0071] Based on the link segmentation table, the link order is determined according to the edge weight level, arrival order difference, dwell time ratio, and continuous passage length. Strong propagation links and slow propagation links are identified, and node transmission restrictions, segment passage restrictions, time period restrictions, and direction restrictions are written to obtain the propagation constraint sequence.

[0072] Based on the propagation topology grid, a time-alignment matching rule is used to perform time-by-time alignment of the concentration sequences of upstream and downstream nodes. The sampling interval is set to 1 hour. The node number, timestamp sequence, concentration value sequence, and edge number are read. Matching is performed when the timestamp difference is less than or equal to 3600 seconds. The edges of adjacent sampling times are bound one by one. The missing value removal rule is used to delete node pairs with empty concentration values ​​or more than 3 consecutive missing values. The direction consistency verification rule is used to compare the edge direction with the growth direction of the time series. When the direction difference is greater than 180 degrees, it is determined to be a reverse misaligned node pair and deleted. The node number, time difference sequence number, edge direction are registered, the concentration difference is calculated, and the number of consecutive occurrences is accumulated to generate a propagation time series table.

[0073] Based on the propagation time series table, a graph attention network is used to perform weight calculation on the adjacency relationship of nodes. The number of attention heads is set to 4, and the node feature dimension is 16. The node number, adjacent node number, edge weight level, concentration difference, and time difference sequence number are read. The node feature vector is input into a linear transformation matrix with a matrix dimension of 16x16. Weighted concatenation is performed on the features of adjacent nodes. The LeakyReLU activation function is used with a negative slope of 0.2 to calculate the attention coefficient of the node pair. The Softmax function is used to normalize the coefficients in the neighborhood of the same node. The arrival order of ocean current fronts is compared, the continuous passage chain of the upwelling zone is screened, and the dwell time of the stagnant water mass is divided. The dwell time threshold is set to 6 hours. The dwell time greater than or equal to 6 hours is classified into the dwell segment, and the dwell time less than 6 hours is classified into the continuous passage segment. The continuous passage segment, the intermittent transition segment, and the dwell segment are divided. The start and end positions of the segment, the duration, and the correspondence of adjacent nodes are written to generate a link segment table.

[0074] Based on the link segmentation table, a multi-index sorting rule is used to perform comprehensive sorting of links. The edge weight level, arrival order difference, dwell time ratio, and continuous passage length are read. The edge weight level is assigned values ​​of 3, 2, and 1. The arrival order difference is normalized with a normalization range of 0 to 1. The dwell time ratio is converted proportionally, and the continuous passage length is standardized by interval. The comprehensive score is calculated with weight coefficients of 0.4, 0.2, 0.2, and 0.2, and the links are sorted in order. A threshold division rule is used to classify links with a comprehensive score greater than 0.7 into strong propagation links and links with a comprehensive score less than or equal to 0.7 into slow propagation links. Node transmission restrictions, segment passage restrictions, time period restrictions, and direction restrictions are written to generate a propagation constraint sequence.

[0075] The graph attention network first performs a neighbor node traversal on the concentration sequence corresponding to each node, reading the upstream node number, downstream node number, edge weight level, time difference sequence number, and concentration difference. It then performs time-by-time matching calculations between the upstream node concentration sequence and the current node concentration sequence to obtain multiple sets of node pair influence values. These multiple sets of influence values ​​are normalized to form a weight distribution. The weight distribution is then weighted and summed with the corresponding upstream node concentration value to generate the node update value. The multiple sets of node update values ​​are then calculated in parallel across multiple channels, recording the influence values ​​of nearby nodes, delayed nodes, and low-weight nodes. The multi-channel calculation results are then concatenated and rearranged to form a set of node propagation influence values, which are then written into the link segmentation calculation process.

