A multi-step risk warning method and system for sea wind wave based on threshold consistency

Abstract

The application discloses a kind of offshore wind wave multi-step risk early warning method and system based on threshold consistency, it is related to marine environment intelligent prediction technical field, the present application is converted into wind grade label and wave grade label with order relationship by the continuous 10 m wind speed and significant wave height according to physical grade rule, constructs wind wave ordered threshold evidence;Utilize monotone ordered constraint to guarantee the logical consistency between different grade threshold exceedance probability, and based on wind, wave grade risk threshold exceedance probability deduces final risk probability, to output future multiple prediction time offshore wind wave risk probability graph, binary risk early warning graph and risk boundary uncertainty graph.The present application can not only predict whether future sea area grid enters risk state, but also make risk probability output and wind wave safety threshold trigger logic keep consistent, provide technical support for short-time multi-step marine wind wave risk identification, early warning display and auxiliary decision-making.

Description

Technical Field

[0001] This invention relates to the fields of intelligent prediction of marine environment, early warning of maritime navigation safety, risk assessment of marine engineering and artificial intelligence, and in particular to a multi-step risk early warning method and system for sea winds and waves based on threshold consistency. Background Technology

[0002] Activities such as maritime navigation, nearshore engineering operations, offshore wind power operation and maintenance, port scheduling, and marine disaster prevention and control are all highly sensitive to short-term wind and wave conditions. Strong winds, high waves, and their combined effects can significantly increase the difficulty of ship maneuvering, reduce navigation stability, and may lead to course deviations, operational interruptions, equipment damage, or even maritime safety accidents. Therefore, in maritime safety decision-making, managers typically focus not only on the specific values ​​of future wind speeds or wave heights, but also on whether future sea conditions will exceed preset safety thresholds, and whether the risk area is likely to persist, expand, shrink, or shift spatially in the near future.

[0003] Existing ocean wind and wave prediction technologies mainly include numerical prediction methods, statistical prediction methods, and data-driven deep learning methods. Numerical prediction methods are typically based on ocean dynamic equations and wave spectrum models, capable of describing physical processes such as wind and wave generation, propagation, and dissipation, and have a relatively clear physical foundation. However, these methods usually rely on complex initial boundary conditions, physical parameterization schemes, and grid integration calculations, resulting in high computational costs. Furthermore, they are easily affected by topography, coastlines, land-sea boundaries, and local wind field disturbances in complex nearshore waters and scenarios with rapidly changing local wind and waves.

[0004] Traditional statistical forecasting methods typically establish empirical relationships between wind and wave variables based on historical observation sequences. While these methods offer advantages such as simple structure and low computational cost, their expressive power is limited, making it difficult to fully characterize the nonlinear, multi-scale, and strongly spatiotemporally coupled features of ocean wind and wave fields. In recent years, deep learning methods such as convolutional neural networks, recurrent neural networks, ConvLSTM, and U-Net-like spatiotemporal models have been widely used for marine environmental forecasting. These methods can automatically learn the spatial structure and temporal evolution of wind and wave fields from large-scale historical data, improving forecasting efficiency and accuracy.

[0005] However, existing data-driven methods for wind and wave forecasting and risk warning still have the following shortcomings.

[0006] First, most existing methods use continuous wind speed, wave height, or sea state level as the prediction target, while actual navigation and maritime operation safety management focuses more on whether future sea states will exceed safety thresholds. Simply predicting continuous variables and then making threshold judgments can easily lead to inconsistencies between the prediction target and the safety decision-making target, making it difficult to directly output the risk probability for operational decisions.

[0007] Second, some methods directly model maritime risk warning as an independent binary classification problem, that is, directly predicting whether a certain grid point is a risk grid point. Although this approach can obtain the risk probability, there is no explicit link between the risk probability and the evidence of wind level thresholds and wave level thresholds. This makes it difficult to trace the risk output back to the corresponding wind and wave safety threshold triggering basis, reducing the logical consistency and interpretability of the warning results.

[0008] Third, wind and wave levels inherently possess an ordered nature. For any grid point in a sea area, the probability of exceeding a higher wind level threshold should not be greater than the probability of exceeding a lower wind level threshold; similarly, the probability of exceeding a higher wave level threshold should not be greater than the probability of exceeding a lower wave level threshold. If the model treats wind and wave levels as a general, unordered classification task, a threshold crossing problem may occur where the probability of exceeding a higher level is greater than the probability of exceeding a lower level, causing the risk derivation logic to contradict physical common sense and the rules of level order.

[0009] Fourth, maritime risk management is usually not a decision-making action at a single moment. Ships slowing down, turning, changing course, avoiding dangerous waters, or suspending maritime operations all require a certain amount of lead time. Single-step risk warnings can only reflect the risk status at a specific future moment, while multi-step risk warnings can show the continuity, expansion, shrinkage, or spatial movement trend of the risk area over the next few hours, which is more in line with actual maritime safety management needs.

[0010] Fifth, existing methods often focus more on prediction accuracy, while paying insufficient attention to probabilistic reliability, consistency of risk thresholds, and uncertainty of risk boundaries. For maritime safety early warning, the probabilistic output should not only be "accurate in prediction," but also have good probabilistic calibrability and interpretability. At the same time, for spatial areas where the risk probability is close to the judgment boundary, uncertainty prompts should also be provided to assist in model error analysis, result display, and manual verification.

[0011] Therefore, it is necessary to propose a new short-term multi-step risk early warning method for the sea, which can simultaneously achieve wind and wave physical level binning, monotonically ordered threshold exceedance probability modeling, threshold consistency risk derivation, multi-step risk probability map output, and risk boundary uncertainty expression, thereby improving the accuracy, reliability, logical consistency, and traceability of sea wind and wave risk early warning. Summary of the Invention

[0012] This invention addresses the problems existing in the prior art by proposing a multi-step risk early warning method and system for marine wind and waves based on threshold consistency.

[0013] The technical means employed in this invention are as follows: A multi-step risk early warning method for marine wind and waves based on threshold consistency includes the following steps: Step 1: Obtain marine meteorological reanalysis data of the target sea area and perform preprocessing to obtain historical multivariate marine meteorological state sequences; Step 2: Construct an effective marine grid mask for the target sea area based on the land-sea distribution of the target sea area; Step 3: Based on the standard ocean description level, convert continuous wind speed and continuous significant wave height into wind level labels and wave level labels with physical meaning and grade order, respectively; Step 4: Based on the wind level label and the wave level label, construct wind and wave risk labels on the effective ocean grid points according to the preset wind level risk threshold and wave level risk threshold; Step 5: Based on the historical multivariate marine meteorological state sequence, the effective marine grid mask, the wind level label, the wave level label, and the wind and wave risk label, construct short-term multi-step prediction samples using a sliding window method. The short-term multi-step prediction samples include a historical input window and a future prediction window. Step 6: Construct a temporal jump-connection fusion ordered regression spatiotemporal model, and use the temporal jump-connection fusion ordered regression spatiotemporal model to process the short-time multi-step prediction sample data input into the historical input window of the short-time multi-step prediction sample, extract the multi-scale spatial structure features and temporal evolution features of the historical wind and wave field, and then obtain the shared spatiotemporal features for wind and wave ordered evidence output and risk probability derivation. Step 7: Based on the shared spatiotemporal features, construct the wind level ordered evidence output head and the wave level ordered evidence output head, and output the wind level threshold exceedance probability and wave level threshold exceedance probability for multiple future prediction times. Step 8: Apply monotonically ordered constraints to the output wind level threshold exceedance probability and wave level threshold exceedance probability respectively to obtain wind and wave ordered evidence; Step 9: Obtain wind triggering evidence and wave triggering evidence corresponding to the wind level risk threshold and the wave level risk threshold from the ordered wind and wave evidence; Step 10: Based on the wind-triggered evidence, the wave-triggered evidence, the wind level risk threshold, and the wave level risk threshold, a multi-step risk probability map is calculated using the threshold consistency rule; Step 11: Based on the multi-step risk probability map, generate a binary risk warning map according to the preset risk probability judgment threshold, and calculate the risk boundary uncertainty map; Step 12: For the temporal jump-connection fusion ordered regression spatiotemporal model, construct a multi-task loss function and optimize the model parameters, and train and validate the temporal jump-connection fusion ordered regression spatiotemporal model based on the short-time multi-step prediction samples to obtain the model with the best performance on the validation set. Step 13: Based on the model with the best performance on the validation set, perform predictive reasoning on the latest historical multivariate marine meteorological field of the target sea area to generate short-term multi-step marine risk warnings. The short-term multi-step marine risk warnings include the ordered threshold exceedance probability of wind and waves at multiple future prediction times, risk probability maps, binary risk warning maps, and risk boundary uncertainty maps.

[0014] The multi-step risk warning system for marine wind and waves based on threshold consistency disclosed in this application includes: a marine meteorological data acquisition and data preprocessing module, used to acquire marine meteorological reanalysis data of the target sea area and perform preprocessing to obtain historical multivariate marine meteorological state sequences; An effective ocean grid mask module is used to construct an effective ocean grid mask for a target ocean area based on the land-sea distribution of the target ocean area; The wind and wave rating sub-module is used to convert continuous wind speed and continuous significant wave height into wind rating labels and wave rating labels with physical meaning and rating order, respectively, according to the standard ocean description rating. The wind and wave risk label construction module is used to construct wind and wave risk labels on the effective ocean grid points based on the wind level label and the wave level label, according to the preset wind level risk threshold and wave level risk threshold. The short-time multi-step sample construction module is used to construct short-time multi-step prediction samples based on the historical multivariate marine meteorological state sequence, the effective marine grid mask, the wind level label, the wave level label, and the wind and wave risk label, using a sliding window method. The short-time multi-step prediction samples include a historical input window and a future prediction window. The shared spatiotemporal feature extraction module is used to construct a temporal jump-connection fusion ordered regression spatiotemporal model. Based on the short-time multi-step prediction sample data input through the historical input window of the short-time multi-step prediction sample, the module processes the data using the temporal jump-connection fusion ordered regression spatiotemporal model to extract the multi-scale spatial structure features and temporal evolution features of the historical wind and wave field, thereby obtaining shared spatiotemporal features for wind and wave ordered evidence output and risk probability derivation. The wind and wave ordered evidence output module is used to construct wind level ordered evidence output header and wave level ordered evidence output header based on the shared spatiotemporal features, and output the wind level threshold exceedance probability and wave level threshold exceedance probability at multiple future prediction times. The monotonic order constraint module is used to apply monotonic order constraints to the output wind level threshold exceedance probability and wave level threshold exceedance probability to obtain evidence of wind and wave order. The wind and wave risk threshold evidence extraction module is used to extract wind triggering evidence and wave triggering evidence corresponding to the wind level risk threshold and the wave level risk threshold from the ordered wind and wave evidence. The threshold consistency risk probability derivation module is used to calculate a multi-step risk probability map based on the wind triggering evidence, the wave triggering evidence, the wind level risk threshold, and the wave level risk threshold, using the threshold consistency rule. The binary risk warning map and risk boundary uncertainty generation module is used to generate a binary risk warning map based on the multi-step risk probability map and according to a preset risk probability judgment threshold, and to calculate the risk boundary uncertainty map. The model training and optimization module is used to construct a multi-task loss function and optimize the model parameters for the temporal jump-connection fusion ordered regression spatiotemporal model, and to train and validate the temporal jump-connection fusion ordered regression spatiotemporal model based on the short-time multi-step prediction samples to obtain the model with the best performance on the validation set. The short-term multi-step risk output module is used to predict and infer the latest historical multivariate marine meteorological field of the target sea area based on the model with the best performance of the validation set, and generate short-term multi-step risk warnings for the sea. The short-term multi-step risk warnings for the sea include the ordered threshold exceedance probability of wind and waves at multiple future prediction times, risk probability map, binary risk warning map and risk boundary uncertainty map.