[0076] In this step, a sliding time window slice is applied to the seawater carbon dioxide concentration change process corresponding to the propagation constraint sequence, and anomalous evolution patterns such as sudden jumps, gradual increases, and reciprocating jumps are identified to obtain anomalous state clusters. The specific steps for generating anomalous state clusters are as follows:

[0077] Based on the propagation constraint sequence, continuous time window concentration segments are cut out, and the segment order is rearranged according to the window length and sliding step size. The starting value, ending value, peak position, valley position, change range, duration and interval between adjacent segments of each segment are recorded to obtain the window segment table.

[0078] Based on the window segment table, an isolated forest is used to arrange the increase order, separate the falling segment and the rising segment, lock the sparse segment according to the interval between adjacent segments, the frequency of repetition, and the degree of amplitude dispersion, merge the continuous segments in the same direction and segments with similar amplitude, write the segment arrangement relationship and concentration position, and obtain the morphological distribution table.

[0079] Based on the morphological distribution table, the segments are classified according to the increase span, duration, number of round trips, and turning interval. Sudden jump segments, gradual rise segments, and reciprocating jump segments are grouped together, and the segment affiliation, position index, adjacency relationship, and aggregation range are written to obtain the abnormal state clusters.

[0080] Based on the propagation constraint sequence, a sliding time window slicing rule is used to perform continuous segment extraction on the concentration sequence. The window length is set to 12 sampling times and the sliding step size is 3 sampling times. The node number, timestamp sequence, concentration value sequence, and constraint mark sequence are read. The window is extracted by incrementing the starting index from 1 to subtracting 12 from the end. The extreme value positioning rule is used to perform point-by-point comparison of the concentration value sequence within each window. The first sample value is recorded as the starting value and the last sample value as the ending value. The index of the maximum value is the peak position and the index of the minimum value is the valley position. The difference calculation rule is used to calculate the change range by subtracting the starting value from the ending value. The duration conversion rule is used to multiply the window length by the sampling interval of 3600 seconds to obtain the duration. The segment interval calculation rule is used to calculate the interval between adjacent segments by subtracting the end time of the current window from the start time of the next window. The window number is written, the segment order is rearranged, and the node association number is written to generate a window segment table.

[0081] Based on the window segment table, an isolated forest is used to perform hierarchical partitioning of window segments. The number of decision trees is set to 200, the number of subsamples to 256, the maximum tree depth to 8 levels, and the random seed to 42. Seven fields are read: start value, end value, peak position, valley position, variation range, duration, and interval between adjacent segments. Fields and splitting thresholds are extracted using a uniform random method. The field threshold is a floating-point number between the current minimum and maximum value of the field. Samples are split layer by layer until the maximum tree depth is reached or only one record is retained at a leaf node. The path length of each window segment within the 200 trees is calculated, and the average number of levels is determined. The increment order is determined by the hierarchy from smallest to largest. Segments with a change magnitude greater than 0 are written into the rising segment, and segments with a change magnitude less than 0 are written into the falling segment. The frequency of recurrence is accumulated using frequency statistics rules, and the amplitude dispersion is calculated by dividing the standard deviation by the mean using the dispersion coefficient calculation rules. The threshold for the interval between adjacent segments is set to 2 sampling times, the threshold for the frequency of recurrence is 3 times, and the threshold for the dispersion coefficient is 0.6. Segments that meet the conditions of an interval greater than 2, a frequency less than 3, and a dispersion coefficient greater than 0.6 are written into the sparse segment. Then, continuous segments in the same direction are merged, segments with similar amplitudes are merged, the segment arrangement relationship is written, and the concentrated position is written to generate a morphological distribution table.