[0015] Compared with existing technologies, the multi-step risk early warning method and system for marine wind and waves based on threshold consistency disclosed in this invention has the following beneficial effects: This invention enables marine wind and wave risk early warning results to simultaneously possess threshold consistency, probabilistic expression capability, spatial continuity, and traceability by classifying wind and wave physical levels, modeling ordered threshold exceedance probability, applying monotonically ordered constraints, deriving threshold consistency risk probability, outputting short-term multi-step risk information, and calculating risk boundary uncertainty. This provides effective technical support for short-term multi-step marine wind and wave risk identification and early warning display. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-step risk early warning method for marine wind and waves based on threshold consistency disclosed in this invention; Figure 2 This is a schematic diagram of the effective ocean grid mask on the target sea area prediction grid in the early warning method disclosed in this invention; Figure 3 This is a distribution diagram of the wind level, wave level, and threshold risk sample composition in the early warning method disclosed in this invention; Figure 4 This is an architecture diagram of the temporal jump-connection fusion ordered regression spatiotemporal model in the early warning method disclosed in this invention; Figure 5 This is a decomposition diagram of the risk warning performance and wind and wave trigger set under different forecast lead times in the early warning method disclosed in this invention; Figure 6 This is a reliability analysis diagram of the risk probability output in the early warning method disclosed in this invention; Figure 7 This is a diagram showing the experimental results of model ablation in the early warning method disclosed in this invention; Figure 8 The early warning method disclosed in this invention provides a typical risk event forecast timeframe. The following is a spatial visualization result. Detailed Implementation

[0017] like Figure 1 As shown, the multi-step risk early warning method for marine wind and waves based on threshold consistency disclosed in this invention includes the following steps: Step 1: Obtain marine meteorological reanalysis data of the target sea area and perform preprocessing to obtain historical multivariate marine meteorological state sequences; Step 2: Construct an effective marine grid mask for the target sea area based on the land-sea distribution of the target sea area; Step 3: Based on the standard ocean description level, convert continuous wind speed and continuous significant wave height into wind level labels and wave level labels with physical meaning and grade order, respectively; Step 4: Based on the wind level label and the wave level label, construct wind and wave risk labels on the effective ocean grid points according to the preset wind level risk threshold and wave level risk threshold; Step 5: Based on the historical multivariate marine meteorological state sequence, the effective marine grid mask, the wind level label, the wave level label, and the wind and wave risk label, construct short-term multi-step prediction samples using a sliding window method. The short-term multi-step prediction samples include a historical input window and a future prediction window. Step 6: Construct a temporal jump-connection fusion ordered regression spatiotemporal model, and use the temporal jump-connection fusion ordered regression spatiotemporal model to process the short-time multi-step prediction sample data input into the historical input window of the short-time multi-step prediction sample, extract the multi-scale spatial structure features and temporal evolution features of the historical wind and wave field, and then obtain the shared spatiotemporal features for wind and wave ordered evidence output and risk probability derivation. Step 7: Based on the shared spatiotemporal features, construct the wind level ordered evidence output head and the wave level ordered evidence output head, and output the wind level threshold exceedance probability and wave level threshold exceedance probability for multiple future prediction times. Step 8: Apply monotonically ordered constraints to the output wind level threshold exceedance probability and wave level threshold exceedance probability respectively to obtain wind and wave ordered evidence; Step 9: Obtain wind triggering evidence and wave triggering evidence corresponding to the wind level risk threshold and the wave level risk threshold from the ordered wind and wave evidence; Step 10: Based on the wind-triggered evidence, the wave-triggered evidence, the wind level risk threshold, and the wave level risk threshold, a multi-step risk probability map is calculated using the threshold consistency rule; Step 11: Based on the multi-step risk probability map, generate a binary risk warning map according to the preset risk probability judgment threshold, and calculate the risk boundary uncertainty map; Step 12: For the temporal jump-connection fusion ordered regression spatiotemporal model, construct a multi-task loss function and optimize the model parameters, and train and validate the temporal jump-connection fusion ordered regression spatiotemporal model based on the short-time multi-step prediction samples to obtain the model with the best performance on the validation set. Step 13: Based on the model with the best performance on the validation set, perform predictive reasoning on the latest historical multivariate marine meteorological field of the target sea area to generate short-term multi-step marine risk warnings. The short-term multi-step marine risk warnings include the ordered threshold exceedance probability of wind and waves at multiple future prediction times, risk probability maps, binary risk warning maps, and risk boundary uncertainty maps.

[0018] The multi-step risk early warning method for marine wind and waves based on threshold consistency disclosed in this invention has the following beneficial effects: This invention enables marine wind and wave risk early warning results to simultaneously possess threshold consistency, probabilistic expression ability, spatial continuity and traceability by classifying wind and wave physical levels, modeling ordered threshold exceedance probability, applying monotonically ordered constraints, deriving threshold consistency risk probability, outputting short-term multi-step risk and calculating risk boundary uncertainty. This provides effective technical support for short-term multi-step marine wind and wave risk identification and early warning display.

[0019] The following specific embodiments illustrate each step of the multi-step risk warning method for marine wind and waves based on threshold consistency disclosed in this application: Step 1: Obtain marine meteorological reanalysis data of the target sea area and perform preprocessing to obtain historical multivariate marine meteorological state sequences; Marine meteorological reanalysis data for the target sea area is acquired, a unified prediction grid corresponding to the target sea area is determined, and the marine meteorological reanalysis data is preprocessed including time alignment, spatial alignment, variable standardization, wave direction variable transformation, and land-sea attribute labeling to obtain a historical multivariate marine meteorological state sequence. In one specific embodiment, the marine meteorological reanalysis data uses ERA 5-hour reanalysis data, with a time range from 2015 to 2021 and a spatial resolution of 0.25°. The target nearshore sea area is selected as the study area, and the data is truncated according to the latitude and longitude range of the target nearshore sea area to form a data set of size [missing information]. The prediction grid.

[0020] Marine meteorological variables include the zonal wind component at 10 m altitude, the meridional wind component at 10 m altitude, 10 m wind speed, significant wave height, mean wave period, wave direction, and sea level pressure. The zonal and meridional wind components at 10 m altitude are used to characterize the east-west and north-south components of the near-shore wind field, respectively; 10 m wind speed is used to construct wind class labels; significant wave height is used to construct wave class labels; and mean wave period, wave direction, and sea level pressure provide marine meteorological background information related to wind and wave evolution.

[0021] To avoid wave direction angle variables in and If a numerical discontinuity occurs nearby, the wave direction is converted into a wave direction sine component and a wave direction cosine component:

[0022] In the formula, This represents the original wave direction angle variable, in degrees. Indicates the sinusoidal component of the wave direction; This represents the wave direction cosine component. This transformation avoids discontinuous jumps in the angle variable at the circular boundary.

[0023] Therefore, the input variables for each spatial grid point at any historical moment include the zonal wind component at 10 m altitude, the meridional wind component at 10 m altitude, the wind speed at 10 m, the significant wave height, the average wave period, the sine component of the wave direction, the cosine component of the wave direction, and the sea level pressure.

[0024] Let the preprocessed historical multivariate marine meteorological state sequence constitute the model input tensor. :

[0025] In the formula, Indicates the input spacetime tensor; Represents the space of real numbers; Indicates the number of samples in the batch; Indicates the length of the historical input time step; Indicates the number of input variable channels; This indicates the number of grid points in the spatial grid along the latitudinal or altitude direction; This indicates the number of grid points in the spatial grid along the longitudinal or transverse direction.

[0026] In some embodiments, the historical input time step length is set to That is, the multivariate marine meteorological field of the past 12 hours is used as the model input; the number of input variable channels is set to The data corresponds to the zonal wind component at 10 m altitude, the meridional wind component at 10 m altitude, the wind speed at 10 m, the significant wave height, the average wave period, the sine component of the wave direction, the cosine component of the wave direction, and the sea level pressure. Each sample is used to generate a wind and wave risk probability map for multiple forecast lead times in the future.

[0027] Step 2: Construct an effective marine grid mask for the target sea area based on the land-sea distribution of the target sea area; Specifically, to avoid interference from land areas in subsequent model training, loss calculation, risk probability output, and evaluation index calculation, this invention constructs an effective ocean grid mask based on the land-sea distribution of the target sea area. The effective ocean grid mask can be as follows: Figure 2 As shown, effective ocean grid points participate in subsequent model training, risk inference, and result evaluation, while land grid points or grid points outside the target sea area do not participate in subsequent calculations.

[0028] Let the spatial grid size of the target sea area be... Define the effective set of ocean grid points as The effective ocean grid mask is:

[0029] In the formula, Indicates spatial location Effective ocean grid mask value at the location; Represents the row index of the spatial grid; The column index representing the spatial grid; Represents the effective set of ocean grid points; This indicates that the grid point is a valid ocean grid point; This indicates that the grid point is either a land grid point or a grid point outside the target sea area.

[0030] In a preferred embodiment, the effective ocean grid set Determined based on the land and sea shelters within the study area. For example... Figure 2 As shown, only valid ocean grid points are retained as the spatial range for model training, risk inference, and evaluation calculations; land grid points or invalid grid points are not involved in the above calculations.

[0031] In the subsequent processes of wind and wave level binning, risk label construction, sliding window sample construction, model training, loss function calculation, risk probability output, binary risk warning map generation, and evaluation index calculation, only those meeting the requirements are considered. Effective ocean grid points are processed. For Land grid points or grid points outside the target sea area can be retained as null values, masked values, or background areas that are not displayed in the model output and visualization results.

[0032] Step 3: Based on the standard ocean description level, convert continuous wind speed and continuous significant wave height into wind level labels and wave level labels with physical meaning and grade order, respectively. Specifically, let the wind speed at a certain time and a certain effective ocean grid point be 10 m. Significant wave height is This invention, based on standard oceanographic description levels, converts continuous wind speed and continuous significant wave height into wind level labels and wave level labels with physical meaning and hierarchical order, respectively. The wind level, wave level, and their sample distribution can be as follows: Figure 3 As shown.

[0033] In a preferred embodiment, wind speed is divided into 13 ordered wind levels from 0 to 12 according to the Beaufort scale, i.e., the number of wind level categories is [number missing]. The boundary set of wind speed levels is:

[0034] In the formula, This represents the boundary set of wind speed levels, where the units of each value in the set are m / s, and is used to classify wind speeds at 10 m. Convert to wind level label .

[0035] For any valid ocean grid point, wind level label Defined as:

[0036] In the formula, Indicates wind level label; This represents a wind speed of 10 m. (From the formula...) It can be seen that the wind speed range corresponding to wind level 6 is: .

[0037] In a preferred embodiment, significant wave heights are divided into 10 ordered wave classes from 0 to 9 according to sea state grades, i.e., the number of wave class categories is [number missing]. For any valid ocean grid point, wave level label Defined as:

[0038] In the formula, Indicates wave level label; Represents the significant wave height, in meters (m). From the formula... It can be seen that the significant wave height range corresponding to wave grade 5 is: In boundary value handling, samples that are exactly equal to the level boundary are classified into lower levels to ensure consistency in level label construction. For example, when wind speed... At that time, the grid point is still classified as wind level 5; when the wind speed... Only then is it classified as a wind level 6 or higher. Similarly, when there is a significant wave height... At that time, the grid point is still classified as wave level 4; when the wave height is significant... Only then will it be classified as wave level 5 or higher.

[0039] Through the aforementioned binning method, continuous wind speed and continuous significant wave height are converted into wind level labels and wave level labels with a clear hierarchical relationship, thereby providing structured wind and wave evidence for subsequent wind and wave ordered threshold exceedance probability modeling, risk threshold evidence extraction, and threshold consistency risk probability derivation.

[0040] Step 4: Based on the wind level label and the wave level label, construct wind and wave risk labels on the effective ocean grid points according to the preset wind level risk threshold and wave level risk threshold; Specifically, based on the wind level and wave level labels obtained in step 3, this invention constructs risk labels for ocean grid points according to preset wind level risk thresholds and preset wave level risk thresholds. For future... At any given prediction time, if the wind level of a valid ocean grid point reaches or exceeds a preset wind level risk threshold, or the wave level reaches or exceeds a preset wave level risk threshold, then that grid point is marked as being in a risky state; otherwise, it is marked as being in a non-risky state. The wind level, wave level, and risk samples triggered by wind, waves, and combined wind and wave events can be represented as follows: Figure 3 As shown, the figure includes the distribution of wind level samples, wave level samples, and the composition distribution of non-risk samples, wind-triggered risk samples, wave-triggered risk samples, and combined wind and wave-triggered risk samples in the training set, validation set, and test set. It is used to illustrate the unbalanced characteristics of wind and wave level samples and that risk events can be formed by strong winds, high waves, or both reaching the risk threshold.

[0041] Let the preset wind level risk threshold be... The preset wave level risk threshold is For the future Predicted time and spatial location Effective ocean grid points, their risk labels Defined as:

[0042] In the formula, Indicates the future number Predicted time and spatial location Risk label at the location; Indicates the wind level label for the corresponding grid point; Indicates the wave level label for the corresponding grid point; This indicates the preset wind level risk threshold; This indicates the preset wave level risk threshold; Indicates the prediction step index; and These represent the row index and column index of the spatial grid, respectively.

[0043] With the introduction of an effective marine grid mask, risk labels are only applied when conditions are met. Constructed on an effective ocean grid. For Land grid points or grid points outside the target sea area do not participate in risk label construction, model training, and evaluation index calculation.

[0044] In a preferred embodiment, the preset wind level risk threshold and the preset wave level risk threshold are respectively set as follows:

[0045] That is, when the wind force is not lower than level 6 or the wave force is not lower than level 5, the corresponding valid ocean grid points are marked as being in a risk state. Combining this with the preferred level boundary in step S3, the risk triggering condition can be further expressed as:

[0046] In the formula, Indicates the future number Predicted time and spatial location 10 m wind speed at the location; This indicates the significant wave height of the corresponding grid point.

[0047] Based on whether the wind and wave levels reach their respective risk thresholds, the risk samples can be further divided into wind-triggered risk samples, wave-triggered risk samples, and combined wind and wave-triggered risk samples. Among them, wind-triggered risk samples are defined as follows:

[0048] The risk sample for wave-triggered events is defined as follows:

[0049] The risk sample triggered by complex wind and waves is defined as follows:

[0050] In the formula, This indicates that only wind-triggered risk samples are labeled; This indicates that only wave-triggered risk sample labels are used; This indicates the risk sample labeling triggered by combined wind and waves. Through the above classification, this invention can not only construct binary risk labels, but also distinguish the wind-triggered, wave-triggered, and combined wind and wave-triggered components of risk events in experimental statistics and result analysis.

[0051] The above , To optimize risk threshold settings, in practical applications or threshold sensitivity analysis, the corresponding risk labels can be reconstructed or the risk probabilities can be re-derived given other wind and wave risk thresholds. The specific wind and wave risk threshold to be used should be determined separately based on the vessel conditions, operational tasks, marine environment, or relevant safety regulations.

[0052] Step 5: Based on the historical multivariate marine meteorological state sequence, the effective marine grid mask, the wind level label, the wave level label, and the wind and wave risk label, construct short-term multi-step prediction samples using a sliding window method. The short-term multi-step prediction samples include a historical input window and a future prediction window. Based on the historical multivariate marine meteorological state sequences, effective marine grid masks, wind level labels, wave level labels, and risk labels obtained in steps 1 to 4, this invention constructs short-term multi-step prediction samples using a sliding window approach. Each sample consists of a historical input window and a future prediction window, used to learn the spatiotemporal evolution information of the wind and wave field over a past period and predict the wind and wave risk status for multiple forecast periods in the future.

[0053] Let the time of the sample anchor point be... The historical input time step length is The number of future prediction steps is The history input window is defined as follows:

[0054] In the formula, Indicated by anchor point time This is the historical input window for the end time. This represents the multivariate marine meteorological field at the first moment in the historical input window; Represents a multivariable marine meteorological field at the anchor point; This indicates the length of the historical input time step.

[0055] The future forecast window is defined as:

[0056] In the formula, Indicated by anchor point time A future forecast window starting from; Indicates the future number Supervision labels for each predicted time point; This represents the number of future prediction steps, where .

[0057] For the future Predicted moment, supervisory label Including wind rating labels, wave rating labels, and risk labels, it can be represented as:

[0058] In the formula, Indicates the future number Wind level label for each predicted moment; Indicates the future number Wave grade label for each predicted moment; Indicates the future number Risk labels for each predicted time.

[0059] For spatial grid location , the future No. The supervision labels for each prediction time point can be further represented as:

[0060] In the formula, Indicates the future number Predicted time and spatial location Supervision label at the location; , and These represent the wind level label, wave level label, and risk label for the corresponding grid point, respectively. Only when... When the grid point's label is used, it participates in subsequent model training, loss calculation, and evaluation metric calculation; when At that time, the grid point is masked as either a land grid point or a grid point outside the target sea area.

[0061] Therefore, the first Each training sample can be represented as:

[0062] In the formula, Indicates the first A short-time multi-step prediction sample; Indicates the first The anchor point time corresponding to each sample; This represents the historical input window for this sample; This represents the future prediction window for the sample; This represents an effective gridded mask for the ocean.

[0063] The entire training sample set can be represented as:

[0064] In the formula, This represents the sample set constructed using a sliding window approach; Indicates the number of samples.

[0065] In a preferred embodiment, the historical input time step length is set to... This means that the multivariate marine meteorological field of the past 12 hours is used as the model input; the number of future prediction steps is set to... This outputs wind and wave risk probability maps for three forecast timeframes: 1 hour, 2 hours, and 3 hours. The historical input window and the future forecast window are as follows:

[0066] In the formula, This indicates a historical input window based on the marine meteorological field over the past 12 hours; This indicates the future forecast windows corresponding to the three forecast lead times of 1 hour, 2 hours, and 3 hours.

[0067] By constructing the above-mentioned sliding window sample, this invention can convert continuous hourly marine meteorological time series into supervised learning samples suitable for training spatiotemporal prediction networks, enabling the model to learn the spatiotemporal evolution of historical wind and wave fields and output risk probability maps for three forecast periods of 1 hour, 2 hours, and 3 hours, thereby providing a data foundation for short-term, multi-step marine wind and wave risk early warning.

[0068] Step 6: Construct a temporal jump-connection fusion ordered regression spatiotemporal model, and use the temporal jump-connection fusion ordered regression spatiotemporal model to process the short-time multi-step prediction sample data input into the historical input window of the short-time multi-step prediction sample, extract the multi-scale spatial structure features and temporal evolution features of the historical wind and wave field, and then obtain the shared spatiotemporal features for wind and wave ordered evidence output and risk probability derivation. Specifically, the historical input window constructed in step 5 The input temporal jump-connected fusion ordered regression spatiotemporal U-Net, i.e., TSF-OR-STUNet, extracts multi-scale spatial structure features and temporal evolution features of historical wind and wave fields, obtaining a shared spatiotemporal feature representation for subsequent wind and wave ordered evidence output and risk probability derivation. This process can be represented as:

[0069] In the formula, Indicates the history input window The extracted shared spatiotemporal feature representation; The parameter is The TSF-OR-STUNet spatiotemporal prediction network; This represents the learnable parameters in the network.

[0070] The overall structure of the temporal jump-connected fusion ordered regression spatiotemporal model is as follows: Figure 4As shown, TSF-OR-STUNet includes a shared two-dimensional encoder, a bottleneck time fusion module, a multi-scale time series jump connection fusion module, a spatial decoder, a wind level ordered evidence output head, a wave level ordered evidence output head, and a threshold consistency risk derivation module.

[0071] In a preferred embodiment, TSF-OR-STUNet first encodes the multivariate ocean meteorological field at each historical moment in the historical input window using a shared two-dimensional encoder to extract multi-scale spatial features. Then, it captures the temporal evolution of the low-resolution large-scale wind and wave field using a bottleneck time fusion module. A multi-scale temporal jump connection fusion module is introduced into the U-Net jump connection path to inject historical dynamic information into the jump connection features at different scales. Finally, the spatial resolution is restored using a spatial decoder to obtain a shared spatiotemporal feature representation. .

[0072] Specifically, the shared two-dimensional encoder includes: Each scale layer, for historical input window The Middle Multivariable marine meteorological field at a historical moment , No. The encoded features of each scale layer can be represented as:

[0073] In the formula, Indicates the first The historical moment at the Spatial coding features at various scales; Indicates the shared two-dimensional encoder in the first... Encoding functions for each scale layer; .

[0074] For all historical moments in the history input window Stacking the scale-encoded features yields the _th ... Features of skip-connected sequences at various scales:

[0075] In the formula, Indicated by anchor point time For the end time, the first Features of historical skip sequences at various scales; Indicates the first Number of channels per scale; and They represent the first Spatial height and width at various scales.

[0076] For the bottleneck layer of the lowest resolution shared 2D encoder, let its historical sequence characteristics be... This invention employs a bottleneck time fusion module to capture the temporal evolution of large-scale wind and wave fields; preferably, the bottleneck time fusion module uses a Convolutional Long Short-Term Memory (ConvLSTM) network, and its fusion process is represented as follows:

[0077] In the formula, This indicates the bottleneck characteristics after bottleneck time fusion. This represents a convolutional long short-term memory network placed at the encoder bottleneck layer. This module is used to model the overall temporal evolution trend of wind field, wave field, pressure field, and other marine meteorological variables within a historical window.

[0078] Traditional U-Net jump connection structures typically only transmit local spatial details from the most recent input moment, making it difficult to fully utilize information on historical risk region boundary movement, morphological changes, and spatial evolution. This invention introduces a temporal jump connection fusion module into the jump connection path at each scale, enabling high-resolution local spatial features to be combined with historical dynamic information, thereby improving the representation of risk region boundaries, morphology, and spatial continuity in multi-step risk warning.

[0079] For the Features of skip sequences at various scales Firstly, adopt Convolution performs channel reduction to decrease the computational cost of temporal modeling.

[0080] The number of channels after dimensionality reduction is defined as:

[0081] In the formula, Indicates the first Features of skip-connected sequences after dimensionality reduction of channels at various scales; Indicates the use of channel dimensionality reduction convolution; This represents the number of channels after dimensionality reduction; Indicates the minimum number of reserved channels; Indicates the channel compression ratio; This indicates rounding down to the nearest integer.

[0082] Then, ConvLSTM is applied to the dimensionality-reduced skip sequence features to obtain the 1st... Historical dynamic aggregation features at various scales:

[0083] In the formula, Indicates the first The temporal aggregated hidden state at the end of the historical input window at each scale; Indicates the first Convolutional Long Short-Term Memory (LSTM) networks in scale-hopping paths.

[0084] Next, the historical dynamic aggregated features are projected back to the original channel dimension to obtain the historical dynamic supplementary features:

[0085] In the formula, Indicates the first Historical dynamic supplementary features projected at various scales; Indicates the use of channel recovery convolution.

[0086] Let the first The jump connection feature at each scale is: ,in Indicates the first Anchor point time at each scale The local spatial features. To avoid excessive interference from historical dynamic information with current local details, this invention employs a gating fusion mechanism to adaptively fuse the most recent local spatial features and historical dynamic supplementary features. The gating weight is defined as follows:

[0087] The fusion of the first The scale-hopping feature is represented as follows:

[0088] In the formula, Indicates the first Each scale of gating weights; This represents the Sigmoid activation function; This indicates splicing along the channel dimension; This represents element-wise multiplication; Indicates the first step after time-sequence jump connection fusion Each scale jump connection feature.

[0089] Through the aforementioned temporal jump-connection fusion method, the model can adaptively introduce historical dynamic information while preserving the local spatial structure at the most recent moment, so that the risk area boundary, spatial form and continuous evolution trend in multi-step risk warning can be expressed more stably.

[0090] In the spatial decoding stage, this invention integrates bottleneck time features. Input spatial decoder, and fuse features with temporal skip connections at each scale during the progressive recovery of spatial resolution. By fusing the data, a shared spatiotemporal feature representation is obtained:

[0091] In the formula, Representation space decoder; This represents the shared spatiotemporal feature representation obtained after decoding. This shared spatiotemporal feature representation simultaneously contains large-scale temporal evolution information of historical wind and wave fields, multi-scale spatial structure information, and local risk area boundary information, which are used for subsequent ordered evidence output of wind level, ordered evidence output of wave level, and derivation of risk probability for threshold consistency.

[0092] In a preferred embodiment, the channel compression ratio of the timing skip-connection fusion module is set to... The minimum number of reserved channels is set to This setting allows for a reduction in the computational cost of ConvLSTM in multi-scale skip paths while preserving temporal dynamics.

[0093] Step 7: Based on the shared spatiotemporal features, construct the wind level ordered evidence output head and the wave level ordered evidence output head, and output the wind level threshold exceedance probability and wave level threshold exceedance probability for multiple future prediction times. Specifically, based on the shared spatiotemporal feature representation obtained in step 6 This invention constructs ordered evidence output heads for wind level and wave level, respectively, to output the probability of exceeding the wind level threshold and the probability of exceeding the wave level threshold at multiple future prediction times. The ordered evidence output head does not directly output the probability of each level category using a common unordered multi-classification method, but instead directly predicts the cumulative threshold exceeding probability of whether the wind level or wave level reaches or exceeds a certain level threshold, thus providing structured wind and wave evidence for the subsequent derivation of the threshold consistency risk probability.

[0094] Let the number of future prediction steps be... The number of wind level categories is The number of wave level categories is For the future... At each predicted time, the wind level threshold index is: Wave level threshold index is Since level 0 does not require setting a threshold for "reaching or exceeding level 0", therefore for Individual wind level categories, output The probability of exceeding the wind level threshold; for Each wave level category, output The probability of exceeding the threshold for wave level.

[0095] For the future At each prediction time, the wind level ordered evidence output head and the wave level ordered evidence output head are based on shared spatiotemporal feature representations. Generate the corresponding shared space response graph:

[0096] In the formula, Indicates the future number Shared spatial response map of wind levels at each predicted moment; Indicates the future number Wave level shared spatial response diagram at each predicted moment; The wind level indicates the orderly output of evidence in the future. The convolution mapping function corresponding to each prediction time step; The wave level indicates the ordered evidence output head in the future. The convolution mapping function corresponding to each prediction time step.

[0097] For spatial location The logit values ​​for wind and wave levels at their corresponding thresholds are defined as follows:

[0098] In the formula, Indicates the future number Predicted time and spatial location The wind level has reached or exceeded the threshold. The logit value; Indicates the future number Predicted time and spatial location The wave level has reached or exceeded the threshold. The logit value; Indicates the wind level in the future. The predicted time, the first Bias terms at each threshold; Indicates the wave level in the future. The predicted time, the first Bias terms at each threshold.

[0099] Based on the threshold logit of wind and wave levels, the exceedance probability of the wind level threshold and the exceedance probability of the wave level threshold are defined as follows:

[0100] In the formula, Indicates the future number Predicted time and spatial location The wind level has reached or exceeded the threshold. The probability of; Indicates the future number Predicted time and spatial location The wave level has reached or exceeded the threshold. The probability of; This represents the Sigmoid activation function; Indicates the future number Predicted time and spatial location Wind level label at the location; Indicates the future number Predicted time and spatial location Wave level label; Represents a probability operator; Indicates in When the wind level reaches or exceeds the level class; Indicates in Under the condition that the wave level reaches or exceeds the first class.

[0101] Therefore, for a single sample, the ordered evidence output of wind level can be expressed as:

[0102] In the formula, Indicates sample The corresponding multi-step wind level threshold exceedance probability set, which only includes the wind level threshold exceedance probability on valid ocean grid points.

[0103] The ordered evidence output of wave levels can be represented as:

[0104] In the formula, Indicates sample The corresponding multi-step wave level threshold exceedance probability set, which only includes wave level threshold exceedance probabilities on valid ocean grid points.

[0105] In a preferred embodiment, the number of wind level categories is: The number of wave level categories is Therefore, the wind level ordered evidence output head outputs 12 wind level threshold exceedance probability maps at each prediction time, corresponding to the probability of the wind level reaching or exceeding level 1 to level 12 respectively; the wave level ordered evidence output head outputs 9 wave level threshold exceedance probability maps at each prediction time, corresponding to the probability of the wave level reaching or exceeding level 1 to level 9 respectively.

[0106] In a preferred embodiment, when the number of future prediction steps is At that time, the wind level ordered evidence output header outputs wind level threshold exceedance probability maps for the next 1 hour, 2 hours, and 3 hours respectively; the wave level ordered evidence output header outputs wave level threshold exceedance probability maps for the next 1 hour, 2 hours, and 3 hours respectively. The above wind and wave threshold exceedance probabilities are used as... Figure 4The ordered threshold evidence on the right side of the model shown is used for the monotonic ordered constraint in step S8, the risk threshold evidence extraction in step S9, and the derivation of the threshold consistency risk probability in step S10.

[0107] Through the aforementioned ordered evidence output method, this invention can directly learn the exceedance probabilities of different wind and wave level thresholds, enabling the model output to be more than just ordinary classification results. Instead, it provides structured wind and wave evidence with hierarchical meaning and risk triggering implications. This structured evidence supports the subsequent derivation of risk probabilities from the wind and wave threshold evidence, improving the logical consistency and traceability of wind and wave risk warning results.