[0082] Based on the morphological distribution table, a threshold classification rule is used to classify the segments. The amplification span, duration, number of round trips, turning interval, arrangement relationship, and concentration position are read. The threshold for amplification span is set to 1.8 times the average change amplitude of the most recent 30 windows, the threshold for duration is 6 sampling times, the threshold for number of round trips is 2 times, and the threshold for turning interval is 3 sampling times. Segments with an amplification span greater than the threshold and a duration less than 6 are written into the abrupt jump segment, segments with an amplification span less than the threshold and a duration greater than or equal to 6 are written into the gradual rise segment, and segments with a number of round trips greater than or equal to 2 and a turning interval less than or equal to 3 are written into the reciprocating jump segment. The adjacency merging rule is used to merge the cluster range of segments with continuous positions and a concentration position difference of no more than 2 window numbers. The segment ownership, position index, adjacency relationship, and cluster range are written according to the segment category. The numbering of segments of the same type is sorted, the numbering of cross-segment adjacencies is registered, and the numbering of cluster boundaries is registered to obtain the abnormal state cluster.

[0083] For isolated forests, the starting point, ending point, peak position, valley position, change amplitude, duration, and interval between adjacent segments are read for each window segment to form a sample vector. Multiple sets of segment records and multiple fields are randomly selected from the sample vector. Splitting conditions are set layer by layer according to the field value range. All sample vectors are repeatedly divided, and the number of layers experienced by each window segment before falling into a leaf node is recorded. The corresponding layers of multiple trees are summarized and the average layer is calculated. The window segments are arranged from small to large according to the average layer. The segments with smaller average layers are assigned to the easily isolated segment set, the segments with medium average layers are assigned to the undecided segment set, and the segments with larger average layers are assigned to the normal segment set. Then, combined with the increase value sorting results, the upward increase records in the easily isolated segment set are separated and written into the rising segment, and the downward increase records in the easily isolated segment set are separated and written into the falling segment. Sparse segments are marked according to the interval between adjacent segments, the frequency of recurrence, and the degree of amplitude dispersion. The segment arrangement relationship and concentration position are written into the table to obtain the morphological distribution table.

[0084] In this step, the changes in seawater carbon dioxide concentration over future periods are extrapolated based on anomaly clusters and propagation constraint sequences, resulting in the evolution trajectory of seawater carbon dioxide concentration. The specific steps for generating the concentration evolution trajectory are as follows:

[0085] Based on the abnormal state clusters and propagation constraint sequences, short window sequences and long window sequences are loaded, adjacent concentration change segments are aligned according to time period positions, the proportion of multiple candidate trajectories is compared, and trajectory segments with high proportion, continuous fluctuations, and concentrated turning points are screened out. The future time period concentration is then arranged to obtain the trajectory bottom sequence.

[0086] Based on the trajectory base sequence, repeated residual segments are backfilled segment by segment, the amplitude corresponding to the peak offset segment is corrected, the slope change of adjacent time periods is aligned, the interval between abrupt intervals and lag intervals is compressed, and the concentration connection order and turning point of future time periods are rearranged to obtain the concentration evolution trajectory.

[0087] Based on anomalous state clusters and propagation constraint sequences, a dual-channel long short-term memory network is used to perform parallel loading of short-window and long-window sequences. The short-window length is set to 24 sampling times, the long-window length to 7 periodic statistical segments, the number of hidden units in the short window is set to 128, the number of hidden units in the long window is set to 64, the batch size is set to 32, and the learning rounds are set to 80 rounds. The anomalous state number, propagation direction marker, time period restriction marker, short-window concentration value, and long-window statistical value are read. The tensor splicing rule is used to align adjacent concentration change segments according to time period position. The attention weighted calculation formula is used to calculate the proportion of multiple candidate trajectories. The weight coefficient is obtained by the dot product of the short-window state vector and the long-window state vector and then normalized by Softmax. The proportion threshold is set to 0.65, the fluctuation continuity threshold is that the absolute value of the adjacent difference is less than 0.12, and the turning point concentration threshold is that the number of turning points within 3 consecutive sampling times is greater than or equal to 2. Trajectory segments with high proportion, continuous fluctuation, and turning point concentration are screened out. The concentration values ​​of the next 12 sampling times are recursively arranged, the trajectory number is written, and the time period index is rearranged to generate the trajectory base sequence.