[0108] Step 8: Apply monotonically ordered constraints to the output wind level threshold exceedance probability and wave level threshold exceedance probability respectively to obtain wind and wave ordered evidence; Specifically, since both wind and wave grades possess inherent ordered properties, for the same prediction time and the same effective ocean grid point, the probability of reaching or exceeding a higher grade threshold should not be greater than the probability of reaching or exceeding a lower grade threshold. For example, the probability of a wind grade reaching or exceeding level 7 should not be greater than the probability of a wind grade reaching or exceeding level 6; the probability of a wave grade reaching or exceeding level 6 should not be greater than the probability of a wave grade reaching or exceeding level 5. To avoid logical overlap between the probabilities of exceeding different grade thresholds, this invention applies a monotonically ordered constraint to the threshold bias term in step 7.

[0109] Specifically, let the variable type be... ,in Represents the wind level variable. Represent the wave level variable; let the variable be... The number of grade categories is Its threshold index is For the future At each prediction time, variables At the threshold The logit at each location can be uniformly represented as:

[0110] In the formula, Indicates the future number Predicted time and spatial location Variable Reaching or exceeding the threshold The logit value; Representing variables In the future Predicted time and spatial location The shared space at the location responds; Representing variables In the future The predicted time, the first Bias terms at each threshold.

[0111] To ensure that the threshold exceedance probability satisfies the hierarchical order relationship, this invention modifies the bias term. The following monotonic parameterization method is adopted:

[0112] In the formula, Representing variables In the future The learnable parameters corresponding to the first threshold bias term at each prediction time; A learnable parameter that controls the difference in adjacent threshold biases; This represents a non-negative activation function.

[0113] The non-negative activation function can be expressed as:

[0114] In the formula, Indicates the input variable. Because... , by formula To the style We can obtain:

[0115] In the formula, Representing the same variable At the same prediction time Under these conditions, the threshold bias terms of different levels satisfy a monotonically non-increasing relationship.

[0116] Furthermore, due to the same variable At the same prediction time and the same spatial location Under these conditions, different thresholds share the same spatial response. And only through the threshold bias term Therefore, we have:

[0117] In the formula, This indicates that at the same valid ocean grid point, the logit values ​​corresponding to different threshold levels satisfy a monotonically non-increasing relationship.

[0118] Furthermore, due to the Sigmoid function It is a monotonically increasing function, and the variable is... The threshold exceedance probability satisfies:

[0119] In the formula, Indicates the future number Predicted time and spatial location Variable Reaching or exceeding the threshold The probability of . Formula This indicates that the probability of exceeding a higher-level threshold is no greater than the probability of exceeding a lower-level threshold.

[0120] In wind level variables In this case, the above monotonic relationship is specifically expressed as:

[0121] In the formula, Indicates the future number Predicted time and spatial location The wind level has reached or exceeded the threshold. The probability, .

[0122] In wave level variable In this case, the above monotonic relationship is specifically expressed as:

[0123] In the formula, Indicates the future number Predicted time and spatial location The wave level has reached or exceeded the threshold. The probability, .

[0124] Through the above-mentioned monotonic and ordered constraints, this invention can ensure that the threshold exceedance probability of wind level and wave level conforms to the level order law, avoid the threshold crossing problem where the exceedance probability of high level threshold is greater than that of low level threshold, make the wind and wave ordered evidence physically reasonable and logically consistent, and provide a reliable foundation for subsequent risk threshold evidence extraction and threshold consistency risk probability derivation.

[0125] Step 9: Obtain wind triggering evidence and wave triggering evidence corresponding to the wind level risk threshold and the wave level risk threshold from the ordered wind and wave evidence; Specifically, based on the wind level threshold exceedance probability and wave level threshold exceedance probability output in step S7, and combined with the monotonic order constraint applied in step S8, this invention extracts wind triggering evidence and wave triggering evidence corresponding to preset risk thresholds from the wind and wave order evidence. The wind triggering evidence is used to represent the probability that the wind level at a certain effective ocean grid point at a certain future prediction time will reach or exceed the preset wind level risk threshold; the wave triggering evidence is used to represent the probability that the wave level at a certain effective ocean grid point at a certain future prediction time will reach or exceed the preset wave level risk threshold.

[0126] Specifically, the preset wind level risk threshold is set as follows: The preset wave level risk threshold is For the future Predicted time and spatial location For valid ocean grid points, wind-triggered evidence is defined as:

[0127] In the formula, Indicates the future number Predicted time and spatial location Evidence triggered by the wind at the location; The wind level ordered evidence output head is at the threshold. The probability of exceeding the threshold at that point; Indicates the future number Predicted time and spatial location Wind level label at the location; This represents the shared spatiotemporal feature representation obtained in step S6; Indicates in Wind level is greater than or equal to the threshold under the condition .

[0128] Wave-triggered evidence is defined as:

[0129] In the formula, Indicates the future number Predicted time and spatial location Evidence triggered by the waves at the location; The output header indicates that the wave level is ordered at the threshold. The probability of exceeding the threshold at that point; Indicates the future number Predicted time and spatial location Wave level label; Indicates in Under the condition that the wave level is greater than or equal to the threshold .

[0130] For a single prediction moment, the wind-triggered evidence map can be represented as:

[0131] In the formula, Indicates sample In the future Wind-triggered evidence maps for each predicted moment, which are generated and used only on valid ocean grid points.

[0132] The wave-triggered evidence diagram can be represented as follows:

[0133] In the formula, Indicates sample In the future A wave-triggered evidence map for each predicted moment, which is generated and used only on valid ocean grid points.

[0134] For all future prediction steps, the wind-triggered evidence sequence can be represented as:

[0135] In the formula, Indicates sample The corresponding multi-step wind-triggered evidence sequence; This indicates the number of steps to predict the future.

[0136] The wave-triggered evidence sequence can be represented as:

[0137] In the formula, Indicates sample The corresponding multi-step wave triggering evidence sequence.

[0138] In a preferred embodiment, the preset wind level risk threshold is: The preset wave level risk threshold is Based on the wind and wave level classification rules in step S3, the wind triggering condition corresponds to:

[0139] The wave trigger condition corresponds to:

[0140] In the formula, Indicates the future number Predicted time and spatial location 10 m wind speed at the location; Indicates the future number Predicted time and spatial location Significant wave height at that location.

[0141] Through the above method, this invention converts multi-level wind and wave prediction results into wind-triggered and wave-triggered evidence directly corresponding to business risk thresholds, enabling subsequent risk probabilities to be derived from the specific wind level threshold exceedance probability and wave level threshold exceedance probability. Compared to directly outputting independent binary risk probabilities, this invention ensures that risk warning results are consistent with the wind and wave safety threshold triggering logic, while retaining the wind and wave threshold evidence upon which the risk probability derivation is based.

[0142] Step 10: Based on the wind-triggered evidence, the wave-triggered evidence, the wind level risk threshold, and the wave level risk threshold, a multi-step risk probability map is calculated using the threshold consistency rule; Specifically, based on the wind-triggered and wave-triggered evidence extracted in step 9, this invention derives the final risk probability according to the wind and wave risk label construction logic in step 4. Since the risk state in step 4 is defined as "the wind level reaches or exceeds the preset wind level risk threshold, or the wave level reaches or exceeds the preset wave level risk threshold", the final risk probability should be consistent with the threshold triggering rule, rather than being directly given by an independent binary risk classification header.

[0143] For the future Predicted time and spatial location For valid ocean grid points, wind-triggered events and wave-triggered events are defined as follows:

[0144] In the formula, Indicates the future number Predicted time and spatial location The wind-triggered event occurred at the location; Indicates the future number Predicted time and spatial location The wave triggers an event; Indicates the wind level label for the corresponding grid point; Indicates the wave level label for the corresponding grid point; This indicates the preset wind level risk threshold; This indicates the preset wave level risk threshold.

[0145] A risk event is defined as the union of wind-triggered events and wave-triggered events.

[0146] In the formula, Indicates the future number Predicted time and spatial location Risk events at the location; This represents a union relation. Consistent with the risk label definition in step 4, a grid point is in a risk state as long as either the wind level or the wave level reaches the corresponding risk threshold.

[0147] The wind-triggered evidence and wave-triggered evidence obtained from step 9 can be represented as follows:

[0148] In the formula, Indicates the future number Predicted time and spatial location The probability that the wind level reaches or exceeds the preset wind level risk threshold; Indicates the future number Predicted time and spatial location The probability that the wave level reaches or exceeds the preset wave level risk threshold; This represents the shared spatiotemporal feature representation obtained in step S6.

[0149] This invention employs a threshold-consistent noisy-OR rule to derive the final risk probability based on wind-triggered and wave-triggered evidence:

[0150] In the formula, Indicates the future number Predicted time and spatial location The risk probability at a given location. This risk probability represents the probability of the corresponding effective ocean grid point in the future. The probability of being in a risky state at a given moment.

[0151] Mode It can also be equivalently expanded as:

[0152] The expanded form of the formula shows that as long as either wind-triggered evidence or wave-triggered evidence is high, the final risk probability will increase; when both wind-triggered evidence and wave-triggered evidence are high, the final risk probability will further increase. Therefore, the final risk probability is consistent with the rule that "any of strong winds or high waves reaching a risk threshold triggers risk."

[0153] For the future At any given prediction time, the risk probability diagram can be represented as:

[0154] In the formula, Indicates sample In the future Risk probability diagram for each predicted time point; This indicates that risk probabilities are generated and used only on valid ocean grid points. For The risk probability of land grid points or grid points outside the target sea area can be set to a null value, a masked value, or a background area that is not displayed.

[0155] For all future prediction steps, the multi-step risk probability diagram sequence can be represented as:

[0156] In the formula, Indicates sample The corresponding multi-step risk probability diagram sequence; This indicates the number of steps to predict the future.

[0157] In a preferred embodiment, the number of future prediction steps is: The model outputs risk probability maps for three forecast periods: 1 hour, 2 hours, and 3 hours.

[0158] In the formula, This represents the spatial distribution of risk probability over the next hour. This represents the spatial distribution of risk probability over the next 2 hours. This represents the spatial distribution of risk probability over the next 3 hours. The multi-step risk probability map described above can reflect the trends of risk areas appearing, strengthening, weakening, expanding, or moving within the short-term forecast window.

[0159] By employing the aforementioned threshold-consistent risk derivation method, this invention eliminates the need for a separate binary risk classification output head. Instead, it directly derives the final risk probability from the exceedance probabilities of wind and wave level thresholds. Consequently, the final risk probability maintains consistency with the wind and wave level threshold triggering rules while preserving the corresponding wind and wave threshold evidence sources, thus improving the logical consistency and traceability of multi-step marine wind and wave risk probability early warning results.

[0160] Step 11: Based on the multi-step risk probability map, generate a binary risk warning map according to the preset risk probability judgment threshold, and calculate the risk boundary uncertainty map; Specifically, based on the multi-step risk probability map obtained in step 10, the present invention further generates a binary risk warning map according to a preset risk probability judgment threshold, and calculates a risk boundary uncertainty map. The binary risk warning map is used to indicate whether each effective ocean grid point is in a risk state at multiple future prediction times; the risk boundary uncertainty map is used to represent the degree of uncertainty when the risk probability approaches the judgment boundary.

[0161] Let the risk probability judgment threshold be... ,in For the future... Predicted time and spatial location Effective ocean grid points, based on risk probability Generate binary risk warning labels:

[0162] In the formula, Indicates the future number Predicted time and spatial location The binary risk warning result at the location; This represents the risk probability of the corresponding grid point; Indicates the threshold for determining the probability of risk; when When, it indicates that the effective ocean grid point is identified as a risk grid point; when When this occurs, it indicates that the valid ocean grid point is determined to be a non-risk grid point.

[0163] For the future At any given prediction time, the binary risk warning map can be represented as:

[0164] In the formula, Indicates sample In the future Binary risk warning map for each predicted time point; This indicates that binary risk warning results are generated only on valid ocean grid points.

[0165] For all future prediction steps, the multi-step binary risk warning graph sequence can be represented as:

[0166] In the formula, Indicates sample The corresponding multi-step binary risk warning chart sequence; This indicates the number of steps to predict the future.

[0167] In a preferred embodiment, the risk probability determination threshold is set to... That is, when the risk probability of a certain effective ocean grid point is not less than 0.5, the grid point is identified as a risk grid point; when the risk probability is less than 0.5, the grid point is identified as a non-risk grid point.

[0168] When the risk probability is re-derived using the adjusted business storm risk threshold In this case, the binary risk warning label can be represented as follows:

[0169] In the formula, This represents the binary risk warning result obtained under the adjusted business risk threshold; This represents the risk probability re-derived after adjusting the business threshold.

[0170] Furthermore, the risk probability determination threshold is set to... In a preferred embodiment, to characterize the degree of uncertainty when the risk probability approaches the decision boundary, the present invention calculates the uncertainty of the risk boundary based on the risk probability. For the future... Predicted time and spatial location For an effective ocean grid point, the uncertainty of the risk boundary is defined as follows:

[0171] In the formula, Indicates the future number Predicted time and spatial location Uncertainty at the risk boundary; This represents the risk probability of the corresponding grid point. (From the formula...) It can be seen that when the risk probability The closer the probability is to 0.5, the higher the uncertainty of the risk boundary; the closer the probability of risk is to 0 or 1, the lower the uncertainty of the risk boundary.

[0172] For the future At any given prediction time, the uncertainty diagram of the risk boundary can be represented as follows:

[0173] In the formula, Indicates sample In the future Uncertainty diagram of the risk boundary at each prediction time.

[0174] For all future prediction steps, the multi-step risk boundary uncertainty diagram sequence can be represented as:

[0175] In the formula, Indicates sample The corresponding multi-step risk boundary uncertainty diagram sequence.

[0176] During the evaluation or visualization phase, the true risk label at future moments... When available, risk prediction and diagnosis results can also be generated based on the binary risk warning results and the true risk labels; the risk prediction and diagnosis results are used to distinguish between correct warnings, correct non-warnings, false alarms, and missed alarms.

[0177] In the formula, Indicates the future number Predicted time and spatial location Risk prediction and diagnosis results at the location; This indicates a correct warning; This indicates a correct, non-warning message; Indicates a false alarm; This indicates a missed warning. This diagnostic result relies on the true risk label and is therefore primarily used for model evaluation, error analysis, and visualization of typical events; it is not essential input information for actual prediction.

[0178] For the future At any given prediction time, the risk prediction diagnostic diagram can be represented as follows:

[0179] In the formula, Indicates sample In the future Risk prediction diagnostic chart for each prediction time.

[0180] In a preferred embodiment, when the number of future prediction steps is This invention can output risk probability maps, binary risk warning maps, and risk boundary uncertainty maps for the next 1, 2, and 3 hours. During model evaluation or typical event visualization stages, when true risk labels are available, a risk prediction diagnostic map can also be generated. Typical risk events are shown in the forecast lead time. The following parameters are considered: true value of wind level, wind level prediction, wind level error, probability of exceeding wind level threshold, ordered uncertainty of wind level; true value of wave level, wave level prediction, wave level error, probability of exceeding wave level threshold, ordered uncertainty of wave level; true value of risk, risk prediction, risk prediction diagnosis, spatial distribution of risk probability and risk boundary uncertainty. Figure 8 As shown in the figure, risk prediction diagnosis is used to distinguish between areas of correct early warning, correct non-early warning, false alarm, and missed alarm; risk boundary uncertainty is used to represent the degree of uncertainty when the risk probability approaches the judgment boundary. This figure illustrates that the present invention can generate a risk probability map based on ordered threshold evidence of wind and waves, and can perform spatial visualization analysis of the risk boundary and prediction error area.

[0181] Through the above methods, this invention can further generate a binary risk warning map that can be directly used for risk display from continuous risk probabilities, and simultaneously output a risk boundary uncertainty map to indicate spatial areas where the risk probability is close to the judgment boundary. In the evaluation stage, the risk prediction and diagnostic map can also identify false alarm and missed alarm areas, providing a basis for model error analysis and interpretation of warning results.

[0182] Step 12: For the temporal jump-connection fusion ordered regression spatiotemporal model, construct a multi-task loss function and optimize the model parameters, and train and validate the temporal jump-connection fusion ordered regression spatiotemporal model based on the short-time multi-step prediction samples to obtain the model with the best performance on the validation set. Specifically, in the model training phase, the present invention performs end-to-end training on the model described in steps 6 to 11 based on the short-term multi-step prediction samples constructed in step 5. The training objectives include wind level ordered evidence learning, wave level ordered evidence learning, and risk warning probability learning, so that the model can output wind and wave threshold exceedance probabilities that conform to the level order relationship, and generate risk probability maps that are consistent with the wind and wave risk threshold triggering rules.

[0183] For the future Predicted time and spatial location Effective ocean grid points, based on real wind level labels. Construct a cumulative binary monitoring label for wind level:

[0184] In the formula, Indicates wind level at the threshold Cumulative binary supervision label at the location; Indicates the actual wind level label; .

[0185] Similarly, based on actual wave level labels Construct cumulative binary supervised labels for wave levels:

[0186] In the formula, Indicates wave level at the threshold Cumulative binary supervision label at the location; Indicates the actual wave level label; .

[0187] The binary cross-entropy loss function is defined as:

[0188] In the formula, Indicates the model's predicted probability; This represents the corresponding binary supervision label.

[0189] Based on the wind level threshold exceedance probability output in step 7 Wind level ordered loss is defined as:

[0190] In the formula, Indicates wind level-based losses; Indicates the number of steps to predict the future; Represents the effective set of ocean grid points; Indicates the number of valid ocean grid points; This indicates the number of wind level categories.

[0191] Based on the wave level threshold exceedance probability output in step 7 The wave-level ordered loss is defined as:

[0192] In the formula, Indicates ordered loss based on wave level; Indicates the number of wave classification categories; This indicates that the wave level has reached or exceeded the threshold. The predicted probability.

[0193] Based on the risk probability derived in step 10 and the real risk labels constructed in step 4 Risk warning loss is defined as:

[0194] In the formula, Indicates risk warning and loss; Indicates the true risk label; This represents the risk probability derived from evidence of wind and wave thresholds. This represents the weight of positive risk samples, used to increase the contribution of risk grid points during training, thereby alleviating the class imbalance problem caused by the relative sparsity of risk samples.

[0195] The model's total loss function is defined as:

[0196] In the formula, Represents the total loss function of the model; Indicates wind level-based losses; Indicates ordered loss based on wave level; Indicates risk warning and loss; , and These represent the weighting coefficients for wind-level ordered losses, wave-level ordered losses, and risk warning losses, respectively.

[0197] In a preferred embodiment, the three loss terms are given the same weight, i.e. Through the aforementioned multi-task loss function, the wind level ordered evidence output head, the wave level ordered evidence output head, the threshold-consistent risk derivation module, and the shared spatiotemporal feature extraction network can be jointly optimized. This is because the risk probability... Derived from wind-triggered and wave-triggered evidence, the risk warning loss can propagate backward along the risk derivation chain to the wind level ordered evidence output head, the wave level ordered evidence output head, and the shared spatiotemporal feature extraction network, thereby prompting the model to pay more attention to wind and wave samples and risk boundary areas that are close to the risk threshold.

[0198] The objective for optimizing model parameters can be expressed as:

[0199] In the formula, This represents the optimal model parameters obtained after training. This represents the learnable parameters in TSF-OR-STUNet; This indicates that when the model parameters are The corresponding total loss function.

[0200] In a preferred embodiment, the AdamW optimizer is used to optimize the model parameters. The parameter update process of the next iteration can be represented as:

[0201] In the formula, Indicates the first Model parameters at the next iteration; This represents the updated model parameters; This represents the gradient of the total loss function with respect to the model parameters; This indicates the AdamW parameter update operator.

[0202] During training, the optimal model parameters can be selected based on the performance on the validation set. After the model is trained, the risk warning results on the validation and test sets are evaluated. For example... Figure 5 As shown, the model can maintain good risk identification performance under different forecast lead times, and can divide the risk prediction results into different components such as wind triggering, wave triggering and combined wind and wave triggering, which can be used to analyze the impact of different risk sources on the early warning results.

[0203] like Figure 6 As shown, the reliability of the risk probability is evaluated by comparing the risk probability output by the model with the actual frequency of risk occurrence. The results show that the risk probability output by the model has good consistency and calibration performance.

[0204] like Figure 7 As shown, ablation experiments were conducted by removing the temporal jump connection fusion module, monotonic ordered constraints, threshold consistent risk derivation module, or risk auxiliary supervision respectively, to compare the risk warning performance under different model settings, so as to illustrate the role of the above key technologies in improving multi-step risk identification capability, hierarchical logical consistency, and risk probability reliability.

[0205] Through the above training and verification methods, this invention can simultaneously optimize ordered evidence of wind level, ordered evidence of wave level, and risk probabilities derived from wind and wave threshold evidence, ensuring that the model training objective aligns with the final risk warning objective. The saved optimal model parameters are used for subsequent short-term multi-step risk probability inference, generation of binary risk warning maps, and output of risk boundary uncertainty maps.

[0206] Step 13: Based on the model with the best performance on the validation set, perform predictive reasoning on the latest historical multivariate marine meteorological field of the target sea area to generate short-term multi-step marine risk warnings. The short-term multi-step marine risk warnings include the ordered threshold exceedance probability of wind and waves at multiple future prediction times, risk probability maps, binary risk warning maps, and risk boundary uncertainty maps.

[0207] Specifically, after the model training is completed, the optimal model parameters obtained in step 12 are used. It performs predictive reasoning on the latest historical multivariate marine meteorological field of the target sea area, and generates wind and wave ordered threshold exceedance probability, risk probability map, binary risk warning map and risk boundary uncertainty map for multiple future prediction times.

[0208] Let the latest anchor point time be Based on the processing methods in steps 1 to 5, construct the latest historical input window:

[0209] In the formula, Indicates the latest anchor point time This is the historical input window for the end time. This represents the multivariate marine meteorological field at the first moment in the historical input window; This represents the multivariable marine meteorological field at the latest anchor point moment; This indicates the length of the historical input time step.

[0210] Enter the latest history window Input the trained TSF-OR-STUNet to obtain the shared spatiotemporal feature representation:

[0211] In the formula, This represents the shared spatiotemporal features corresponding to the latest historical input window; Indicates the use of optimal parameters TSF-OR-STUNet; This represents the optimal model parameters obtained after training.

[0212] Based on shared spatiotemporal feature representation Following steps S7 and S8, generate the wind level threshold exceedance probability and wave level threshold exceedance probability for multiple future prediction times:

[0213] In the formula, Indicates the latest anchor point time Next, the next chapter Predicted time and spatial location The wind level has reached or exceeded the threshold. The probability of; This indicates that the wave level at the corresponding location has reached or exceeded the threshold. The probability of; , ; Indicates in Latest anchor point time under the condition Downwind level reaches or exceeds class; Indicates in Latest anchor point time under the condition The next wave level reaches or exceeds the first class.

[0214] Based on the preset wind level risk threshold and preset wave level risk threshold Extract wind-triggered evidence and wave-triggered evidence according to step S9:

[0215] In the formula, Indicating the future in the latest reasoning process Predicted time and spatial location Evidence triggered by the wind at the location; This indicates evidence of wave triggering at the corresponding location.

[0216] Further, following the threshold consistency noisy-OR rule in step S10, derive the future... Predicted time and spatial location Risk probability at the location:

[0217] In the formula, Indicating the future in the latest reasoning process Predicted time and spatial location The probability of risk at a given location. This probability is derived from the probability of exceeding the wind level risk threshold and the probability of exceeding the wave level risk threshold, and is consistent with the rule that "risk is triggered when either strong wind or high wave reaches the risk threshold".

[0218] For the future The risk probability diagram for each predicted time point is as follows:

[0219] In the formula, Indicates the latest anchor point time The next chapter Risk probability diagram for each predicted time point; This indicates that risk probabilities are generated and used only on valid ocean grid points.

[0220] For all future prediction steps, the multi-step risk probability diagram sequence is represented as:

[0221] In the formula, Indicates the latest anchor point time The following is a multi-step risk probability graph sequence; This indicates the number of steps to predict the future.

[0222] Based on the preset risk probability judgment threshold Generate a binary risk warning map according to step S11:

[0223] In the formula, Indicating the future in the latest reasoning process Predicted time and spatial location The binary risk warning result at the location; This represents the threshold for determining the risk probability. In a preferred embodiment, .

[0224] Further, calculate the uncertainty of the risk boundary according to step 11:

[0225] In the formula, Indicating the future in the latest reasoning process Predicted time and spatial location The risk boundary is uncertain. When the risk probability... The closer the probability is to 0.5, the higher the uncertainty of the risk boundary; the closer the probability of risk is to 0 or 1, the lower the uncertainty of the risk boundary.

[0226] In a preferred embodiment, the historical input time step length is The number of future prediction steps is At this point, the present invention uses the latest multivariate marine meteorological field of the past 12 hours as input and outputs risk probability maps, binary risk warning maps, and risk boundary uncertainty maps for the next 1 hour, 2 hours, and 3 hours:

[0227] In the formula, , and These represent the risk probability diagrams for the next 1 hour, 2 hours, and 3 hours, respectively. , and These represent the binary risk warning maps for the corresponding forecast lead time; , and These represent the risk boundary uncertainty diagrams for the corresponding forecast lead time.

[0228] When new hourly marine meteorological data are continuously acquired, the anchor point time can be updated using a sliding window method. The latest historical input window is reconstructed, and the above predictive reasoning process is repeated to form a continuously updated short-term, multi-step risk output. This process can be represented as:

[0229] In the formula, This indicates the time interval for data updates. Under hourly data conditions, One hour is available.

[0230] During the model evaluation or typical event visualization phase, when the true risk labels for future moments are available, a risk prediction diagnostic map can be generated according to step S11. This map is used to distinguish between areas with correct early warnings, correct non-early warnings, false alarms, and missed warnings. This risk prediction diagnostic map is mainly used for model evaluation, error analysis, and visualization, and is not essential input information for the actual prediction inference phase.

[0231] Through the above-mentioned predictive reasoning method, this invention can utilize the trained TSF-OR-STUNet to output wind and wave threshold evidence, risk probability maps, binary risk warning maps, and risk boundary uncertainty maps for multiple forecast periods in the future, based on the latest historical multivariate marine meteorological field. This provides model output results for short-term multi-step marine wind and wave risk identification and early warning display.

[0232] Furthermore, since the wind and wave risk thresholds used in different application scenarios or evaluation conditions may differ, this invention supports the re-deriving of risk probabilities based on given wind level and wave level risk thresholds without retraining the model. Specifically, the system can directly select wind triggering evidence and wave triggering evidence corresponding to the given threshold from the wind level threshold exceedance probability and wave level threshold exceedance probability output in step 7, and recalculate the risk probability according to the threshold consistency risk derivation rules described in step 10.

[0233] It should be noted that the specific wind and wave risk thresholds used should be determined separately based on vessel conditions, operational tasks, marine environment, or relevant safety regulations. This invention provides a risk probability re-derivation mechanism based on given wind and wave risk thresholds, rather than limiting a single safety threshold to a specific business scenario.