[0088] Based on the trajectory base sequence, a frequency band residual correction rule is used to perform segment-by-segment correction on the concentration connection order of future time periods. The residual segment length is set to 6 sampling times, and the periodic retrieval range is set to 4 to 12 sampling times. The predicted concentration value, historical measured concentration value, peak position, valley position, and adjacent slope value are read. The residual sequence is generated by difference operation. The first 3 frequency band components of the amplitude are extracted by discrete Fourier transform. When the frequency band amplitude is greater than 0.08, it is written into the repeated residual segment and segment-by-segment backfilling is performed. The local peak correction rule is used to perform increase or decrease correction on the amplitude corresponding to the peak offset segment. The adjustment coefficient is set to 1.15 and the adjustment coefficient is set to 0.87. The slope alignment rule is used to compare the slope changes of adjacent time periods. When the slope difference is greater than 0.2, the transition point is written by linear interpolation. The interval compression rule is used to compress the interval between the jump interval and the lag interval to 0.6 times the original interval. The concentration connection order of future time periods is rearranged, the turning position is rewritten, and the trajectory boundary number is registered to generate the concentration evolution trajectory.

[0089] In this step, based on the evolution trajectory of seawater carbon dioxide concentration and anomalous state clusters, threshold nodes, diffusion links, boundary nodes, and regional connectivity relationships are identified, and hierarchical early warning results are further generated. The specific steps for generating regional early warning units are as follows:

[0090] Based on the concentration evolution trajectory and anomalous state clusters, the nodes exceeding the threshold are identified, and the contribution ratios of upstream sources are ranked one by one. The continuous periods of exceeding the limit are accumulated, and the relationship between the starting point, duration, and fallback position of exceeding the limit and the anomalous clusters is recorded. The diffusion conditions and time period distribution of nodes are summarized to obtain a diffusion criterion table.

[0091] In this embodiment, the warning levels preferably include at least Level 1, Level 2, and Level 3. The warning levels can be classified based on whether the seawater carbon dioxide concentration exceeds a preset threshold, the rate of increase in concentration, the duration of continuous exceedance, and the number of diffusion nodes in the area. Specifically, a Level 1 warning can be determined when the seawater carbon dioxide concentration exceeds the preset threshold but the rate of increase is small, the duration of continuous exceedance is short, and the number of diffusion nodes is limited; a Level 2 warning can be determined when the seawater carbon dioxide concentration significantly exceeds the preset threshold and the rate of increase is large, the duration of continuous exceedance reaches the preset intermediate level requirement, and the diffusion range covers multiple adjacent nodes; a Level 3 warning can be determined when the seawater carbon dioxide concentration significantly exceeds the preset threshold and the rate of increase continues to expand, the duration of continuous exceedance is long, and the diffusion range covers multiple marine zoning units or key ecologically sensitive areas. Technical personnel can adjust the above thresholds according to the historical background, regulatory requirements, and ecological sensitivity of different marine areas.

[0092] In addition to regional early warning units, the final output can also simultaneously output tiered early warning results. The tiered early warning results preferably include the early warning level, trigger time period, set of nodes exceeding the threshold, set of diffusion links, set of boundary nodes, and corresponding coverage area.

[0093] Based on the diffusion criterion table, the diffusion link is traced back along the high-proportion direction to lock the boundary nodes and turning points, adjacent areas are spliced ​​together, the same-source diffusion segments are merged, and the correspondence between the warning level, trigger time period and coverage area is written to generate regional warning units.

[0094] Based on the concentration evolution trajectory and anomalous state clusters, a propagation contribution inversion rule is used to lock each node exceeding the threshold. The risk threshold is set as the average concentration of the most recent 30 sampling times multiplied by 1.25. The node number, predicted concentration value, upstream node number, edge weight value, anomalous state number, and time period index are read. The weighted proportion calculation formula is used to multiply the weight value of each upstream node by the corresponding concentration increment and divide it by the sum of upstream sources to rank the contribution proportions. A continuous segment accumulation rule is used to perform sequential traversal of sampling times exceeding the risk threshold. The allowable value for continuous exceedance is set to 1 sampling time. When the interruption exceeds 1 time, a new segment is segmented and continuous exceedance time periods are accumulated. A position write-back rule is used to register the correspondence between the exceedance start point, duration, fallback position and anomalous cluster. A criterion summary rule is used to perform joint writing of the number of upstream sources, contribution proportion order, continuous exceedance length, and anomalous state number. The node diffusion conditions, time period distribution, and segment number sorting are summarized to generate a diffusion criterion table.

[0095] Based on the diffusion criterion table, a directed link backtracking rule is used to perform layer-by-layer tracing for high-proportion directions. The contribution proportion backtracking threshold is set to 0.18. The node number, upstream source number, contribution proportion value, direction code, segment number, and time period index are read. A reverse traversal is performed along the edges with a contribution proportion greater than or equal to 0.18. The boundary judgment rule is used to compare the number of adjacent nodes, the termination mark of the connection, and the segment turning mark. When the number of adjacent nodes is equal to 1 and the termination mark is 1, the boundary node is written. When the number of changes in the direction code is greater than or equal to 1, the turning node is written. The connected piece splicing rule is used to merge adjacent area number differences of no more than 1 and continuous time period index segments. The same source merging rule is used to merge diffusion segments with the same source number. The warning level threshold is set as follows: the number of covered nodes is greater than or equal to 5 and the continuous duration is greater than or equal to 6 sampling times to write to level 3; the number of covered nodes is greater than or equal to 3 and the continuous duration is greater than or equal to 4 sampling times to write to level 2; and the rest are written to level 1. The warning level, trigger time period, and coverage correspondence are written to generate regional warning units.

[0096] Please see Figure 2 An AI-based seawater carbon dioxide concentration prediction and early warning system is used to execute the aforementioned AI-based seawater carbon dioxide concentration prediction and early warning method. The system includes:

[0097] Data acquisition and processing module: used to acquire seawater carbon dioxide concentration monitoring data of the target sea area and auxiliary parameter data related to changes in seawater carbon dioxide concentration, and perform time alignment, spatial registration, missing value repair and outlier removal processing to generate a basic dataset of seawater carbon dioxide concentration.

[0098] Marine network construction module: Based on the basic dataset of seawater carbon dioxide concentration and target sea area data, it completes the delineation of ocean basin zoning boundaries, verification of the first and last nodes of ocean current channels, verification of adjacent positions of air-sea exchange interfaces, identification of transmission orientation, deletion of breakpoints and writing of edge weights, and establishes a propagation topology grid.

[0099] Propagation constraint module: Based on the propagation topology grid, align the concentration sequences of upstream and downstream nodes, register the time difference, direction, and concentration difference, combine the graph attention network to assign adjacency weights, distinguish between passing sections and stopping sections, and obtain the propagation constraint sequence;

[0100] Anomaly clustering module: Based on the propagation constraint sequence, it cuts out continuous time window concentration segments, registers the start and end values, peak and valley positions, amplitude and duration, and combines isolated forest to divide the segment hierarchy, and clusters abrupt jump segments, gradual rise segments and repetitive jump segments to obtain anomaly state clusters;

[0101] Trajectory extrapolation module: Based on the abnormal state cluster and propagation constraint sequence, it loads short window sequence and long window sequence, selects high proportion trajectory, expands the concentration sequence for future time period, backfills residual fragments and corrects peak shift, and obtains concentration evolution trajectory;

[0102] Early warning generation module: Based on the concentration evolution trajectory and abnormal state clusters, it locks the threshold-exceeding nodes, accumulates continuous over-limit periods, traces back the diffusion links and boundary nodes, splices up regional connected pieces, and generates regional early warning units.