[0234] Let the adjusted business risk level threshold be: The adjusted business wave level risk threshold is ,in:

[0235] In the formula, This indicates the adjusted business risk level threshold; This indicates the adjusted risk threshold for the business wave level; Indicates the number of wind level categories; Indicates the number of wave class categories.

[0236] For the future Predicted time and spatial location For valid ocean grid points, wind-triggered evidence corresponding to the adjusted wind level risk threshold can be directly selected from the ordered wind level evidence output:

[0237] In the formula, This indicates the adjusted business risk level threshold. Next, the next chapter Predicted time and spatial location Evidence triggered by the wind at the location; This indicates that the wind level has reached or exceeded the threshold. The probability of exceeding the threshold.

[0238] At the same time, wave triggering evidence corresponding to the adjusted wave level risk threshold can be directly selected from the ordered wave level evidence output:

[0239] In the formula, This indicates the adjusted business wave level risk threshold. Next, the next chapter Predicted time and spatial location Evidence triggered by the waves at the location; This indicates that the wave level has reached or exceeded the threshold. The probability of exceeding the threshold.

[0240] Based on the adjusted wind-triggered and wave-triggered evidence, the risk probability is re-derived according to the same threshold-consistent noisy-OR rule as in step S10:

[0241] In the formula, This indicates that under the adjusted business risk threshold, the future [number]th [year]... Predicted time and spatial location The risk probability is re-derived.

[0242] Mode It can also be equivalently expanded as:

[0243] The expanded form of the formula shows that after adjusting the business threshold, the final risk probability remains consistent with the logic that "any wind or wave reaching the adjusted risk threshold will trigger the risk".

[0244] For the future At each prediction time, the risk probability diagram after adjusting the threshold can be represented as follows:

[0245] In the formula, This indicates that under the adjusted business risk threshold, the sample... In the future Risk probability diagram for each predicted time.

[0246] For all future prediction steps, the multi-step risk probability map sequence after threshold adjustment can be represented as:

[0247] In the formula, This represents a multi-step risk probability diagram sequence re-derived under the adjusted business risk threshold. This indicates the number of steps to predict the future.

[0248] In a preferred embodiment, the default risk threshold is , When a more conservative risk warning result is needed, the business threshold can be adjusted to a lower level, for example, adjusting the wind level risk threshold and wave level risk threshold to [a lower level]. , When the business scenario allows for continued navigation or operation under high wind and wave conditions, the business threshold can be adjusted to a higher level, for example... , .

[0249] Through the above method, this invention can directly utilize the output wind and wave ordered threshold exceedance probabilities to re-derive risk probability maps for different business thresholds while keeping the model parameters unchanged. Compared to a fixed binary classification risk output model, this invention has greater flexibility in threshold adjustment and business adaptability.

[0250] This invention achieves short-term, multi-step risk probability early warning at sea driven by wind and wave physical level evidence through threshold-consistent ordered wind and wave evidence modeling, and has the following technical effects.

[0251] First, this invention ensures that the risk probability output aligns with the sea wind and wave safety threshold triggering logic. Traditional methods typically predict continuous wind speed or wave height first, then determine whether a safety threshold is exceeded through post-processing, or directly model the risk warning as an independent binary classification task. These methods easily lead to inconsistencies between the model training objective, the risk output result, and the actual safety threshold rules. This invention first outputs the wind level threshold exceedance probability and the wave level threshold exceedance probability, then extracts wind triggering evidence and wave triggering evidence based on preset wind level and wave level risk thresholds, and derives the final risk probability through a threshold-consistent noisy-OR rule, ensuring that the risk probability directly corresponds to the business logic of "risk is triggered when either the wind level or wave level reaches the risk threshold."

[0252] Secondly, this invention improves the logical consistency and physical rationality of wind and wave level prediction results. Both wind and wave levels have a natural ordering relationship; the probability of exceeding a higher level threshold should not be greater than the probability of exceeding a lower level threshold. This invention, through monotonic ordering constraints, ensures that the exceedance probabilities of wind and wave level thresholds satisfy a monotonic relationship, avoiding the threshold crossing problem where the exceedance probability of a higher level threshold is greater than that of a lower level threshold. This makes the wind and wave evidence output by the model more consistent with physical level patterns.

[0253] Third, this invention improves the traceability of risk probability output. Since the final risk probability is not directly given by an independent binary classification output head, but rather derived from the exceedance probabilities of wind level and wave level risk thresholds, each risk probability can be traced back to the corresponding wind-triggered and wave-triggered evidence. This method makes the risk probability no longer an isolated classification result, but a probabilistic output with clear wind and wave threshold bases, which helps improve the interpretability and credibility of maritime wind and wave risk warning results.

[0254] Fourth, this invention can output short-term multi-step risk probability maps. Compared with single-step risk warnings, multi-step risk probability maps can simultaneously provide the spatial distribution of risks over multiple forecast periods, allowing users to observe the emergence, intensification, weakening, expansion, or movement trends of risk areas within short-term forecast windows. In a preferred embodiment, this invention uses the multivariate marine meteorological field of the past 12 hours as input and outputs risk probability maps for three forecast periods: 1 hour, 2 hours, and 3 hours, thereby improving the lead time and continuity of short-term maritime risk identification.

[0255] Fifth, this invention enhances the expressive power of risk area boundaries and spatial structures. The preferred model, TSF-OR-STUNet, extracts multi-scale spatial features through a shared two-dimensional encoder, captures the temporal evolution of large-scale wind and wave fields through a bottleneck time fusion module, and introduces a temporal jump connection fusion module into the jump connection path to inject historical dynamic information into jump connection features at different scales. Thus, the model can retain the local spatial structure of the most recent moment while incorporating historical wind and wave field evolution information, which helps improve the expressive power of risk area boundaries, spatial morphology, and continuous change processes in multi-step risk probability maps.

[0256] Sixth, this invention can generate a risk boundary uncertainty map to indicate spatial regions where the risk probability is close to the decision boundary. For maritime risk early warning, grid points with a risk probability close to 0.5 typically represent risk boundary regions where the model's decision is more uncertain. This invention calculates the risk boundary uncertainty based on the risk probability; the closer the risk probability is to 0.5, the higher the uncertainty; the closer the risk probability is to 0 or 1, the lower the uncertainty. This output can help identify uncertain regions near the risk boundary, providing a reference for model error analysis, risk result display, and manual review.

[0257] Seventh, this invention enables the re-derivation of risk probabilities under different wind and wave risk thresholds. Since the model outputs multi-level threshold exceedance probabilities for wind and wave levels, while keeping the model parameters unchanged, the corresponding threshold probabilities can be directly selected from existing ordered evidence of wind and waves based on the given wind and wave risk thresholds, and the risk probabilities can be re-derived using the noisy-OR rule. Compared to a fixed binary risk output model, this invention maintains consistency in risk derivation logic under different threshold settings, improving adaptability when threshold settings change.

[0258] Eighth, this invention enables the learning of ordered wind and wave evidence and the learning of risk probability to mutually reinforce each other through multi-task training. This invention simultaneously constructs ordered wind level loss, ordered wave level loss, and risk warning loss, allowing the model to both learn the probability of exceeding the wind and wave level threshold and directly optimize the risk probability derived from the wind and wave threshold evidence. Since the risk warning loss can propagate backward along the noisy-OR risk derivation chain to the ordered wind level evidence output head, the ordered wave level evidence output head, and the shared spatiotemporal feature extraction network, the model can pay more attention to wind and wave samples close to the risk threshold and risk boundary regions, thereby improving the consistency between the risk warning target and the model training target.

[0259] Ninth, in the 5-hour ERA reanalysis data experiment of the target nearshore waters used in this embodiment, the preferred model of this invention achieved good comprehensive early warning effect and probabilistic reliability in the three-step marine wind and wave risk early warning task. Experimental results show that the preferred model achieved a PR-AUC of 0.992, a CSI / IoU of 0.911, and an ECE of 0.826 × 10⁻⁶ under representative experimental settings. - ³, Brier's score reached 0.990 × 10⁸ - ², and maintains a relatively balanced pattern of false alarms and missed alarms. The above results demonstrate that, under the experimental conditions described above, the present invention can improve the ability to identify risk areas, the reliability of risk probability, and the stability of multi-step wind and wave risk warning results.

[0260] The method of this invention converts continuous 10 m wind speed and significant wave height into physically meaningful ordered levels, constructs wind level threshold exceedance probability and wave level threshold exceedance probability, and ensures logical consistency between threshold exceedance probabilities of different levels through monotonic ordered constraints; further, it derives the final risk probability based on wind level risk threshold exceedance probability and wave level risk threshold exceedance probability, so that the risk probability is consistent with the triggering logic of wind and wave safety threshold.

[0261] This invention can output maritime risk probability maps and binary risk warning maps for multiple predicted future time points. It can not only identify future risk areas but also reflect the continuity, expansion, shrinkage, or spatial movement trends of risk areas through multi-step risk maps. Since the final risk probability is derived from wind and wave threshold evidence, the risk probability can be traced back to the corresponding wind-triggered and wave-triggered evidence, thereby improving the logical consistency and traceability of the risk warning results.

[0262] Simultaneously, this invention can output a risk boundary uncertainty map to indicate spatial areas where the risk probability is close to the judgment boundary. During model evaluation or typical event visualization stages, when true risk labels are available, a risk prediction diagnostic map can also be generated to distinguish between correctly warned, correctly non-warning, false alarm, and missed alarm areas. Therefore, this invention can provide technical support for short-term, multi-step ocean wind and wave risk identification, early warning display, error analysis, and decision support.

[0263] This invention can be applied to scenarios such as auxiliary early warning for ship navigation safety, port and waterway risk alerts, safety management of nearshore engineering operations, auxiliary decision-making for offshore wind power operation and maintenance, risk display for offshore platform operations, short-term risk identification of marine disasters, and intelligent shipping systems. The method can output wind and wave risk probability maps and binary risk warning maps for multiple forecast periods in the future based on historical marine meteorological fields, and can output risk boundary uncertainty maps when needed, providing relevant business systems with short-term, multi-step wind and wave risk identification, early warning display, and auxiliary decision-making information.

[0264] In the field of maritime navigation safety, this invention can provide crew members, shipping managers, or intelligent navigation systems with the spatial distribution of wind and wave risk probabilities within a short-term forecast window, assisting in determining whether there are risk areas ahead triggered by strong winds or high waves. Through multi-step risk probability maps for the next 1, 2, and 3 hours, users can observe the spatial changing trends of risk areas, thus providing a reference for navigation decisions such as deceleration, turning, partial route changes, or increasing avoidance distances.

[0265] In the field of port and waterway management, this invention can provide short-term spatial distribution information on wind and wave risks for port scheduling, waterway navigation management, and maritime supervision. Through binary risk warning maps and risk probability maps, managers can identify risk areas that may affect navigation safety in the near future, providing auxiliary information for vessel entry and exit arrangements, waterway control, pilotage scheduling, and navigation safety alerts.

[0266] In the fields of nearshore engineering and offshore operations, this invention can provide short-term wind and wave risk trend assessment for construction vessels, offshore platforms, nearshore engineering operations, and offshore wind power operation and maintenance. The risk probability map can be used to indicate areas where the risk may increase, decrease, or shift in the near future; the risk boundary uncertainty map can serve as an auxiliary reference, indicating spatial areas where the risk probability is close to the judgment boundary, providing a reference for selecting work windows, suspending operations, and managing on-site safety.

[0267] In the field of marine disaster prevention and emergency management, this invention can be used for short-term, multi-step probabilistic identification of strong winds, high waves, and their corresponding risk areas. Compared to risk assessment at a single moment, multi-step risk probability maps can show the temporal changes of risk areas, helping emergency management personnel understand the persistence and spatial evolution trends of risk areas, and providing auxiliary information for nearshore disaster prevention and mitigation, emergency response, and risk assessment.

[0268] In intelligent shipping and maritime risk management platforms, this invention can serve as a short-term wind and wave risk probability output module, used in conjunction with electronic charts, vessel route planning systems, marine meteorological information systems, or maritime regulatory platforms. The module can provide risk probability maps and binary risk warning maps for multiple forecast periods in the future, and can also provide risk boundary uncertainty maps as needed, thereby enhancing the platform's ability to probabilistically express and visualize maritime wind and wave risks.

[0269] A multi-step risk early warning system for marine wind and waves based on threshold consistency includes: The marine meteorological data acquisition and preprocessing module is used to acquire marine meteorological reanalysis data of the target sea area and perform preprocessing to obtain historical multivariate marine meteorological state sequences. An effective ocean grid mask module is used to construct an effective ocean grid mask for a target ocean area based on the land-sea distribution of the target ocean area; The wind and wave rating sub-module is used to convert continuous wind speed and continuous significant wave height into wind rating labels and wave rating labels with physical meaning and rating order, respectively, according to the standard ocean description rating. The wind and wave risk label construction module is used to construct wind and wave risk labels on the effective ocean grid points based on the wind level label and the wave level label, according to the preset wind level risk threshold and wave level risk threshold. The short-time multi-step sample construction module is used to construct short-time multi-step prediction samples based on the historical multivariate marine meteorological state sequence, the effective marine grid mask, the wind level label, the wave level label, and the wind and wave risk label, using a sliding window method. The short-time multi-step prediction samples include a historical input window and a future prediction window. The shared spatiotemporal feature extraction module is used to construct a temporal jump-connection fusion ordered regression spatiotemporal model. Based on the short-time multi-step prediction sample data input through the historical input window of the short-time multi-step prediction sample, the module processes the data using the temporal jump-connection fusion ordered regression spatiotemporal model to extract the multi-scale spatial structure features and temporal evolution features of the historical wind and wave field, thereby obtaining shared spatiotemporal features for wind and wave ordered evidence output and risk probability derivation. The wind and wave ordered evidence output module is used to construct wind level ordered evidence output header and wave level ordered evidence output header based on the shared spatiotemporal features, and output the wind level threshold exceedance probability and wave level threshold exceedance probability at multiple future prediction times. The monotonic order constraint module is used to apply monotonic order constraints to the output wind level threshold exceedance probability and wave level threshold exceedance probability to obtain evidence of wind and wave order. The wind and wave risk threshold evidence extraction module is used to extract wind triggering evidence and wave triggering evidence corresponding to the wind level risk threshold and the wave level risk threshold from the ordered wind and wave evidence. The threshold consistency risk probability derivation module is used to calculate a multi-step risk probability map based on the wind triggering evidence, the wave triggering evidence, the wind level risk threshold, and the wave level risk threshold, using the threshold consistency rule. The binary risk warning map and risk boundary uncertainty generation module is used to generate a binary risk warning map based on the multi-step risk probability map and according to a preset risk probability judgment threshold, and to calculate the risk boundary uncertainty map. The model training and optimization module is used to construct a multi-task loss function and optimize the model parameters for the temporal jump-connection fusion ordered regression spatiotemporal model, and to train and validate the temporal jump-connection fusion ordered regression spatiotemporal model based on the short-time multi-step prediction samples to obtain the model with the best performance on the validation set. The short-term multi-step risk output module is used to predict and infer the latest historical multivariate marine meteorological field of the target sea area based on the model with the best performance of the validation set, and generate short-term multi-step risk warnings for the sea. The short-term multi-step risk warnings for the sea include the ordered threshold exceedance probability of wind and waves at multiple future prediction times, risk probability map, binary risk warning map and risk boundary uncertainty map.