[0103] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence, characterized in that, Includes the following steps: S1: Acquire seawater carbon dioxide concentration monitoring data and auxiliary parameter data related to changes in seawater carbon dioxide concentration in the target sea area. The auxiliary parameter data includes water temperature, salinity, pH value, total alkalinity, ocean current velocity, and ocean current direction. Perform time alignment, spatial registration, missing value repair, and outlier removal on the acquired data to generate a basic dataset of seawater carbon dioxide concentration. S2: Based on the aforementioned seawater carbon dioxide concentration dataset and target sea area range data, delineate the boundaries of the ocean basin partition, verify the beginning and end nodes of the ocean current channel, check the adjacent positions of the air-sea exchange interface, identify the transmission direction of carbon sink nodes and carbon source nodes, delete transmission breakpoints, write the edge weight level, and establish a propagation topology grid. S3: Based on the propagation topology grid, align the seawater carbon dioxide concentration sequences of upstream and downstream nodes, use a graph attention network to compare the arrival order of ocean current fronts, filter out the upwelling band passage chains, delineate the dwelling time of stagnant water masses, identify strong propagation chains and slow propagation chains, and obtain the propagation constraint sequence. S4: Based on the propagation constraint sequence, cut out continuous time window segments of seawater carbon dioxide concentration, use isolated forests, arrange the increase order, separate the falling segment and the rising segment, lock the sparse segment, and group the sudden jump segment, the gradual rise segment and the reciprocating jump segment to obtain the abnormal state cluster. S5: Based on the anomalous state cluster and propagation constraint sequence, load the short window sequence and long window sequence, select the high proportion trajectory, expand the seawater carbon dioxide concentration sequence for future periods, backfill the repeated residual segments, correct the peak shift segments, and obtain the seawater carbon dioxide concentration evolution trajectory. S6: Based on the evolution trajectory of seawater carbon dioxide concentration and the cluster of abnormal states, lock the threshold nodes, rank the upstream contribution ratios, accumulate the continuous over-limit periods, trace back the diffusion links and boundary nodes, splice the regional connectivity pieces, and generate graded early warning results by combining the seawater carbon dioxide concentration threshold, concentration growth rate, continuous over-limit duration and regional diffusion range.

2. The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to claim 1, characterized in that, The basic dataset of seawater carbon dioxide concentration includes seawater carbon dioxide concentration monitoring values, time stamps, spatial location stamps, and auxiliary parameter stamps. The propagation topology grid includes basin partitioning units, unidirectional transmission links, and edge weight hierarchy stamps. The propagation constraint sequence includes propagation order, propagation strength classification, and dwell time period stamps. The abnormal state clusters include abrupt jump state segments, gradual rise state segments, and reciprocating jump state segments. The seawater carbon dioxide concentration evolution trajectory includes short window change trajectory, long window change trajectory, and offset correction trajectory. The graded early warning results include the set of over-threshold nodes, diffusion link set, and boundary node set corresponding to different early warning levels.

3. The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to claim 1, characterized in that, The specific steps for generating the propagation topology mesh are as follows: Based on the aforementioned seawater carbon dioxide concentration dataset and target sea area range data, the boundaries of the ocean basin are delineated, the positions of the first and last nodes of the ocean current channel are checked segment by segment, the correspondence between adjacent positions of the air-sea exchange interface is verified, the transmission orientation of carbon sink nodes and carbon source nodes is identified, the reachable direction, disconnection position and adjacent order are recorded, and the bottom table of the connection edge is obtained. Based on the aforementioned edge table, delete the edges corresponding to transmission breakpoints, write high transmission level, medium transmission level and low transmission level, merge continuous links in the same direction, eliminate reverse conflict links, correct the order of adjacent nodes and boundary connection relationship, and establish a propagation topology mesh.