[0270] Furthermore, the temporal jump-connection fusion ordered regression spatiotemporal model includes a shared two-dimensional encoder, a bottleneck time fusion module, a multi-scale temporal jump-connection fusion module, and a spatial decoder; The shared two-dimensional encoder is used to encode the multivariate marine meteorological field at each historical moment entered in the historical input window and extract multi-scale spatial features. The process is as follows: Assume a shared two-dimensional encoder includes Each scale layer, for historical input window The Middle Multivariable marine meteorological field at a historical moment , No. The encoded features of each scale layer can be represented as:

[0271] In the formula, Indicates the first The historical moment at the Spatial coding features at various scales; Indicates the shared two-dimensional encoder in the first... Encoding functions for each scale layer; ; For all historical moments in the history input window Stacking the scale-encoded features yields the _th ... Features of skip-connected sequences at various scales:

[0272] In the formula, Indicated by anchor point time For the end time, the first Features of historical skip sequences at various scales; Indicates the first Number of channels per scale; and They represent the first Spatial height and width at various scales; The bottleneck time fusion module is used to fuse multi-scale spatial features to obtain bottleneck features after bottleneck time fusion, and the process is as follows: For the bottleneck layer of the lowest resolution shared 2D encoder, let its historical sequence characteristics be... The bottleneck time fusion module uses a Convolutional Long Short-Term Memory (ConvLSTM) network, and its fusion process is represented as follows:

[0273] In the formula, This indicates the bottleneck characteristics after bottleneck time fusion. This represents a convolutional long short-term memory network set in the bottleneck layer of the encoder; The multi-scale temporal skip connection fusion module is used to fuse historical skip connection sequence features in multi-scale spatial features to obtain scale skip connection features. The process is as follows: For the Features of skip sequences at various scales ,use Convolution performs channel reduction to decrease the computational cost of temporal modeling.

[0274] The number of channels after dimensionality reduction is defined as:

[0275] In the formula, Indicates the first Features of skip-connected sequences after dimensionality reduction of channels at various scales; Indicates the use of channel dimensionality reduction convolution; This represents the number of channels after dimensionality reduction; Indicates the minimum number of reserved channels; Indicates the channel compression ratio; Indicates rounding down; Then, ConvLSTM is applied to the dimensionality-reduced skip sequence features to obtain the 1st... Historical dynamic aggregation features at various scales:

[0276] In the formula, Indicates the first The temporal aggregated hidden state at the end of the historical input window at each scale; Indicates the first Convolutional long short-term memory networks in scale-hopping paths; Next, the historical dynamic aggregated features are projected back to the original channel dimension to obtain the historical dynamic supplementary features:

[0277] In the formula, Indicates the first Historical dynamic supplementary features projected at various scales; Indicates the use of channel recovery convolution; Let the first The jump connection feature at each scale is: ,in Indicates the first Anchor point time at each scale Local spatial features; A gated fusion mechanism is employed to adaptively fuse recent local spatial features and historical dynamic supplementary features. The gate weights are defined as follows:

[0278] The fusion of the first The scale-hopping feature is represented as follows:

[0279] In the formula, Indicates the first Each scale of gating weights; This represents the Sigmoid activation function; This indicates splicing along the channel dimension; This represents element-wise multiplication; Indicates the first step after time-sequence jump connection fusion Each scale jump connection feature; The spatial decoder is used to fuse the bottleneck features obtained after bottleneck temporal fusion and the scale skip connection features to obtain shared spatiotemporal features. The process is as follows:

[0280] In the formula, Representation spatial decoder; This represents the shared spatiotemporal features obtained after decoding.

[0281] The specific functions of each module in the threshold consistency-based multi-step risk warning system for marine wind and waves disclosed in this invention correspond to the functions implemented in each step of the threshold consistency-based multi-step risk warning method for marine wind and waves disclosed in this invention, and will not be described in detail in this application.

[0282] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-step risk early warning method for marine wind and waves based on threshold consistency, characterized in that, Includes the following steps: Step 1: Obtain marine meteorological reanalysis data of the target sea area and perform preprocessing to obtain historical multivariate marine meteorological state sequences; Step 2: Construct an effective marine grid mask for the target sea area based on the land-sea distribution of the target sea area; Step 3: Based on the standard ocean description level, convert continuous wind speed and continuous significant wave height into wind level labels and wave level labels with physical meaning and grade order, respectively. Step 4: Based on the wind level label and the wave level label, construct wind and wave risk labels on the effective ocean grid points according to the preset wind level risk threshold and wave level risk threshold; Step 5: Based on the historical multivariate marine meteorological state sequence, the effective marine grid mask, the wind level label, the wave level label, and the wind and wave risk label, construct short-term multi-step prediction samples using a sliding window method. The short-term multi-step prediction samples include a historical input window and a future prediction window. Step 6: Construct a temporal jump-connection fusion ordered regression spatiotemporal model, and use the temporal jump-connection fusion ordered regression spatiotemporal model to process the short-time multi-step prediction sample data input into the historical input window of the short-time multi-step prediction sample, extract the multi-scale spatial structure features and temporal evolution features of the historical wind and wave field, and then obtain the shared spatiotemporal features for wind and wave ordered evidence output and risk probability derivation. Step 7: Based on the shared spatiotemporal features, construct the wind level ordered evidence output head and the wave level ordered evidence output head, and output the wind level threshold exceedance probability and wave level threshold exceedance probability for multiple future prediction times. Step 8: Apply monotonically ordered constraints to the output wind level threshold exceedance probability and wave level threshold exceedance probability respectively to obtain wind and wave ordered evidence; Step 9: Obtain wind triggering evidence and wave triggering evidence corresponding to the wind level risk threshold and the wave level risk threshold from the ordered wind and wave evidence; Step 10: Based on the wind-triggered evidence, the wave-triggered evidence, the wind level risk threshold, and the wave level risk threshold, a multi-step risk probability map is calculated using the threshold consistency rule; Step 11: Based on the multi-step risk probability map, generate a binary risk warning map according to the preset risk probability judgment threshold, and calculate the risk boundary uncertainty map; Step 12: For the temporal jump-connection fusion ordered regression spatiotemporal model, construct a multi-task loss function and optimize the model parameters, and train and validate the temporal jump-connection fusion ordered regression spatiotemporal model based on the short-time multi-step prediction samples to obtain the model with the best performance on the validation set. Step 13: Based on the model with the best performance on the validation set, perform predictive reasoning on the latest historical multivariate marine meteorological field of the target sea area to generate short-term multi-step marine risk warnings. The short-term multi-step marine risk warnings include the ordered threshold exceedance probability of wind and waves at multiple future prediction times, risk probability maps, binary risk warning maps, and risk boundary uncertainty maps.

2. The multi-step risk early warning method for marine wind and waves based on threshold consistency according to claim 1, characterized in that: The constructed temporal jump-connection fusion ordered regression spatiotemporal model includes a shared two-dimensional encoder, a bottleneck time fusion module, a multi-scale temporal jump-connection fusion module, and a spatial decoder; The shared two-dimensional encoder is used to encode the multivariate marine meteorological field at each historical moment entered in the historical input window and extract multi-scale spatial features. The process is as follows: Assume a shared two-dimensional encoder includes Each scale layer, for historical input window The Middle Multivariable marine meteorological field at a historical moment , No. The encoded features of each scale layer are represented as follows: In the formula, Indicates the first The historical moment at the Spatial coding features at various scales; Indicates the shared two-dimensional encoder in the first... Encoding functions for each scale layer; ; For all historical moments in the history input window Stacking the scale-encoded features yields the ... Features of skip-connected sequences at various scales: In the formula, Indicated by anchor point time For the end time, the first Features of historical skip sequences at various scales; Indicates the first Number of channels per scale; and They represent the first Spatial height and width at various scales; Represents the space of real numbers; Indicates the length of the historical input time step; The bottleneck time fusion module is used to fuse multi-scale spatial features to obtain bottleneck features after bottleneck time fusion, and the process is as follows: For the bottleneck layer of the lowest resolution shared 2D encoder, let its historical sequence characteristics be... The bottleneck time fusion module uses a Convolutional Long Short-Term Memory (ConvLSTM) network, and its fusion process is represented as follows: In the formula, This indicates the bottleneck characteristics after bottleneck time fusion. This represents a convolutional long short-term memory network set in the bottleneck layer of the encoder; The multi-scale temporal skip connection fusion module is used to fuse historical skip connection sequence features in multi-scale spatial features to obtain scale skip connection features. The process is as follows: For the Features of skip sequences at various scales ,use Convolution performs channel reduction to decrease the computational cost of temporal modeling. The number of channels after dimensionality reduction is defined as: In the formula, Indicates the first Features of skip-connected sequences after dimensionality reduction of channels at various scales; Indicates the use of channel dimensionality reduction convolution; This represents the number of channels after dimensionality reduction; Indicates the minimum number of reserved channels; Indicates the channel compression ratio; Indicates rounding down; The features of the dimension-reduced skip sequence are fused using ConvLSTM to obtain the 1st... Historical dynamic aggregation features at various scales: In the formula, Indicates the first The temporal aggregated hidden state at the end of the historical input window at each scale; Indicates the first Convolutional long short-term memory networks in scale-hopping paths; Projecting the historical dynamic aggregated features back to the original channel dimension yields the historical dynamic supplementary features: In the formula, Indicates the first Historical dynamic supplementary features projected at various scales; Indicates the use of channel recovery convolution; Let the first The jump connection feature at each scale is: ,in Indicates the first Anchor point time at each scale Local spatial features; A gated fusion mechanism is employed to adaptively fuse recent local spatial features and historical dynamic supplementary features. The gate weights are defined as follows: The fusion of the first The scale-hopping feature is represented as follows: In the formula, Indicates the first Each scale of gating weights; This represents the Sigmoid activation function; This indicates splicing along the channel dimension; This represents element-wise multiplication; Indicates the first step after time-sequence jump connection fusion Each scale jump connection feature; The spatial decoder is used to fuse the bottleneck features obtained after bottleneck temporal fusion and the scale jump connection features to obtain shared spatiotemporal features, as follows: In the formula, Representation space decoder; This represents the shared spatiotemporal features obtained after decoding.

3. The multi-step risk early warning method for marine wind and waves based on threshold consistency according to claim 2, characterized in that: The specific process of step 7 is as follows: For the future At each prediction time, the wind level ordered evidence output head and the wave level ordered evidence output head are based on shared spatiotemporal feature representations. Generate the corresponding shared space response graph: In the formula, Indicates the future number Shared spatial response map of wind levels at each predicted moment; Indicates the future number Wave level shared spatial response diagram at each predicted moment; The wind level indicates the orderly output of evidence in the future. The convolution mapping function corresponding to each prediction time step; The wave level indicates the ordered evidence output head in the future. The convolution mapping function corresponding to each prediction time step; For spatial location The logit values ​​for wind and wave levels at their corresponding thresholds are defined as follows: In the formula, Indicates the future number Predicted time and spatial location The wind level has reached or exceeded the threshold. The logit value; Indicates the future number Predicted time and spatial location The wave level has reached or exceeded the threshold. The logit value; Indicates the wind level in the future. The predicted time, the first Bias terms at each threshold; Indicates the wave level in the future. The predicted time, the first Bias terms at each threshold; Based on the threshold logit of wind and wave levels, the exceedance probabilities of the wind level threshold and wave level threshold are respectively: In the formula, Indicates the future number Predicted time and spatial location The wind level has reached or exceeded the threshold. The probability of; Indicates the future number Predicted time and spatial location The wave level has reached or exceeded the threshold. The probability of; This represents the Sigmoid activation function; Indicates the future number Predicted time and spatial location Wind level label at the location; Indicates the future number Predicted time and spatial location Wave level label; Represents a probability operator; Indicates in When the wind level reaches or exceeds the level class; Indicates in Under the condition that the wave level reaches or exceeds the first class; For a single sample, the ordered evidence output of wind level is represented as follows: In the formula, Indicates sample The corresponding multi-step wind level threshold exceedance probability set, which only includes the wind level threshold exceedance probability on valid ocean grid points; The wave level ordered evidence output is represented as follows: In the formula, Indicates sample The corresponding multi-step wave level threshold exceedance probability set, which only includes wave level threshold exceedance probabilities on valid ocean grid points.

4. The multi-step risk early warning method for marine wind and waves based on threshold consistency according to claim 3, characterized in that: The specific process of step 8 is as follows: Define variable type ,in Represents the wind level variable. Represent the wave level variable; let the variable be... The number of grade categories is Its threshold index is For the future At each prediction time, variables At the threshold The logit at this location is represented as: In the formula, Indicates the future number Predicted time and spatial location Variable Reaching or exceeding the threshold The logit value; Representing variables In the future Predicted time and spatial location The shared space at the location responds; Representing variables In the future The predicted time, the first Bias terms at each threshold; Among them, the bias term The following monotonic parameterization method is adopted: In the formula, Representing variables In the future The learnable parameters corresponding to the first threshold bias term at each prediction time; A learnable parameter that controls the difference in adjacent threshold biases; Represents a non-negative activation function; The non-negative activation function can be expressed as: In the formula, Indicates input variables; because , by formula To the style We can obtain: In the formula, Representing the same variable At the same prediction time Below, the threshold bias terms of different levels satisfy a monotonically non-increasing relationship; Due to the same variable At the same prediction time and the same spatial location Under these conditions, different thresholds share the same spatial response. And only through the threshold bias term Therefore, we have: In the formula, This indicates that at the same valid ocean grid point, the logit values ​​corresponding to different threshold levels satisfy a monotonically non-increasing relationship; Due to the Sigmoid function It is a monotonically increasing function, and the variable is... The threshold exceedance probability satisfies: In the formula, Indicates the future number Predicted time and spatial location Variable Reaching or exceeding the threshold The probability of; In wind level variables In this case, the above monotonic relationship is specifically expressed as: In the formula, Indicates the future number Predicted time and spatial location The wind level has reached or exceeded the threshold. The probability, ; In wave level variable In this case, the above monotonic relationship is specifically expressed as: In the formula, Indicates the future number Predicted time and spatial location The wave level has reached or exceeded the threshold. The probability, .

5. The multi-step risk early warning method for marine wind and waves based on threshold consistency according to claim 4, characterized in that: The process of step 9 is as follows: Let the preset wind level risk threshold be... The preset wave level risk threshold is For the future Predicted time and spatial location For valid ocean grid points, wind-triggered evidence is defined as: In the formula, Indicates the future number Predicted time and spatial location Evidence triggered by the wind at the location; The wind level ordered evidence output head is at the threshold. The probability of exceeding the threshold at that point; Indicates the future number Predicted time and spatial location Wind level label at the location; Representation of shared spatiotemporal features; Indicates in Wind level is greater than or equal to the threshold under the condition ; Wave-triggered evidence is defined as: In the formula, Indicates the future number Predicted time and spatial location Evidence triggered by the waves at the location; The output header indicates that the wave level is ordered at the threshold. The probability of exceeding the threshold at that point; Indicates the future number Predicted time and spatial location Wave level label; Indicates in Under the condition that the wave level is greater than or equal to the threshold ; For a single prediction moment, the wind-triggered evidence map is represented as follows: In the formula, Indicates sample In the future Wind-triggered evidence maps for each predicted moment, and evidence maps are only generated and used on valid ocean grid points; The wave-triggered evidence diagram is represented as follows: In the formula, Indicates sample In the future Wave-triggered evidence maps for each predicted moment, and evidence maps are only generated and used on valid ocean grid points; For all future prediction steps, the wind-triggered evidence sequence is represented as: In the formula, Indicates sample The corresponding multi-step wind-triggered evidence sequence; Indicates the number of steps to predict the future; The wave-triggered evidence sequence is represented as follows: In the formula, Indicates sample The corresponding multi-step wave triggering evidence sequence.

6. The multi-step risk early warning method for marine wind and waves based on threshold consistency according to claim 5, characterized in that: The process of constructing the multi-task loss function and optimizing the model parameters in step 12 is as follows: For the future Predicted time and spatial location Effective ocean grid points, based on real wind level labels. Construct a cumulative binary monitoring label for wind level: In the formula, Indicates wind level at the threshold Cumulative binary supervision label at the location; Indicates the actual wind level label; ; Based on real wave level labels Construct cumulative binary supervised labels for wave levels: In the formula, Indicates wave level at the threshold Cumulative binary supervision label at the location; Indicates the actual wave level label; ; The constructed multi-task loss function includes a wind level ordered loss function, a wave level ordered loss function, and a risk warning loss function; wherein, the wind level ordered loss function, the wave level ordered loss function, and the risk warning loss function are all constructed based on the binary cross-entropy loss function; The binary cross-entropy loss function is: In the formula, Indicates the model's predicted probability; This represents the corresponding binary supervision label; Based on the probability of exceeding the wind level threshold The wind level-based losses are as follows: In the formula, Indicates wind level-based losses; Indicates the number of steps to predict the future; Represents the effective set of ocean grid points; Indicates the number of valid ocean grid points; Indicates the number of wind level categories; Based on the probability of exceeding the wave level threshold The wave level ordered loss is: In the formula, Indicates ordered loss based on wave level; Indicates the number of wave classification categories; This indicates that the wave level has reached or exceeded the threshold. The predicted probability; Based on risk probability and real risk labels The risk warning loss is: In the formula, Indicates risk warning and loss; Indicates the true risk label; This represents the risk probability derived from evidence of wind and wave thresholds. Indicates the weight of positive risk samples; The model's total loss function is: In the formula, Represents the total loss function of the model; Indicates wind level-based losses; Indicates ordered loss based on wave level; Indicates risk warning and loss; , and These represent the weighting coefficients for wind-level ordered losses, wave-level ordered losses, and risk warning losses, respectively. The objective for optimizing model parameters can be expressed as: In the formula, This represents the optimal model parameters obtained after training. This represents the learnable parameters in the temporal jump-connected fusion ordered regression spatiotemporal model; This indicates that when the model parameters are The corresponding total loss function.

7. The multi-step risk early warning method for marine wind and waves based on threshold consistency according to claim 6, characterized in that: Risk labeling of effective ocean grid points in step 4 for: In the formula, Indicates the future number Predicted time and spatial location Risk label at the location; Indicates the wind level label for the corresponding grid point; Indicates the wave level label for the corresponding grid point; This indicates the preset wind level risk threshold; This indicates the preset wave level risk threshold; Indicates the prediction step index; and These represent the row index and column index of the spatial grid, respectively; Based on whether the wind and wave levels reach the corresponding risk thresholds, the risk samples are divided into wind-triggered risk samples, wave-triggered risk samples, and combined wind and wave-triggered risk samples; among them, wind-triggered risk samples are defined as follows: The risk sample for wave-triggered events is defined as follows: The risk sample triggered by complex wind and waves is defined as follows: In the formula, This indicates that only wind-triggered risk samples are labeled; This indicates that only wave-triggered risk sample labels are used; This indicates a risk sample label triggered by complex winds and waves.

8. The multi-step risk early warning method for marine wind and waves based on threshold consistency according to claim 1, characterized in that: The historical multivariate marine meteorological state sequence includes marine meteorological variables such as zonal wind component at 10m altitude, meridional wind component at 10m altitude, 10m wind speed, significant wave height, mean wave period, wave direction, and sea level pressure.

9. A system for implementing the multi-step risk early warning method for marine wind and waves based on threshold consistency as described in any one of claims 1-8, characterized in that, include: The marine meteorological data acquisition and preprocessing module is used to acquire marine meteorological reanalysis data of the target sea area and perform preprocessing to obtain historical multivariate marine meteorological state sequences. An effective ocean grid mask module is used to construct an effective ocean grid mask for a target ocean area based on the land-sea distribution of the target ocean area; The wind and wave rating sub-module is used to convert continuous wind speed and continuous significant wave height into wind rating labels and wave rating labels with physical meaning and rating order, respectively, according to the standard ocean description rating. The wind and wave risk label construction module is used to construct wind and wave risk labels on the effective ocean grid points based on the wind level label and the wave level label, according to the preset wind level risk threshold and wave level risk threshold. The short-time multi-step sample construction module is used to construct short-time multi-step prediction samples based on the historical multivariate marine meteorological state sequence, the effective marine grid mask, the wind level label, the wave level label, and the wind and wave risk label, using a sliding window method. The short-time multi-step prediction samples include a historical input window and a future prediction window. The shared spatiotemporal feature extraction module is used to construct a temporal jump-connection fusion ordered regression spatiotemporal model. Based on the short-time multi-step prediction sample data input through the historical input window of the short-time multi-step prediction sample, the module processes the data using the temporal jump-connection fusion ordered regression spatiotemporal model to extract the multi-scale spatial structure features and temporal evolution features of the historical wind and wave field, thereby obtaining shared spatiotemporal features for wind and wave ordered evidence output and risk probability derivation. The wind and wave ordered evidence output module is used to construct wind level ordered evidence output header and wave level ordered evidence output header based on the shared spatiotemporal features, and output the wind level threshold exceedance probability and wave level threshold exceedance probability at multiple future prediction times. The monotonic order constraint module is used to apply monotonic order constraints to the output wind level threshold exceedance probability and wave level threshold exceedance probability to obtain evidence of wind and wave order. The wind and wave risk threshold evidence extraction module is used to extract wind triggering evidence and wave triggering evidence corresponding to the wind level risk threshold and the wave level risk threshold from the ordered wind and wave evidence. The threshold consistency risk probability derivation module is used to calculate a multi-step risk probability map based on the wind triggering evidence, the wave triggering evidence, the wind level risk threshold, and the wave level risk threshold, using the threshold consistency rule. The binary risk warning map and risk boundary uncertainty generation module is used to generate a binary risk warning map based on the multi-step risk probability map and according to a preset risk probability judgment threshold, and to calculate the risk boundary uncertainty map. The model training and optimization module is used to construct a multi-task loss function and optimize the model parameters for the temporal jump-connection fusion ordered regression spatiotemporal model, and to train and validate the temporal jump-connection fusion ordered regression spatiotemporal model based on the short-time multi-step prediction samples to obtain the model with the best performance on the validation set. The short-term multi-step risk output module is used to predict and infer the latest historical multivariate marine meteorological field of the target sea area based on the model with the best performance on the validation set, and generate short-term multi-step risk warnings for the sea. The short-term multi-step risk warnings for the sea include the ordered threshold exceedance probability of wind and waves at multiple future prediction times, risk probability map, binary risk warning map, and risk boundary uncertainty map.

10. The system according to claim 9, characterized in that: The temporal jump-connection fusion ordered regression spatiotemporal model includes a shared two-dimensional encoder, a bottleneck time fusion module, a multi-scale temporal jump-connection fusion module, and a spatial decoder. The shared two-dimensional encoder is used to encode the multivariate marine meteorological field at each historical moment entered in the historical input window and extract multi-scale spatial features. The process is as follows: Assume a shared two-dimensional encoder includes Each scale layer, for historical input window The Middle Multivariable marine meteorological field at a historical moment , No. The encoded features of each scale layer can be represented as: In the formula, Indicates the first The historical moment at the Spatial coding features at various scales; Indicates the shared two-dimensional encoder in the first... Encoding functions for each scale layer; ; For all historical moments in the history input window Stacking the scale-encoded features yields the ... Features of skip-connected sequences at various scales: In the formula, Indicated by anchor point time For the end time, the first Features of historical skip sequences at various scales; Indicates the first Number of channels per scale; and They represent the first Spatial height and width at various scales; Represents the space of real numbers; Indicates the length of the historical input time step; The bottleneck time fusion module is used to fuse multi-scale spatial features to obtain bottleneck features after bottleneck time fusion, and the process is as follows: For the bottleneck layer of the lowest resolution shared 2D encoder, let its historical sequence characteristics be... The bottleneck time fusion module uses a Convolutional Long Short-Term Memory (ConvLSTM) network, and its fusion process is represented as follows: In the formula, This indicates the bottleneck characteristics after bottleneck time fusion. This represents a convolutional long short-term memory network set in the bottleneck layer of the encoder; The multi-scale temporal skip connection fusion module is used to fuse historical skip connection sequence features in multi-scale spatial features to obtain scale skip connection features. The process is as follows: For the Features of skip sequences at various scales ,use Convolution performs channel reduction to decrease the computational cost of temporal modeling. The number of channels after dimensionality reduction is defined as: In the formula, Indicates the first Features of skip-connected sequences after dimensionality reduction of channels at various scales; Indicates the use of channel dimensionality reduction convolution; This represents the number of channels after dimensionality reduction; Indicates the minimum number of reserved channels; Indicates the channel compression ratio; Indicates rounding down; Then, ConvLSTM is applied to the dimensionality-reduced skip sequence features to obtain the 1st... Historical dynamic aggregation features at various scales: In the formula, Indicates the first The temporal aggregated hidden state at the end of the historical input window at each scale; Indicates the first Convolutional long short-term memory networks in scale-hopping paths; Next, the historical dynamic aggregated features are projected back to the original channel dimension to obtain the historical dynamic supplementary features: In the formula, Indicates the first Historical dynamic supplementary features projected at various scales; Indicates the use of channel recovery convolution; Let the first The jump connection feature at each scale is: ,in Indicates the first Anchor point time at each scale Local spatial features; A gated fusion mechanism is employed to adaptively fuse recent local spatial features and historical dynamic supplementary features. The gate weights are defined as follows: The fusion of the first The scale-hopping feature is represented as follows: In the formula, Indicates the first Each scale of gating weights; This represents the Sigmoid activation function; This indicates splicing along the channel dimension; This represents element-wise multiplication; Indicates the first step after time-sequence jump connection fusion Each scale jump connection feature; The spatial decoder is used to fuse the bottleneck features obtained after bottleneck temporal fusion and the scale jump connection features to obtain shared spatiotemporal features, as follows: In the formula, Representation space decoder; This represents the shared spatiotemporal features obtained after decoding.

Images