4. The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to claim 1, characterized in that, The specific steps for generating the propagation constraint sequence are as follows: Based on the propagation topology grid, the seawater carbon dioxide concentration sequences of upstream and downstream nodes are aligned, and nodes are matched and connected according to adjacent sampling times. Node pairs with missing values ​​and reverse misaligned node pairs are removed. The node number, time difference number, connection direction, concentration difference, and number of consecutive occurrences are recorded to obtain the propagation time series table. Based on the propagation timeline, a graph attention network is used to compare the arrival order of ocean current fronts, filter out continuous upwelling chains, divide the stationary water mass stationary time period according to the stationary duration threshold, and list continuous passage sections, intermittent transition sections and stationary sections. Write the start and end positions of the sections, the duration, and the correspondence between adjacent nodes to obtain the link segmentation table. Based on the link segmentation table, the link order is determined according to edge weight level, arrival order difference, dwell time ratio, and continuous passage length. Strong propagation links and slow propagation links are identified, and node transmission restrictions, segment passage restrictions, time period restrictions, and direction restrictions are written to obtain the propagation constraint sequence.

5. The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to claim 4, characterized in that, The graph attention network first performs an adjacency node traversal on the seawater carbon dioxide concentration sequence corresponding to each node, reading the upstream node number, downstream node number, edge weight level, time difference sequence number, and concentration difference. It then performs time-by-time matching calculations between the seawater carbon dioxide concentration sequences of the upstream nodes and the current node to obtain multiple sets of node pair influence values. These influence values ​​are then normalized to form a weight distribution. The weight distribution is then weighted and summed with the corresponding upstream node concentration values ​​to generate node update values. Finally, these node update values ​​are processed in parallel across multiple channels, recording the influence values ​​of nearby nodes, delayed nodes, and low-weight nodes. The multi-channel calculation results are then concatenated and rearranged to form a set of node propagation influence values, which are then written into the link segmentation calculation process.

6. The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to claim 1, characterized in that, The specific steps for generating the abnormal state cluster are as follows: Based on the propagation constraint sequence, continuous time window segments of seawater carbon dioxide concentration are cut out, and the segments are rearranged according to the window length and the sliding step size. The starting value, ending value, peak position, valley position, change range, duration and interval between adjacent segments of each segment are recorded to obtain the window segment table. Based on the window segment table, an isolated forest is used to arrange the increase order, separate the falling segment and the rising segment, lock the sparse segment according to the interval between adjacent segments, the frequency of recurrence, and the degree of amplitude dispersion, merge the continuous segments in the same direction and segments with similar amplitude, write the segment arrangement relationship and concentration position, and obtain the morphological distribution table. Based on the morphological distribution table, the segments are classified according to the increase span, duration, number of round trips, and turning interval. Sudden jump segments, gradual rise segments, and reciprocating jump segments are grouped together, and the segment affiliation, position index, adjacency relationship, and aggregation range are written to obtain the abnormal state cluster.

7. The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to claim 6, characterized in that, The isolated forest process begins by reading the starting point, ending point, peak position, valley position, change range, duration, and interval between adjacent segments for each window segment, forming a sample vector. Multiple sets of segment records and multiple fields are randomly selected from the sample vector. Splitting conditions are set layer by layer according to the field value range. All sample vectors are repeatedly divided, recording the number of layers each window segment experiences before falling into a leaf node. The corresponding layers of multiple trees are summarized, and the average layer is calculated. Window segments are arranged from smallest to largest based on the average layer. Segments with smaller average layers are assigned to the easily isolated segment set, segments with medium average layers are assigned to the undecided segment set, and segments with larger average layers are assigned to the normal segment set. Then, combined with the increase value sorting results, upward increase records in the easily isolated segment set are separated and written into the rising segment, and downward increase records in the easily isolated segment set are separated and written into the falling segment. Sparse segments are marked according to the interval between adjacent segments, frequency of repetition, and degree of amplitude dispersion, and written into the segment arrangement relationship and concentration position to obtain a morphological distribution table.

8. The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to claim 1, characterized in that, The specific steps for generating the concentration evolution trajectory are as follows: Based on the aforementioned abnormal state clusters and propagation constraint sequences, short window sequences and long window sequences are loaded, adjacent seawater carbon dioxide concentration change segments are aligned according to time period positions, the proportions of multiple candidate trajectories are compared, and trajectory segments with high proportions, continuous fluctuations, and concentrated turning points are screened out. The seawater carbon dioxide concentration arrangement for future time periods is then expanded to obtain the trajectory bottom sequence. Based on the trajectory base sequence, repeated residual segments are backfilled segment by segment, the amplitude corresponding to the peak offset segment is corrected, the slope changes of adjacent time periods are aligned, the interval between abrupt jump intervals and lag intervals is compressed, and the connection sequence and turning point of seawater carbon dioxide concentration in future time periods are rearranged to obtain the evolution trajectory of seawater carbon dioxide concentration.

9. The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to claim 1, characterized in that, The specific steps for generating the aforementioned regional early warning unit are as follows: Based on the evolution trajectory of seawater carbon dioxide concentration and the cluster of abnormal states, the nodes exceeding the threshold are identified, and the contribution ratios of upstream sources are ranked one by one. The continuous periods of exceeding the limit are accumulated, and the relationship between the starting point, duration, and fallback position of exceeding the limit and the abnormal cluster is recorded. The diffusion conditions and time distribution of the nodes are summarized to obtain a diffusion criterion table. Based on the aforementioned diffusion criterion table, the diffusion link is traced back along the high-proportion direction to lock boundary nodes and turning points. Adjacent regions are connected, and diffusion segments from the same source are merged. The warning levels are divided according to the seawater carbon dioxide concentration threshold, concentration growth rate, duration of continuous exceedance, and regional diffusion range. The correspondence between the warning level, trigger time period, and coverage area is written to generate a graded warning result.

10. A seawater carbon dioxide concentration prediction and early warning system based on artificial intelligence, characterized in that, The method for predicting and warning seawater carbon dioxide concentration based on artificial intelligence according to any one of claims 1-9, wherein the system comprises: Data acquisition and processing module: used to acquire seawater carbon dioxide concentration monitoring data of the target sea area and auxiliary parameter data related to changes in seawater carbon dioxide concentration, and perform time alignment, spatial registration, missing value repair and outlier removal processing to generate a basic dataset of seawater carbon dioxide concentration. Marine network construction module: Based on the aforementioned seawater carbon dioxide concentration dataset and target sea area range data, completes the delineation of ocean basin zoning boundaries, verification of the beginning and end nodes of ocean current channels, verification of adjacent positions of air-sea exchange interfaces, identification of transmission orientation, deletion of breakpoints and writing of edge weights, and establishes a propagation topology mesh. Propagation constraint module: Based on the propagation topology grid, align the seawater carbon dioxide concentration sequences of upstream and downstream nodes, register the time difference, direction, and concentration difference, and combine the graph attention network to allocate adjacency weights to distinguish between passage sections and stopping sections, thereby obtaining the propagation constraint sequence; Anomaly Clustering Module: Based on the propagation constraint sequence, it cuts out continuous time window segments of seawater carbon dioxide concentration, registers the start and end values, peak and valley positions, amplitude and duration, divides the segments into hierarchical levels in combination with isolated forest, and clusters abrupt jump segments, gradual rise segments and repetitive jump segments to obtain anomaly state clusters; Trajectory extrapolation module: Based on the anomalous state cluster and propagation constraint sequence, short window sequence and long window sequence are loaded, high proportion trajectory is selected, seawater carbon dioxide concentration sequence for future time period is expanded, residual fragments are backfilled and peak shift is corrected to obtain the evolution trajectory of seawater carbon dioxide concentration. Early warning classification module: Based on the evolution trajectory of seawater carbon dioxide concentration and abnormal state clusters, it locks the threshold-exceeding nodes, accumulates the continuous over-limit period, traces back the diffusion link and boundary nodes, splices the regional connectivity pieces, and generates classified early warning results by combining the seawater carbon dioxide concentration threshold, concentration growth rate, continuous over-limit duration and regional diffusion range.

Images